WO2000065985A2 - Method and system for knowledge guided hyperintensity detection and volumetric measurement - Google Patents
Method and system for knowledge guided hyperintensity detection and volumetric measurement Download PDFInfo
- Publication number
- WO2000065985A2 WO2000065985A2 PCT/US2000/011457 US0011457W WO0065985A2 WO 2000065985 A2 WO2000065985 A2 WO 2000065985A2 US 0011457 W US0011457 W US 0011457W WO 0065985 A2 WO0065985 A2 WO 0065985A2
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- location
- identifying
- imaging
- imaging scan
- brain
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30016—Brain
Definitions
- This invention relates generally to the field of magnetic resonance imaging ("MRI") and, in particular, to the automation of the interpretation of information present in MRI images to detect brain lesions. More specifically, the present invention relates to a method and/or system of detecting hyperintense regions in MRI images that are suspected of being related to various brain pathologies, such as Alzheimer's Disease, Multiple Sclerosis and like neurodegenerative conditions and/or determining volumetric measurements of cerebral anatomical regions.
- MRI magnetic resonance imaging
- Cline et al present a technique to produce a 3D segmentation of the head employing TI and T2 images.
- This method requires an initial visual classification of a relatively small sample of the tissue types in the TI and T2 images by an experienced radiologist to begin.
- the accuracy of this initial input directly affects the accuracy of the succeeding classification algorithm, which classifies the remaining tissue based on a bivariate distribution model for each tissue type.
- the possible classification outputs are; background, brain, CSF, WM, lesion, tumor, arteries or veins.
- the resulting surfaces are filtered twice using a 3D diffusion filter to smooth discontinuities created by misclassifications and to improve rendering.
- a 3D connectivity algorithm then extracts surfaces from this smoothed, segmented data set.
- a dividing cubes algorithm is used to process the voxels marked in the connectivity algorithm for display. This method is rated with only preliminary results for three selected patients; one normal, one MS case and one tumor. It is important to reiterate that success of the classification depends upon the operator's level of expertise. This method may be most useful as a surgical planning tool, or perhaps as a visualization tool, which appears to be the original intent.
- Hohol et al.7 have adopted a similar technique. This method begins by manually isolating the intracranial cavity using the PD and T2 images. Regions of interest (ROI) are then generated for brain parenchyma and CSF to aid the Expectation Maximization (EM) tissue classifier. Each image for a patient is registered to a time reference image, and the EM classifier is used to classify each pixel of each image as either WM, GM, CSF, or lesion tissue. This classification includes an unspecified correction for partial voluming artifacts. Individual lesions are fixed by the use of a 4 dimensional connectivity algorithm that is assumed to similar to Cline ⁇ with the addition of a time dimension. With use of the time dimension element, lesion volumes are reported to change over time. No verification of the imaging technique is reported. However, the authors do report lesion burden correlation with the neuropsychological test scores of MS patients.
- ROI Regions of interest
- EM Expectation Maximization
- Wicks et l employ two manually selected thresholds to perform white matter lesion (WML) detection.
- the first is set to separate brain tissue from the skull, while the second is set to identify all areas definable as lesions having intensities greater than the brain tissue.
- the threshold is set somewhat lower in value than that is required to identify the lesions.
- the outlines of the identified lesions are superimposed on the original T2 image to allow manual correction of misclassification. It is reported that this requirement poses no additional burden since it prevents the operator from having to outline each lesion manually. It is also stated that the established threshold can be used for any successive serial study of one patient provided that the intensity histograms from later studies are scaled to match that of the initial study.
- this semi-automated thresholding technique generates approximately half the inter-observer variation than purely manual outlining technique, but it is stated that visual assessment of the delineation of the lesions between the manual technique and this semi-automatic one are "equally plausible".
- the semi-automatic double threshold method is probably suitable for use with small patient sets in a non-critical research environment.
- Zijdenbos et alA describe a semi-automatic method based on pattern recognition and built around a back-propagation artificial neural network (ANN). This is done to minimize the required input training points needed to achieve a successful tissue classification.
- the method begins with the use of an inter-cranial contour algorithm to remove the skull and incidental CSF, followed by an intensity correction algorithm to remove the shading, or intensity inhomogeneity artifact. This correction requires the operator to select 10 to 20 points in the WM that therefore strongly influences the detected lesion load.
- the intensity correction is followed by the same diffusion filter mentioned in Cline6 to enhance the signal/noise ratio.
- tissue classification occurs using the ANN with 3 input nodes (T1,T2,PD) and five output nodes; background, WM, GM, CSF, and WML.
- T1,T2,PD input nodes
- WML weighted layer
- One sample of each class input tissue class is presented in sequence to the ANN until convergence (segmentation) occurs.
- postprocessing is required to remove the WML that occur in close proximity to GM along the sulci and correct the classification errors caused by misregistration of the TI image.
- all WML smaller than 10 pixels are eliminated as they are assumed to be the result of noise or misregistration in the TI image. This semi-automated technique is compared with a completely manual method using two different observers.
- the classifier is able to separate CSF and WM.
- the kNN classifier is used to detect pixels representing lesion pathology. Consequently, the image is classified into WM, CSF, and WML without reference to GM. All pixels labeled as lesions are then highlighted on the MRI image for operator verification. This process is reported as requiring several iterations involving operator refinement of the classifications. It is reported that inter-rater reliability for this semi-automatic method is roughly half of that reported by the purely manual method, while showing a 25% improvement over the manual method with respect to intra-rater reliability. Therefore, very little diagnostic leverage is obtained since expert knowledge is required to pass judgment on lesion identification. It however does appear to be an excellent tool for documenting disease progression if used by an experienced neuroradiologist.
- Mitchell et al ⁇ report another method which is a manual outlining technique that employs a novel use of the multispectral data to enhance clinical diagnosis.
- This method begins with routine outlining of the MS lesions by a radiologist.
- a tissue intensity analysis is subsequently performed with the assumption that the CSF characteristics will not change over time and thus can be used to calibrate, or normalize, the image intensities of the PD and T2 slices.
- the same technique is applied to "normal appearing" WM regions to develop a cluster mean to better differentiate it from "abnormal” WM.
- An equivalent WM-GM (eW-G) spectrum is formed by projecting this 2D multispectral feature data onto a ID axis using a principle component analysis.
- the unique feature of this technique is that the measured changes in the eW-G spectrum correlation well with disease progression especially when lesion size changes are not apparent. It is stated that the use of the eW- G spectra could relax the requirement for accurately defining lesion boundaries in the measure of disease progression. This technique would be applicable as a research or diagnostic tool to identify disease progression based on the "eW-G" spectra, while specifically useful in diagnosis of MS with its relatively large diffuse lesions.
- Samarasekera et al ⁇ > report a semi-automatic method that utilizes the fuzzy connectedness principles to achieve lesion segmentation.
- This method begins by selecting 10 centrally located slices from within the imaged head and constructing an intensity histogram. From this, an empirical intensity threshold is set by finding the first pixel value which is greater in intensity than the second highest peak in the histogram, but has a bin count that is 93% less than that second highest peak. This threshold is used to create a binary volume containing only the brightest appearing pixels. The original 3D volume image is next thresholded at a low value to create another binary image corresponding to skull tissue. These two binary volume images are subsequently analyzed with a connected component algorithm and any components connected in both binary images are considered blood vessels and eliminated.
- Any remaining voxels in the first binary volume are thought to represent true and false positive lesions.
- a fuzzy connectedness algorithm is applied to the original volume image, using the previously derived binary volume image, as a template to determine membership grades based on neighborhood connectedness and intensity similarity. Fuzzy objects formed on the basis of the fuzzy threshold (40%) are tested for connection to the scalp binary volume image, and if connected are eliminated. Any fuzzy object with a membership of greater than 1500, or less than 7 voxels is also eliminated. Finally, any remaining objects are now superimposed on slice images for operator approval as lesions. Test results, using four neuroradiologists, show that this method has a detection sensitivity of 97% compared to the radiologists, and a false negative volume fraction of 1.3%, with false positives virtually non-existant.
- Mitchell et al ⁇ describe a semi-automatic technique to segment MS lesions in 3D data sets using only the T2 and PD weighted images.
- the technique generates 2D histograms of operator defined ROI to allow feature space classification of these regions using either an interactive kNN classification algorithm or a maximum likelihood (ML) classifier.
- ML maximum likelihood
- two operator selected thresholds determine the classification parameters for each cluster based on proximity of the threshold to the data to be clustered.
- mean and covariance values from the tissue ROI are used to establish those same parameters for the data clusters.
- a principle component analysis is used to establish an elliptical region about the mean for classification of each cluster.
- Knowledge supplied by the operator is used to establish confidence intervals and vary the classification thresholds accordingly.
- the resulting tissue classes are assigned colors and applied to the 3D image.
- This method is not able to directly segment lesion without operator direction, but once manually detected, is able to quantify their volumes.
- the accuracy of the method's ability to calculate volumes is measured against phantom studies with the assumption that larger measured volumes have less error.
- the technique requires an extensive amount of specialized knowledge from the operator, and is likely not to be useful except to test the concept of the classifying tools.
- Vinitski et al $ describe a similar method also based on a kNN classifier.
- the RF inhomogeneity is corrected in the T1,T2 and PD images by applying a correction matrix developed from imaging an oil filled cylinder.
- the 3D anisotropic diffusion filter is then applied to remove partial volume effects on image voxels, while leaving edges and small lesions relatively distinct.
- An expert observer is required to identify multiple samples of 8 different tissue types (as per Cline ⁇ ) to initiate the kNN classifier, which employed a "k" of 20 for 2D classification, and a "k” of 40 for 3D classification.
- the resulting tissue classifications are then judged by 5 board certified neuroradiologists for accuracy.
- Pannizzo et al ⁇ report a semi-automated method of MS lesion quantification that is histogram driven.
- the method using an edge following algorithm to remove skull tissue that is subsequently judged by an operator. And, if needed, selects a new threshold in case of rejection.
- a histogram of the remaining brain tissue is generated and presumed bimodal.
- the brain tissue is segmented into two distributions; a central distribution corresponding to WM/GM tissue, and higher intensity pixels that correspond to MS lesions and periventricular effusions (PVE). Thresholds are calculated from the histogram where fitted lines to the sides of the central distribution cross the horizontal axis. The operator is again asked to judge the performance of this step and make the appropriate corrections if needed.
- Kapouleas ⁇ 7 describes an automated system that requires the use of a model or atlas to remove false positive lesions.
- the brain tissue is segmented from the skull using the PD images, while lesions are detected using a simple threshold in the T2 slices.
- An initial removal of false positives lesions is completed by measuring the average intensity of the each brain slice based on the outline derived during the PD segmentation.
- the threshold is adjusted such that all pixels not brighter than 30% of the average in both the PD and T2 slice are rejected.
- a 3D representation is constructed from the stacked PD segmentations and a "locally deformable" 3D geometric atlas is made to fit the imaged brain. Using the locations defined by this atlas, the remaining false positives are removed.
- This method is judged for validity by comparing lesion output counts/locations with that of radiologists.
- the technique achieves a reported 87% agreement with the radiologists for axial images while falling to 78% agreement for coronal images.
- the lack of detail in the description of the method and the relatively low reported detection sensitivities lead to the conclusion that this tool may not be as useful as the manual technique presented by Samaraskera ⁇ .
- Li et al. ⁇ report a knowledge-based method built around the fuzzy c-means algorithm (FCM).
- FCM fuzzy c-means algorithm
- the algorithm relies exclusively on intensity class relationships derived from the T2 slice and is restricted to those 5 slices lying immediately above or below the central axial slice in a image set.
- the analysis begins with tissue classification based on the unsupervised fuzzy-c means (UFCM) algorithm that yields 10 possible output classifications.
- the authors report that the high number of output classes reduce the changes of misclassify tumor tissue from otherwise normal tissue (WM/GM).
- the encoded knowledge comprises the distributions derived found from case studies, relating class center intensities to the presence or absence of tissue type.
- the classification is dependent upon the proximity of tissues with respect to the lateral ventricles.
- the class centers define "focus of attention" areas in the slice being analyzed, which are used to remove skull tissue and locate WM. Once identified, WM is subjected to a shape analysis to test for abnormalities (tumor detection). With WM identified and presence of tumor tissue accepted or rejected, CSF is then identified followed by GM identification. There is no relative performance measurement reported for this technique since a limited number of slices/patients are analyzed as proof of concept. However, it appears most useful in detecting tumors of size large enough to cause some distortion in the expected shape of the classified WM.
- Warfield et ⁇ /.19 demonstrate a method that employs an anatomical atlas to segment cortical and sub-cortical structures and distinguish WM lesions.
- This method subjects the input slice to the 3D diffusion filter previously mentioned to correct for inhomogeneity.
- the skull is removed semi- automatically and the intracranial tissue is segmented by an EM algorithm.
- the anatomical atlas is co-registered with a classified 3D image and elastic matching is used to fit anatomical structures to the classified brain.
- the initial EM classifier and the elastically matched atlas can adequately identify all structures in the brain except for the cortex.
- the cortex is segmented by the use of a seed growing algorithm that is dependent upon a separate model.
- the WM and any incident lesions are segmented by removing the GM and CSF tissues previously identified.
- a two-class minimum distance classifier is used on the identified WM.
- the performance of the cortex segmentation is compared against 5 different raters and is reported to achieve a 95% success rate against a best rater and 96% accuracy as measured with respect to a "standard" cortex.
- the method is unique in that it segments all possible GM tissue before attempting to segment WM/WML thus avoiding the problem inherent in all methods that segment WM lesions based on pixel intensity criteria. There is no published information concerning its ability to distinguish WM lesions so the method cannot be judged.
- Johnston et alT ⁇ present an involved method of MS lesion detection that is limited to slices occurring adjacent to the central axial slice as in Li 18 above.
- This method employs PD and T2 slices and performs 8 bit scaling on the data.
- the skull is removed and the image with the remaining intracranial tissue is filtered with a non-linear low-pass filter to correct for RF inhomogeneity.
- a variation of the iterated conditional modes (ICM) algorithm is applied to perform a preliminary 3D segmentation. This step yields a separate 3D image for each tissue type (4) and assigns to each voxel an 8 bit intensity value that represents the probability that the voxel belongs to that tissue class.
- the method requires initial operator selection of pure tissue samples to generate histograms for each tissue type.
- the first step is the merging of the two (PD & T2) 3D WM probability maps generated by ICM.
- the second is the re-application of ICM to this merged data set to generate a WM/WM lesion mask allowing the segmentation of lesions from WM without interference from GM intensities.
- Goldberg-Zimring et al?- demonstrate an automated technique using an ANN classifier combined with an anatomical map. This method is built on the assumptions that MS lesions in T2 and PD weighted scans are much brighter than other brain tissue, that non MS regions of the brain will be very large or very small in size, that MS lesions are relatively circular, and that most of the MS lesions occur almost exclusively in the periventricular WM.
- the subject image is first normalized so that its maximum pixel intensity is equal to 1.
- a sliding window thresholder is then applied to the image and returns a "1" for the central pixel of the window if that image pixel has a value greater than 0.5, otherwise is "0".
- the ANN is supplemented by an unspecified post-processing stage that removed artifacts caused by ANN.
- the reported sensitivity of 0.87 and specificity of 0.96 for this technique is based on a group of 45 images.
- the technique produces a vast number of false positive lesions before the application of the ANN and might be significantly improved by employing a more sophisticated thresholding technique.
- the lack of specification of the final post-processing stage that corrects the ANN output raises some questions about the veracity of the sensitivity and specificity reports.
- DeCarli et al.2 describes a fully automatic global thresholding method for segmenting WM hyperintensities.
- a segmentation ⁇ of the PD image is performed to identify cerebral cortex, CSF, and brain volumes.
- PD image pixels representing brain tissue are added to T2 image pixels representing brain tissue and a histogram of the combined images is generated.
- the histogram is modeled as a Gaussian allowing application of standard statistical moments to the pixel distribution. All pixels having intensities greater than 3 standard deviations above the mean for this distribution are considered to be WM hyperintensities.
- Brunetti et al.23- 5 present a simple technique (“Quantitative Magnetic Color Imaging”) that can be considered fully automated as the operator is only asked to view the final output image.
- This method employs TI, T2 and PD images and calculates relaxation rate maps for each, based on the selection of different repetition times (TR) parameters during initial image acquisition. These parametric maps are then color coded (red for 1/T1, green for 1/T2, blue for PD) and combined into one multispectral color image. The colored image is displayed using a predefined color map in which the colors directly reflect the calculated parameter values. A violet-blue color on the map indicates suspected lesions.
- such calculated volumetric measurements are used, for example, to indicate atrophy during a pathological process.
- a pathological process optionally includes a dementia evidencing a physical change in the brain.
- a dementia evidencing a physical change in the brain.
- the instant invention detects lesion tissue in subcortical regions of a brain, which may indicate a dementia, such as, Alzheimer's disease, Parkinson's disease, Huntington's disease, and a purely aging-associated dementia.
- a system and/or method of providing a cerebral region template generated from a patient's own morphology optionally enhances accurate location of regions of interest in the brain and accurate analysis thereof.
- the present invention provides a method and/or system for detecting hyperintense regions in MRI images which addresses the problems associated with human subjectivity and existing thresholding techniques.
- the present invention introduces the concept of the use of knowledge guided rules and methods for automatically locating certain anatomical regions and detecting hyperintensities associated with same, which in turn are candidates for possible lesions.
- Yet another feature and advantage of the present invention is to provide a method and/or system for processing the information and/or data provided by MRI images to detect and identify hyperintensities for disease screening, and thereby serve as a diagnostic aid in helping neuroradiologists verify disease diagnosis and severity, to arrive at a prognosis, or to follow the possible effects of therapy (e.g., drug therapy).
- therapy e.g., drug therapy
- KGHID knowledge-guided hyperintensity detection
- the method/system herein described optionally requires no more than a reliable initial segmentation of brain tissues into classes of, for example, cerebral spinal fluid, white matter, gray matter and mixed boundary tissue. KGHID is then able to identify subcortical structures and hyperintense lesions using these tissue classes and encoded anatomical knowledge. This knowledge consists of pixel intensity relationships as found in the classified tissues.
- Another feature and advantage of this method or system is its ability to detect automatically lesions confined within white matter tissue around the lateral ventricles or within desired subcortical structures. Lesions in these areas are thought to be attributable to various neurodegenerative diseases.
- the method detects hyperintensities within various tissues in the brain, including but not limited to potential lesions in white matter, a periventricular ring, and possible lesions in lenticular nuclei.
- the structures already delineated as part of the analysis form the basis for further implementing lesion detection within the caudate nuclei and the thalamus.
- the method/system requires no initial assumptions about the relative health of the brain being analyzed.
- the present technique has been proven to work with in vivo and in vitro brain images.
- the method/system of the present invention is preferable applied to axial images. It could also be adapted for use with coronal, sagittal, or like images.
- the method/system of the present invention requires no operator intervention nor reference to an anatomical atlas, it is ideally suited to real-time imaging. While the algorithm described herein is written in IDL, a computer language suited to prototyping, it could be easily ported into C or other software languages. It is a feature and advantage of the instant invention to provide a system and/or method for determining volumetric measurements of cerebral regions, automatically or free of subjective intervention by a user.
- the instant invention calculates volumetric measurements of subcortical regions of a brain.
- volumetric measurements are used, for example, to indicate atrophy during a pathological process.
- a pathological process optionally includes a dementia evidencing a physical change in the brain.
- I have recognized atrophy or volume shrinkage in a hippocampus coincident with an Alzheimer's disease process.
- the instant invention detects lesion tissue in subcortical regions of a brain, which may indicate a dementia, such as, Alzheimer's disease, Parkinson's disease, Huntington's disease, and a purely aging-associated dementia.
- a cerebral region template generated from a patient's own morphology.
- such a template optionally enhances accurate location of regions of interest in the brain and accurate analysis thereof.
- the instant invention provides a method of interpreting at least one imaging scan of a patient.
- the method includes the following sequential, non-sequential, or sequence-independent steps.
- a processor for example, (a) identifies a location of at least one cerebral region in the imaging scan based, at least in part, on a relative location of a lateral ventricle.
- the identifying step (a) is free of human intervention and/or is automatic.
- such a step at least substantially eliminates human subjectivity, which may vary from user to user.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the processor (b) identifies a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans, the identifying step (b) is free of human intervention and/or is automatic.
- the processor (c) determines a volumetric measurement for at least one of the cerebral regions.
- the processor (d) identifies a suspected presence of white matter lesion tissue in the imaging scan based on a knowledge base relating voxel intensity distributions and a spatial relationship of cerebral anatomical structures.
- At least one cerebral region includes at least one of a caudate nucleus, a lenticular nucleus, a thalamus, a hippocampus, a periventricular ring, white matter, a brain stem, and a cerebellum.
- the identifying step (a) includes determining a location of the caudate nucleus, at least in part, from the relative location of the lateral ventricle.
- the identifying step (a) includes determining a location of the thalamus, at least in part, from at least one of a relative location of the caudate nucleus and the relative location of the lateral ventricle.
- the identifying step (a) includes determining a location of the lenticular nucleus, at least in part, from at least one of the relative location of the caudate nucleus and a relative location of the thalamus.
- the identifying step (a) includes determining a location of the hippocampus, at least in part, from the relative location of the lateral ventricle.
- the identifying step (a) includes determining a location of the periventricular ring, at least in part from at least one of the location of the lateral ventricle, the location of the caudate nucleus, and the location of the thalamus.
- the identifying step (d) includes identifying white matter regions in the at least one imaging scan, eliminating therefrom voxels identifiable as at least one of white matter, a perivascular space, and cerebrospinal fluid.
- the white matter regions include at least one of an internal capsule, a frontal lobe, an anterior temporal lobe, an anterior parietal lobe, a posterior parietal lobe, an occipital lobe, and a posterior temporal lobe.
- the processor generates a template unique to the patient, the template including each identified at least one cerebral region.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the processor (b) identifies a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans, using the generated template.
- an automated method of identifying suspected lesions in a brain includes the following sequential, non-sequential, sequence- independent steps.
- a processor (a) provides a magnetic resonance image (MRI) of a patient's head, including a plurality of slices of the patient's head, which MRI comprises a multispectral data set that can be displayed as an image of varying pixel intensities.
- the processor (b) identifies a brain area within each slice to provide a plurality of masked images of intracranial tissue.
- the processor (c) applies a segmentation technique to at least one of the masked images to classify the varying pixel intensities into separate groupings, which potentially correspond to different tissue types.
- the processor (d) refines the initial segmentation into the separate groupings of at least the first masked image obtained from step (c) using one or more knowledge rules that combine pixel intensities with spatial relationships of anatomical structures to locate one or more anatomical regions of the brain.
- the processor (e) identifies, if present, the one or more anatomical regions of the brain located in step (d) in other masked images obtained from step (c).
- the processor (f) further refines the resulting knowledge rule-refined images from steps (d) and (e) to locate suspected lesions in the brain.
- the magnetic resonance image includes a multispectral data set including proton density weighted (PDw), TI weighted (Tlw) and T2 weighted (T2w) acqusitions.
- the slices are taken in the axial, coronal, or sagittal planes of the patient's head.
- the varying pixel intensities are classified into at least four separate groupings, which potentially correspond to at least four different tissue types, including a first tissue type, a second tissue type, a third tissue type, and a fourth tissue type.
- the first tissue type comprises cerebrospinal fluid.
- the second tissue type comprises white matter.
- the third tissue type comprises gray matter.
- the fourth tissue type comprises white matter hyperintensities.
- the anatomical regions of the brain include at least one of lateral ventricles, caudate nuclei, lenticular nuclei, hippocampus, brain stem, cerebellum, and thalamus.
- the suspected lesions include hyperintense lesions embedded within the white matter.
- the suspected lesions comprise a periventricular ring.
- an apparatus for interpreting at least one imaging scan of a patient includes (a) first means for identifying a location of at least one cerebral region in the imaging scan based, at least in part, on a relative location of a lateral ventricle.
- the first identifying means (a) is free of human intervention.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the apparatus further includes (b) second means for identifying a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans.
- the second identifying means (b) is free of human intervention.
- the apparatus includes (c) means for determining a volumetric measurement for at least one of the each cerebral region.
- the apparatus includes (d) third means for identifying a suspected presence of white matter lesion tissue in the imaging scan based on a knowledge base relating voxel intensity distributions and a spatial relationship of cerebral anatomical structures.
- at least one cerebral region includes at least one of a caudate nucleus, a lenticular nucleus, a thalamus, a hippocampus, a periventricular ring, white matter, a brain stem, and a cerebellum.
- the first identifying means (a) includes means for determining a location of the caudate nucleus, at least in part, from the relative location of the lateral ventricle.
- the first identifying means (a) includes means for determining a location of the thalamus, at least in part, from at least one of a relative location of the caudate nucleus and the relative location of the lateral ventricle.
- the first identifying means (a) includes means for determining a location of the lenticular nucleus, at least in part, from at least one of the relative location of the caudate nucleus and a relative location of the thalamus.
- the first identifying means (a) includes means for determining a location of the hippocampus, at least in part, from the relative location of the lateral ventricle.
- the first identifying means (a) includes means for determining a location of the periventricular ring, at least in part from at least one of the location of the lateral ventricle, the location of the caudate nucleus, and the location of the thalamus.
- the third identifying means d) includes means for identifying white matter regions in the at least one imaging scan, eliminating therefrom voxels identifiable as at least one of white matter, a perivascular space, and cerebrospinal fluid.
- the white matter regions include at least one of an internal capsule, a frontal lobe, an anterior temporal lobe, an anterior parietal lobe, a posterior parietal lobe, an occipital lobe, and a posterior temporal lobe.
- the apparatus includes means for generating a template unique to the patient, the template including each identified at least one cerebral region.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the apparatus optionally further include (b) fourth means for identifying a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans, using the generated template.
- a computer readable medium including instructions being executed by a computer.
- the instructions instruct the computer to execute an interpretation of at least one imaging scan of a patient.
- the instructions include the following sequential, non-sequential, or sequence- independent steps.
- a processor (a) identifies a location of at least one cerebral region in the imaging scan based, at least in part, on a relative location of a lateral ventricle.
- the identifying instruction (a) is free of human intervention.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the processor (b) identifies a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans.
- the identifying instruction (b) is free of human intervention.
- the processor (c) determines a volumetric measurement for at least one of the each cerebral region.
- the processor (d) identifies a suspected presence of white matter lesion tissue in the imaging scan based on a knowledge base relating voxel intensity distributions and a spatial relationship of cerebral anatomical structures.
- At least one cerebral region includes at least one of a caudate nucleus, a lenticular nucleus, a thalamus, a hippocampus, a periventricular ring, white matter, a brain stem, and a cerebellum.
- the identifying instruction (a) includes determining a location of the caudate nucleus, at least in part, from the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the thalamus, at least in part, from at least one of a relative location of the caudate nucleus and the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the lenticular nucleus, at least in part, from at least one of the relative location of the caudate nucleus and a relative location of the thalamus.
- the identifying instruction (a) includes determining a location of the hippocampus, at least in part, from the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the periventricular ring, at least in part from at least one of the location of the lateral ventricle, the location of the caudate nucleus, and the location of the thalamus.
- the identifying instruction (d) includes identifying white matter regions in the at least one imaging scan, eliminating therefrom voxels identifiable as at least one of white matter, a perivascular space, and cerebrospinal fluid.
- the white matter regions include at least one of an internal capsule, a frontal lobe, an anterior temporal lobe, an anterior parietal lobe, a posterior parietal lobe, an occipital lobe, and a posterior temporal lobe.
- the processor generates a template unique to the patient, the template including each identified at least one cerebral region.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the processor (b) identifies a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans, using the generated template.
- a computer system for interpreting at least one imaging scan of a patient is provided.
- the computer system includes a processor.
- the computer system also includes a memory storing a computer program controlling operation of the processor.
- the program includes instructions for causing the processor to effect the following sequential, non-sequential, or sequence-independent steps.
- the processor (a) identifies a location of at least one cerebral region in the imaging scan based, at least in part, on a relative location of a lateral ventricle.
- the identifying instruction (a) is free of human intervention.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the processor (b) identifies a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans.
- the identifying instruction (b) is free of human intervention.
- the processor (c) determines a volumetric measurement for at least one of the each cerebral region.
- the processor (d) identifies a suspected presence of white matter lesion tissue in the imaging scan based on a knowledge base relating voxel intensity distributions and a spatial relationship of cerebral anatomical structures.
- At least one cerebral region includes at least one of a caudate nucleus, a lenticular nucleus, a thalamus, a hippocampus, a periventricular ring, white matter, a brain stem, and a cerebellum.
- the identifying instruction (a) includes determining a location of the caudate nucleus, at least in part, from the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the thalamus, at least in part, from at least one of a relative location of the caudate nucleus and the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the lenticular nucleus, at least in part, from at least one of the relative location of the caudate nucleus and a relative location of the thalamus.
- the identifying instruction (a) includes determining a location of the hippocampus, at least in part, from the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the periventricular ring, at least in part from at least one of the location of the lateral ventricle, the location of the caudate nucleus, and the location of the thalamus.
- the identifying instruction (d) includes identifying white matter regions in the at least one imaging scan, eliminating therefrom voxels identifiable as at least one of white matter, a perivascular space, and cerebrospinal fluid.
- the white matter regions include at least one of an internal capsule, a frontal lobe, an anterior temporal lobe, an anterior parietal lobe, a posterior parietal lobe, an occipital lobe, and a posterior temporal lobe.
- the processor generates a template unique to the patient, the template including each identified at least one cerebral region.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the processor (b) identifies a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans, using the generated template.
- the computer system includes a computer including the processor.
- the computer is optionally communicatable with a user via a computer network.
- the computer includes a Web server.
- an internet appliance in accordance with another embodiment of the instant invention, includes a thin client programmably connected via a computer network to a single web hosting facility.
- the single web hosting facility includes a server communicatable with a user via said thin client.
- the server is in communication with a processor and a computer readable medium including instructions being executed by a processor.
- the instructions instruct the computer to execute an interpretation of at least one imaging scan of a patient.
- the instructions includes the following sequential, non-sequential, or sequence- independent steps.
- the processor (a) identifies a location of at least one cerebral region in the imaging scan based, at least in part, on a relative location of a lateral ventricle.
- the identifying instruction (a) is free of human intervention.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the processor (b) identifies a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans.
- the identifying instruction (b) is free of human intervention.
- the processor (c) determines a volumetric measurement for at least one of the each cerebral region.
- the processor (d) identifies a suspected presence of white matter lesion tissue in the imaging scan based on a knowledge base relating voxel intensity distributions and a spatial relationship of cerebral anatomical structures.
- the at least one cerebral region includes at least one of a caudate nucleus, a lenticular nucleus, a thalamus, a hippocampus, a periventricular ring, white matter, a brain stem, and a cerebellum.
- the identifying instruction (a) includes determining a location of the caudate nucleus, at least in part, from the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the thalamus, at least in part, from at least one of a relative location of the caudate nucleus and the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the lenticular nucleus, at least in part, from at least one of the relative location of the caudate nucleus and a relative location of the thalamus.
- the identifying instruction (a) includes determining a location of the hippocampus, at least in part, from the relative location of the lateral ventricle.
- the identifying instruction (a) includes determining a location of the periventricular ring, at least in part from at least one of the location of the lateral ventricle, the location of the caudate nucleus, and the location of the thalamus.
- the identifying instruction (d) includes identifying white matter regions in the at least one imaging scan, eliminating therefrom voxels identifiable as at least one of white matter, a perivascular space, and cerebrospinal fluid.
- the white matter regions include at least one of an internal capsule, a frontal lobe, an anterior temporal lobe, an anterior parietal lobe, a posterior parietal lobe, an occipital lobe, and a posterior temporal lobe.
- the processor generates a template unique to the patient, the template including each identified at least one cerebral region.
- the at least one imaging scan includes a plurality of consecutive imaging scans.
- the processor (b) identifies a location of each cerebral region of the at least one cerebral region in a successive imaging scan of the plurality of imaging scans based, at least in part, on a location of a corresponding cerebral region in a preceding imaging scan of the plurality of imaging scans, using the generated template.
- Figure 1 is a sample intensity distribution graph for pixels or voxels relating to white matter hyperintensities in a proton density weighted data acquisition;
- Figure 2 is a sample, axial MRI scan of a patient's head
- Figure 3 is a sample, axial MRI scan, showing at least a lateral ventricle
- Figure 4 is a sample, axial MRI scan, showing at least caudate nuclei and lenticular nuclei
- Figure 5 is a sample, axial MRI scan, showing at least a periventricular ring
- Figure 6 is a sample, axial MRI scan, showing at least hyperintense lesions embedded within white matter
- Figure 7 is a sample, axial MRI scan, showing at least VGC segmentation of gray matter and white matter hyperintensities;
- Figure 8 is sample, axial MRI scan, showing lesion detection output consistent with the instant invention;
- Figure 9 is an illustrative embodiment of a flowchart consistent with the instant invention
- Figure 10 is another illustrative embodiment of a flowchart consistent with the instant invention
- Figure 11 is an illustrative embodiment of a computer and assorted peripherals
- Figure 12 is an illustrative embodiment of internal computer architecture consistent with the instant invention
- Figure 13 is an illustrative embodiment of a memory medium; and Figure 14 is an illustrative embodiment of a computer network architecture consistent with the instant invention.
- a procedure is here, and generally, conceived to be a self-consistent sequence of steps leading to a desired result. These steps are those requiring physical manipulations of physical quantities. Usually, though not necessarily, these quantities take the form of electrical or magnetic signals capable of being stored, transferred, combined, compared and otherwise manipulated. It proves convenient at times, principally for reasons of common usage, to refer to these signals as bits, values, elements, symbols, characters, terms, numbers or the like. It should be noted, however that all of these and similar terms are to be associated with the appropriate physical quantities and are merely convenient labels applied to these quantities.
- the manipulations performed are often referred to in terms, such as adding or comparing, which are commonly associated with mental operations performed by a human operator. No such capability of a human operator is necessary, or desirable in most cases, in any of the operations described herein which form part of the present invention; the operations are machine operations.
- Useful machines for performing the operation of the present invention include general purpose digital computers or similar devices.
- the present invention also relates to an apparatus for performing these operations.
- This apparatus may be specially constructed for the required purpose or it may comprise a general purpose computer as selectively activated or reconfigured by a computer program stored in a computer.
- the procedures presented herein are not inherently related to a particular computer or other apparatus.
- Various general purpose machines may be used with programs written in accordance with the teachings herein, or it may prove more convenient to construct more specialized apparatus to perform the required method steps. The required structure for a variety of these machines will appear from the description given.
- an automated method of identifying suspected lesions in a brain comprises: (a) providing a magnetic resonance image ("MRI") of a patient's head, including a plurality of slices of the patient's head, which MRI comprises a multispectral data set that can be displayed as an image of varying pixel intensities; (b) identifying a brain area within each slice to provide a plurality of masked images of intracranial tissue; (c) applying a segmentation technique to at least one of the masked images to classify the varying pixel intensities into separate groupings, which potentially correspond to different tissue types; (d) refining the initial segmentation into the separate groupings of at least the first masked image obtained from step (c) using one or more knowledge rules that combine pixel intensities with spatial relationships of anatomical structures to locate one or more anatomical regions of the brain; (e) if present, identifying the one or more anatomical regions of the brain located in step (d) in other
- the magnetic resonance image comprises a multispectral data set including proton density weighted (PDw), TI weighted (Tlw) and T2 weighted (T2w) acqusitions, and the slices are taken in the axial, coronal, or sagittal planes of the patient's head.
- the varying pixel intensities may be classified into a number of, preferably at least four, separate groupings, which potentially correspond to at least four different tissue types, including a first tissue type, a second tissue type, a third tissue type, and a fourth tissue type.
- the first tissue type comprises cerebrospinal fluid ("CSF")
- the second tissue type comprises white matter ("WM")
- the third tissue type comprises gray matter (“GM”)
- the fourth tissue type comprises white matter hyperintensities ("WMHI”).
- the anatomical regions of the brain which can be automatically located by the present method, include, but are not limited to, lateral ventricles, caudate nuclei, lenticular nuclei, or thalamus. Consequently, suspected lesions can be detected which comprise hyperintense lesions embedded within the white matter or which may comprise a periventricular ring.
- further refinement of the resulting knowledge rule- refined images of at least the third and fourth tissue types is carried out. Suspected lesions can be identified in subcortical regions using the present automated technique.
- the present invention comprises an automated technique to segment WM hyperintense lesion using TI, T2 and PD images.
- the present method begins by using a multiple threshold technique to remove the skull pixels from the image.
- the resulting multispectral data is submitted to a data clustering tool that utilizes FCM and a cluster partition validity index to recursively refine the tissue classification in the image.
- the cluster output consists of WM, GM, CSF and a fourth tissue type which contains, in part, the WM hyperintense lesions.
- the detection algorithm uses anatomical and empirical knowledge to selectively identify and analyze the subcortical structures (e.g., thalamus and basal ganglia) for lesion involvement. Further, it employs spatial reasoning based on the defined brain anatomy and statistical thresholding to detect and segment WM lesions. Only preliminary results are available. Kappa statistic indicates substantial agreement when compared to an expert neuroradiologist for a limited set of cases.
- the present method can be enhanced by incorporating the anisotropic filter mentioned in Wicks26 ; some variant of the CSF based intensity correction described by Mitchell 1 1 , and possibly the histogram matching method of Wang° ⁇ Addition of these techniques may increase the robustness of the algorithm and perhaps decrease its report of false positives.
- the algorithm can be enhanced by incorporating a database to capture past and present measures, therefore becoming useful in serial studies or to validate results of drug trials.
- the knowledge rules of the present invention comprise factors taking into account pixel intensity and spatial relationships of anatomical features.
- the knowledge rules may also utilize results of initial segmentation programs, including conventional segmentation techniques.
- Bright regions on magnetic resonance images (MRI) of athe brain have been associated with normal cognitive decline (Boone et al., 1992; Ylikoski et al., 1993 Mirsen et al., 1001) as well as the neurodegenerative disease in patients (Sullivan et al., 1990; Leifer et al., 1990; Filippi et al., 1995).
- Previous studies have used computerized methods, such as region of interest tracing (Kitaki et al., 1997) and threshold determination (Decarli et al., 1995; MacKay et al., 1996), to quantify WMHI.
- KGHID knowledge-guided hyperintensity detection
- MRI Data Acquisition A complete MRI of the brain including images in the axial, coronal, and sagittal planes, is acquired on a 1.5 Tesla Signa Advantage scanner (GE Medical Systems, Milwaukee, WI).
- the multispectral MRI dataset for each brain slice consists of proton density weighted (Pdw) (TR/TE e ff 550/27 msec), and T2 weighted (T2w) (TR/TE e ff3, 000/ 102.2 msec) acquisitions with an acquisition matrix of 256x192, a field of view of 22 cm, and a slice thickness of 5mm.
- Pdw proton density weighted
- T2w T2 weighted
- Analyzed images are obtained in the axial plan, where the first of nine consecutive slices are identified by the concurrent presence of the lateral ventricle, thalamus, basal ganglia (e.g., head of the caudate and lenticular nuclei), and the genu of the corpus calloswn (see, e.g., Figure 2).
- the selected volume generally just superior to the eye orbit through to the apex of the brain, allows maximum exposure to white matter while minimizing the effects of the extraneous tissue.
- the unsupervised masking of the intracranial tissue is empirically developed. It uses a logarithmic transformation of the gray values in the Pdw slice to effect a better separation of the dark background pixels from the brighter pixels of the imaged head. Once each pixel in the raw Pdw image is replaced by its natural logarithm, the image is thresholded by retaining only those pixels that constitute the top 30% of the intensities in this image, the rest being set to value zero. This step has shown to result in a reliable mask for the whole head.
- the intensities of the Pdw image are now shifted lower in value by effectively dividing the image intensities by a factor of 10 and then replacing each pixel by its natural logarithm. These steps serve to move the darkest pixel in the original Pdw image to zero so that the natural boundary between the skull and intracranial tissue can be more easily found with a simple threshold.
- the image is then thresholded such that the lowest 30% of the resulting intensities are set to zero.
- a region labeling algorithm is now applied to the result followed by an erosion operation ensuring that any imaged tissue connecting the skull and brain not removed during thresholding is severed. Once these residual tissue fragments are removed, the final mask is created by applying a dilation operation to the remaining pixels.
- the resulting dilation is converted to a binary image and applied as a multiplicative mask to the original multispectral image set (Pdw, T2w, Tlw). This process is repeated for each of the nine consecutive slices acquired for each patient, since the brain area may vary with the slice.
- the following optional final step in the unsupervised masking is executed if the previous steps did not result in an adequate brain segmentation.
- the success or failure is judged by a simple measurement of the masked brain area.
- the desired intracranial mask for example, comprise from about 65% to about 83% of the total image area in a 256x256 pixel image with a 22cm field of view. Any deviation from this range at the end of the first stage leads to several iterations with small incremental adjustments in threshold values until a mask with area less than 83% of the whole head for that slice is obtained.
- the masking technique described above provides an automatic and unsupervised means for providing masked images to the VGC module. Some parametric adjustments may be needed to apply this technique to other data sets, but this masking technique works reliably and without operator intervention on the data sets used in this laboratory.
- the VGC algorithm (Bensaid et al., 1996) is an enhanced version of FCM.
- FCM is an iterative clustering algorithm that finds a cluster center for each class (Bezdek et a., 1993).
- the optimization procedure in FCM inherently leads to classes that are approximately of the same size, and the resulting segmentation may not be anatomically meaningful (Velthuizen et al., 1995).
- Bensaid et al (1996) developed a refinement of FCM using a validity measure called VGC.
- the validity measure is essentially a fuzzy and normalized version of the Wilk's lambda statistic, which is used in linear discriminant analysis (Devijver and Kittler, 1982).
- VGC operates on the output of FCM and iteratively tries to improve the validity of the partitioning by splitting and merging clusters. At each iteration, the modified partition is evaluated and is retained if it constitutes an improvement. This method has been applied successfully to brain tumor images (Valthuizen, 1995). Because of its capability to detect small classes, VGC is a suitable clustering tool for application to MRI hyperintensities. In an embodiment, although not every embodiment, of the instant invention, I apply VGC to the masked image sets and search for four classes: cerebrospinal fluid (CSF: 1st class), white matter (WM; 2nd class), gray matter (GM; 3rd class), and a fourth class tissue type composed partially of WMHI.
- CSF cerebrospinal fluid
- WM white matter
- GM gray matter
- 3rd class gray matter
- VGC differentiates the four classes better than FCM, it is not effective in producing a reliable segmentation of the hyperintensities.
- the WMHI lesions vary in intensity over the image, which can be understood as partial volume effects, and some image nonuniformity due to the characteristics of the MRI coil (Wicks et al., 1992).
- the initial segmentation obtained by application of VGC to the masked data is refined by KGHID.
- KGHID The knowledge rules of KGHID are based on intensities combined with the spatial relationship of anatomical structures. MRI pixel intensities vary greatly from one patient to the next and can also exhibit significant interslice variation within the same patient (Wicks et al., 1992). Since the image data have a high signal-to-noise ratio, the distributions of intensities for each cluster approximate Gaussians (Bernstein et al., 1989; Henkelman, 1985; Parker et al., 1987; Gudbjartsson and Platz, 1995), as illustrated, by way of example, in Figure 1. Z-scores (Zar, 996) are then a meaningful measure of the intensity distributions.
- Ijr is the pixel intensity in the k tn spectral image, ° c and c Z ⁇ flp ⁇ ), and Z2(I lw) scores are used in specific rules described below.
- KGHID analyses begin with the identification of the lateral ventricles, caudate nuclei, lenticular nuclei, and the thalamus in the first image of each patient case. To aid the interpretation of the following discussion, relevant structures are identified in Fig. 2. Once identified, the locations of these structures are used to identify their possible presence in subsequent slices. Details of these localization procedures are discussed and illustrated below.
- Localization of the lateral ventricles begins by applying an iterative region growing operation to an initial seed of 1 pixel positioned at the midpoint of the intracranial tissue, calculated as the point in the image that represents the average x and y coordinates of the tissue. With each growing iteration, horizontal lines are generated, initially 50 pixels in length and stacked in single-pixel layer increments in the y direction. With the generation of each new line, the length is adjusted in the x direction, limited by the most lateral occurrences of CSF that lie medial to the WM comprising the internal capsules. The extension in the y direction continues until WM pixels are detected (the genu and splenium of the corpus callosum). This region- growing process creates a mask that roughly resembles the shape of the lateral ventricles.
- CSF pixels identified by the VGC segmentation are grouped into regions by a region-labeling operation (Russ, 1995) (also known as "blobbing").
- the ventricular CSF comprising the lateral ventricles is identified and isolated as those regions that have a nonempty intersection of these regions with the mask just generated, for example, as shown in Figure 2.
- Lateral ventricle (yellow).
- Relevant structure labels are as follows: 1) lateral ventricles (yellow); 2) thalamus; 3) basal ganglia (occurs bilaterally and includes the lenticular nuclei and head of the caudate nuclei); 4) genu of the corpus callosum; 5) internal capsule (occurs bilaterally); 6) external capsule (occurs bilaterally).
- the lateral ventricle is divided both vertically and horizontally, guided by its anatomical shape.
- the vertical division is defined as the midpoint between its maximum and minimum location with respect to the x-axis.
- the horizontal division occurs where its width is minimum.
- the upper half of the lateral ventricle is further divided into three distinct regions defined as superior, central, and inferior to the caudate curvatures, for example, as shown in Figure 3. These curvatures, occurring bilaterally with respect to the vertical division, appear as concave regions in the exterior wall of the lateral ventricle.
- the lateral ventricle divided as superior (green), central (red), and interior (yellow) to the caudate curvatures the nuclei are grouped with respect to their y- coordinates.
- Each subgroup of pixels then represents a horizontal line extending from the medial edge outward.
- Individual subgroups are analyzed by searching for the occurrence of WM in the vicinity of a (roughly vertical) line joining the maximum x-coordinates of the superior and inferior borders.
- a spline curve is fit to those subgroups in which WM is found, thus fixing the lateral edge of the caudate nuclei, for example, as shown in Figure 4.
- the superior border of each curvature is defined by the point where the x-coordinates of the pixels comprising the lateral ventricle reach their absolute maximum distances from the vertical dividing line.
- the inferior border is defined by the presence of the thalamostriate vein immediately adjacent to the wall of the lateral ventricles and by 4th class z-scores indicating pixels with bright appearance on the T2w and Pdw images.
- the central region of the caudate curvature is then defined as the area between its superior and inferior borders.
- the caudate nuclei are identified as segmented clusters of GM and 4th class tissue located immediately adjacent to and confined within the central region of the caudate curvatures defined above. While the medial edge and body of each nucleus are defined solely by their locations within their respective curvatures, the lateral edges occur where their tissues converge with the WM comprising the internal capsule. To find this edge, pixels comprising the nuclei are grouped with respect to their y-coordinates. Each subgroup of pixels then represents a horizontal line extending from the medial edge outward. Individual subgroups are analyzed by searching for the occurrence of WM in the vicinity of a roughly vertical line joining the maximum x-coordinates of the superior and inferior borders. To account for any missing expected WM within these subgroups, a spline curve is fit to those subgroups in which WM is found, thus fixing the lateral edge of the caudate nuclei, as shown, by way of example, in Figure 4.
- the lenticular nuclei are identified using knowledge of their bilateral locations, positioned generally lateral and slightly inferior to the caudate nuclei as viewed in the axial plane. These nuclei are clusters composed of GM and 4th class tissue that are bordered by WM pixels making up the internal and external capsules, as shown, by way of example, in Figure 4. The demarcation of the lenticular nuclei from the surrounding WM is accomplished in a manner that is similar to the method used to depict the lateral edge of the caudate nuclei. Lesions confined within the lenticular nuclei are identified as small clusters of hyperintense tissue with 4th class pixel z-scores indicating their bright appearance on the T2w and PDw images, as shown by way of example, in Figure 4.
- the thalamus Viewed in the axial plane, the thalamus is identified as segmented clusters of GM and 4th class tissue bilaterally positioned inferior to the lower edge of the caudate curvatures and immediately adjacent to the lateral ventricle. Clusters of GM and 4th class tissue, identified by VGC segmentation, are grouped by region labeling. The thalamus is identified as the intersection of these regions with the confined area of interest.
- KGHID proceeds to classify lesion tissue in the image beginning with the periventricular ring.
- These lesions are identified by iteratively applying a morphological dilation operator (Haralick et al., 1986) to the previously defined edge of the lateral ventricle. Each dilation generates a single pixel ring around the previous single layer.
- a morphological dilation operator Harmonic et al., 1986
- Each dilation generates a single pixel ring around the previous single layer.
- all pixels identified as caudate nuclei and thalamus are removed from the single pixel ring and the remaining pixels are judged for suitability as lesion tissue via their respective 4th class z- scores of the PDw and T2w images.
- KGHID identifies WMHI contained entirely within regions identified as WM.
- the effects of CFG confined within the sulci and the GM tissue immediately surrounding these pockets of CSF are removed.
- the image of the remaining intracranial tissue is subjected to an iterative morphological erosion operator are returned to the eroding image.
- the erosion process continues until all pixels in the final layer are identified as either WM by the original VGC segmentation, pixels comprising the border between GM and WM with low 4th class z- scores in the PDw and T2w images, or pixels making up the lateral borders of the lenticular nuclei
- Large clusters are divided, if required, into separate regions by excluding hypointense embedded pixels with low 4th class z-scores indicating WM on the PDw and T2W images. Pixel clusters are removed from further consideration as WMHI lesions if not surrounded primarily by WM identified either by VGC segmentation or pixels with z-scores indicating WM on the PDw and T2w images. In addition, pixel clusters representing perivascular spaces, with z-scores in the second class that possibly represent the hypointensities of CSF on the T2w image, are also removed.
- the remaining regions in the image are subjected to a geometrical shape analysis which judges their circularity by forming the quotient of the cluster area divided by the square of the cluster perimeter (Jahne, 1995; Castleman, 1996). This measure of circularity is insensitive to cluster size, and is a number ranging from 1/(4 ) for a perfect circle, becoming substantially smaller as the cluster becomes more elongated. Clusters deemed essentially linear in shape are rejected from further analysis. The prevailing clusters are then identified as tissue comprising WMHI lesions, as shown, by way of example, in Figure 6.
- Region of Interest 32% 81% 0.10 Region of interest 38% 84% 0.13 ⁇
- Table I presents the nine-slice average for sensitivity (Metz, 1978), specificity (Metz, 1978), and kappa (Cohen, 1960) statistics of the results generated by KGHID, region of interest, and thresh- holding methods as compared to ground truth determined by an experienced neuroradiologist (FRM).
- FRM experienced neuroradiologist
- the sensitivity of KGHID is at least twice as great as that of the other methods, thus ensuring greater confidence in its ability to make true positive detections.
- the relatively high specificity of KGHID ensures a low number of false positive detections.
- the kappa value shows that KGHID is in high agreement with the ground trust image (Landis and Koch, 1977).
- VGC VGC-to-vehicle
- Increasing the number of classification bins from four to five causes VGC to split the lesion tissue among three (or more) competing classes without a clear advantage to further analysis. For this reason, segmentation by VGC is limited to four tissue classes.
- KGHID uses the 3rd and 4th class tissue of VGC and encoded brain anatomy to extract hyperintense lesion tissues. The anatomical decision criteria are based upon extensive consultations with neuroradiologists and empirical refinement, thus contributing to a confident reclassification of lesion tissues from within these two rather broad categories.
- Tables II and III show the effect of KGHID on the VGC segmentation.
- Table II shows the distribution of the pixels in each of the VGC classes over the final KGHID assignments.
- the WMHI, preiventricular ring, and basal ganglia lesions make up the class of "lesions.” Since KGHID is designed to detect and accurately segment WMHI, the majority of pixel reassignments occur in VGC's 4th class and consequently the KGHID lesion class.
- Table III shows the distribution of VGC classes over the labeled anatomical structures. It is immediately apparent that the anatomical structures are very small (e.g., less than a percent of the pixels in an image) and are unlikely to be detected with a global segmentation method. Table III also shows that the composition of the structures in terms of VGC classes varies. KGHID takes advantage of local characteristics in the immediate neighborhood of the structures to differentiate tissues, rather than relying solely on MRI intensities. As indicated by the results above, KGHID demonstrates several advantages over other computerized techniques. To be useful, these techniques must exhibit reasonable sensitivity, specificity, and agreement with expert judgment. As shown in Table I, only KGHID demonstrates adequate agreement with the judgment of an expert radiologist.
- KGHID shows a clear advantage in the localization of brain anatomy and in precisely reclassifying the WMHI, preiventricular ring, and basal ganglia lesions.
- This method using hybrid clustering and knowledge based techniques, is operator independent and therefore removes the effects of human subjectivity inherit in other quantification methods.
- the results of this method are compared with the results of the manual computerized methods on 42 AD patients. These preliminary results suggest that the thresholding method underestimates hyperintense quantities, while the method of the present invention closely approximates region tracing in both hyperintense quantity and lesion identification.
- the sensitivity measure describes the fraction of patients with disease that are detected by the diagnostic test in question.
- the specificity measure describes the fraction of patients who are correctly identified as having no disease. 11.2. lnterrater Reliability Measurement Using the Kappa Statistic
- the kappa statistic (Cohen, 1960) is one such measure that determines the extent of agreement between two or more judges exceeding that which could be expected to occur purely by chance.
- the kappa statistic is used when two or more judges consider the same entity and express a judgment regarding that entity.
- WMHI White matter hyperintense
- AD Alzheimer's Disease
- WMHI white matter hyperintense
- DWML deep white matter lesions
- PVL periventricular lesions
- MRI magnetic resonance image
- KGHID Knowledge Guided Hyperintensity Detection
- KGHID to quantify WMHI lesion load on MRI (e.g., 90 images) of 10 patients diagnosed with probable AD.
- Magnetic resonance imaging is a noninvasive technique offering high resolution analysis of fundamental brain anatomy and/or pathology. Specific disease conditions are increasingly associated with volume changes in specific brain structures, and/or the presence of focal bright spots on MRI images (i.e., hyperintensities).
- This application introduces an objective, automatic and rapid computer algorithm for post-acquisition analysis of brain MRI studies; the Knowledge-Guided MRI Analysis Program. This computer algorithm automatically recognizes brain regions, computes volumes and identifies hyperintensities for each region.
- the output from the algorithm discriminated the two diagnostic categories with sensitivity and specificity averaging 80%. It is a feature of the instant invention to include all of the brain from apex to foramen magnum, and fine tune the detection of presently identified and additional brain regions (e.g., cerebellum, brain stem). It is another feature of the instant invention to provide analysis of MRI studies, with an emphasis on dementias, and identify associations between regional volumes and/or hyperintensity burdens with specific diagnostic categories. From these associations, discriminant function equations are optionally developed to aid in differential diagnosis. The program optionally provides a rapid initial screen of all brain MRI studies to highlight regions that are outside of the age- and/or gender-adjusted normal range of values, thereby, suggesting areas to be scrutinized first by the neuroradiologist.
- a goal of the instant invention includes developing KGMAP into a general analysis program for conventional MRI studies.
- the output routinely provides volumetric data for a number of brain regions and indicate the locations of hyperintensities. This output is optionally provided with normative data much like routine blood chemistry results are presently displayed.
- a computer program is disclosed herein which identifies and regionally localizes hyperintensities (HYPs) and measures the volumes of multiple brain regions on human brain magnetic resonance imaging studies.
- This program identifies brain regions and detects HYPs without global thresholding, archival maps or user interaction. It is knowledge-guided, building on encoded neuroanatomical relationships and MRI characteristics to classify MRI pixels into specific brain regions. It also integrates MRI information from proton density, T2 and TI scan sequences in formulating a final analysis.
- the Knowledge Guided MRI Analysis Program (“KGMAP") is compiled in IDL and can be run on both DOS and UNIX platforms and presently runs in about 3 minutes for a typical MRI study.
- NIR plans to establish age and gender adjusted norms for brain region volume and HYP data for comparison with neural disorders.
- discriminant function information such as that proposed here for measuring neural changes consistent with AD dements as distinct from from Non -AD dements
- the output from the KGMAP algorithm optionally aids the physician(s) in assigning diagnostic categories. Pilot data are presented indicating that the KGMAP is equal to or better than other automated or semi-automated methods of identifying HYPs in MRI studies. Data are also presented summarizing the brain region volume data from 50 cases.
- An optional feature of the instant invention includes application thereof beyond the axial planes (13) to include the entire brain.
- Such a modified program optionally adds volumetric measurement of the hypothalamus, brain stem and cerebellum, as well as extend the analysis of total brain, gray mater, white matter and CSF spaces.
- Another optional feature of the instant invention is to identify disorders for which diagnosis can be facilitated with the application of the algorithm. Examples include additional AD and Non- AD dementia cases.
- the KGMAP includes sensitivity and specificity values of 80% or greater in distinguishing these two groups.
- a series of other brain MRI studies is obtained from local MRI centers and analyzed for regional HYP density and volume changes using the KGMAP algorithm. Association between regional and HYP volumes, post-scan diagnosis and patient history is optionally evaluated by means comparisons followed by logistic regression (as for the pilot data).
- Another optional feature of the instant invention includes collection of MRI studies for neurologically normal individuals and quantify volumes and HYPs.
- this population is diverse so that gender, age and ethnicity differences in region and HYP volumes can be estimated.
- a series of normative values are optionally established. This optionally permits KGMAP to provide output for each MRI analyzed indicating the normal range of values (e.g., adjusted for age, gender, etc., as necessary), and highlighting those values that lie outside the normal range.
- the predictor algorithms are optionally applied to indicate those cases with values consistent with specific disease categories.
- hippocampal volume is also reported to decline in other clinical conditions. These include epilepsy (Webb et al, 1999) schizophrenia (Wright et al, 2000) and alcoholism (Agartz et al, 1999) or traumatic brain injury (Bigler et al, 1997). Similar arguments can be made regarding the size of the lateral ventricles, and other CSF spaces used as indices of brain atrophy. Still, measurement of regional volumes of temporal lobe structures does have significant diagnostic utility, classifying normals from cognitively impaired individuals correctly 75% of the time (Convit et al, 1997). Thus, while hippocampal volumes may have good sensitivity in detecting dementias, by itself, specificity is optionally restricted given the other clinical conditions where reduced hippocampal volume is reported.
- Measurement of brain volumes may also have value in the diagnosis of other neurological and psychiatric disorders.
- the progression of multiple sclerosis benefits from volumetric studies of lesions, and may benefit further from the analysis of TI weighted images revealing hypointense "black holes" (Truyen et al, 1996; Adams et al, 1999; Dastidar et al, 1999).
- Changes in regional brain volumes via MRI analysis have also been reported in obsessive- compulsive disorder (Szeszko et al, 1999), bipolar disorder (Strakowski et al, 1999) and psychotics using neuroleptic medications (Gur et al, 1998).
- Bright regions on magnetic resonance images of the brain known as HYPs or unidentified bright objects have been associated with neurodegenerative disease (Kertesz et al 1988; Herholdz et al, 1990; Meguro et al, 1990; Leifer et al, 1990; Filippi et al, 1995) and to a lesser extent with normal age-associated cognitive decline (Boone et al, 1992; Ylikoski et al, 1993; Mirsen et al, 1991; Smith et al, 2000).
- HYPs may be associated with a decline of cognitive performance especially with respect to performance on timed attentional and visuopractical tasks (Rao et al, 1996).
- HYPs may be helpful in differentiating degenerative dementias (such as Alzheimer's) from dementias with primarily vascular causes (multi-infarct dementia, Binswanger's disease; Barclay et al, 1992; Rao et al, 1996; Libon et al, 1998).
- HYP involvement in MS typically consists of rather large diffuse white matter lesions as opposed to the small punctate lesions remarked upon by dementia researchers.
- a review of neuroimaging results suggests that although all confirmed MS patients exhibit cerebral white matter lesions at autopsy, and roughly 95% of living patients exhibit HYPs on T2 weighted MRI scans, only 40-60 % of these patients display measurable cognitive faults, covering a rather broad array of cognitive symptoms.
- Visual interpretation and categorical assignment can provide insight to lesion load within distinct anatomical regions of the brain, but lacks objectivity or the ratio scale measurement required by parametric statistics.
- CERAD investigators reported that visual inte ⁇ retation and categorical assignment of white matter lesions is highly subjective (Davis et al, 1992). In spite of efforts to standardize MRI imaging of CERAD subjects and the utilization of a uniform rating scale, the frequency, locations and severity of white matter lesions resulted in an unsatisfactory intraclass correlation.
- the tracing ROI technique can provide information as to the lesion load within a distinct brain region but introduces operator subjectivity in lesion identification and border detection.
- the threshold procedure in which HYPs are defined as all pixels above a given intensity threshold, provides measurement consistency. Yet, in many cases, this approach requires adjustments by the operator for different images and ignores the inherent differences in pixel intensities in each brain region. For example, HYP lesions described as 'punctate' often consist of pixel intensities that lie within a very close range, whereas 'diffuse' regions often host a considerable variety of pixel intensities. Even the landmark study of Guttman et al (1998) using a sophisticated thresholding approach to specify tissue types remarked that it is conceivable some HYPs are misclassified as gray matter in their protocol.
- KGMAP Using the instant KGMAP algorithm to detect HYPs optionally aids in resolving the confusion regarding the possible associations between HYP abnormalities and specific disease states.
- KGMAP is automatic; no "blind observers" are required.
- KGMAP for example, uses a local thresholding operation comparing HYPs to the pixel intensity of the structure in which it resides.
- KGMAP is quantitative avoiding the subjectivity of rating scales.
- KGMAP provides regional information regarding the HYPs, which for some disorders may be more important than the total HYP burden (Moore et al, 1996).
- the KGMAP program optionally significantly aids in establishing a reliable method for identification and quantification of these MRI abnormalities.
- the instant invention is applied to brain MRIs obtained from local centers to begin ascertaining specific diagnostic capabilities of the technique.
- my pilot data indicates possible utility in discriminating AD from Non-AD dements.
- application of the instant invention optionally includes autopsy (where feasible) to confirm the clinical diagnosis pathologically.
- Table 4 indicates the sensitivity, specificity and kappa scores of the KGMAP analysis of HYPs compared to two computerized methods commonly used to quantify HYPs on brain MRI images; 1) pixel thresholding and 2) region of interest ("ROI") manual tracing.
- Pixel Thresholding The quantification of HYPs by a pixel thresholding method (described by DeCarli, 1995) is duplicated with the exception that CSF is segmented using the automated routine of the present invention instead of using an independent thresholding procedure.
- HYP lesions are identified by a manual tracing method performed by two radiology residents.
- a ground truth image is constructed with all real lesions, including some not identified by any of the computerized methods.
- lesions identified by one or more computer-based methods are further evaluated for size and border detection accuracy.
- the results of each method are analyzed against ground truth to determine the quantity of pixels identified as true positive (e.g., pixels identified as lesions that concur with ground truth) , false positive (e.g., pixels identified as lesions that did not concur with ground truth lesion) and false negative (e.g., pixels not identified as lesion but are identified as lesion on the ground truth image).
- true positive e.g., pixels identified as lesions that concur with ground truth
- false positive e.g., pixels identified as lesions that did not concur with ground truth lesion
- false negative e.g., pixels not identified as lesion but are identified as lesion on the ground truth image.
- the sensitivity, specificity and Kappa rates are calculated for each method.
- the values reported in Table 1 reflect the average values from 10 MRI studies (e.g., 9 images per subject) analyzed by each method.
- the results in Table 4 indicate that the KGMAP is very accurate at detecting HYPs with both good sensitivity and good specificity.
- the overall kappa score also validates use of this program to identify HYPs.
- the results of global pixel thresholding have good specificity (e.g., few false positives), but are poor at detecting HYPs that are on darker backgrounds (e.g., highlighting the advantages of local thresholding), reflected as low sensitivity scores.
- the two neuroradiology residents performing the ROI tracing tended to be better in terms of sensitivity than the global thresholding approach, but still are not as accurate as the KGMAP.
- the kappa values for thresholding and ROI methods did not inspire confidence as an adjunct for detailed analysis of MRI images by a neuroradiologist.
- the high degree of concordance between the KGMAP and ground truth indicate this software is a rapid, automated and quantitative method for accurately identifying HYPs in MRI images.
- HYPs could be useful for differential diagnosis.
- prior studies done without the KGMAP might not be sufficiently sensitive, quantitative or regionally specific enough to detect associations with diagnostic categories.
- the data available did not strongly support use of HYPs alone in differential diagnosis as shown above, for example in the MRI Hyperintensities in Aging and Disease section.
- dementia of the Alzheimer type, low hippocampal volume is quantitative measure obtained from MRIs that showed association with the disease, for example as discussed in the MRI volume measurements section above.
- the program is optionally modified to include volumes of the structures studied, and extend the software to conservatively identify hippocampal tissue.
- S.D is the sample standard deviation.
- CV. is the coefficient of variation; (S.D./mean)*100
- Table 6 Volumetric and Hyperintensity Data for AD and NON-AD Patients (Means + SEM). Values are presented in mm ⁇ rather than cm ⁇ to present the HYP data as whole numbers. *** p ⁇ .001, AD versus Non-AD
- a MANOVA is computed on the outcomes with diagnostic category as the between subjects factor and total brain volume, with CSF, as a covariate, or separately with age as a covariate. However, because the results with or without the brain volume or patient age corrections are the same, the uncorrected values are shown in the table.
- the volumes for the AD group are lower than the Non-AD group.
- the putamen HYP volumes are greater in the AD cases.
- the algorithm can be successfully applied to a series of standard MRI scans.
- the HYP detection capability is equal to or superior to other computer-based methods for measuring HYP volumes, particularly regarding sensitivity.
- the algorithm in its present form can discriminate between AD and Non-AD cases based upon volume and HYP information.
- these data are sufficiently robust that the effects remained when corrected for age or brain volume in covariate analyses.
- the association of apolipoprotein E genotype with the volume measures is also examined. There is no association of apolipoprotein E allele frequency on any of the parameters measured.
- the algorithm provides the basis with which to predict that ultimately the algorithm can provide sensitivities and specificities of 80% or greater in categorizing dementia into AD and Non-AD groups, especially when all discriminating variables are combined into a discriminant function equation.
- the KGMAP optionally aids in identifying other discrete dementia categories such as Lewy- body dementias, Pick's disease or AIDS dementias.
- disease categories other than dementias for MRI volume features which discriminate them from other related disorders, and/or normal cases of the same age and gender are evaluated.
- the present algorithm, KGMAP recognizes HYPs on brain MRIs and quantifies and localizes them without requiring user interaction or global thresholding.
- the present program runs unassisted and completes the analysis of a total MRI study in, for example, about 3 minutes.
- the KGMAP measures HYP area within the periventricular ring, the thalamus, the lenticular nuclei (i.e., globus pallidus and/or putamen), the caudate nuclei, and the subcortical white matter. It measures the volumes of all of these structure, plus CSF spaces, gray matter and hippocampus. Extension of the algorithm will include determination of volumes in the hypothalamus, cerebellum and brainstem by including all 5 mm axial planes containing brain tissue in the skull. This, for example, involves generating code to identify the added regions in the corresponding MRI images.
- KGMAP Much of the logic written into KGMAP in its current version (nearly 10,000 lines of code) is optionally used to identify and analyze the additional regions.
- KGMAP begins its analysis in the MRI image just superior to the eye orbit in the axial plane.
- KGMAP uses this image to initially identify the regions that are analyzed first (e.g., caudate and lenticular nuclei, lateral ventricle and corresponding periventricular ring, and/or thalamus).
- the initial identification of these regions relies on the encoded knowledge of brain anatomy written into the logic of KGMAP. That is, localizing any one region relies on its relative location to another region.
- the location of the caudate nuclei aids in localizing the lenticular nuclei, in that the lenticular nuclei lie just slightly inferior and lateral to the caudate nuclei when viewed in the axial plane.
- KGMAP uses the pixel locations of these identified regions to aid in locating these regions on the subsequent image planes.
- this process is repeated in each subsequent slice to the apex of the brain.
- KGMAP will be modified to identify regions of interest that exist below the plane just superior to the eye orbit to recognize the remainder of brain tissue through to the brain stem.
- the location of the caudate nuclei aids in the localization of the hippocampus in that the hippocampus lies inferior to the caudate head as it enters the temporal lobe.
- the hippocampus is optionally localized by its most medial position relative to the white matter of the temporal lobe and its most superior border (for instance, as viewed in the axial plane) delineated by the inferior horn of the lateral ventricle.
- the KGMAP algorithm can be performed using films of the MRIs which are scanned into the computer, but is more accurate if the digitized MRI files are used.
- each participating MRI center is optionally provided with a CD-writer that can be adapted to the computer retaining the electronic MRI data. For instance, each evening, the MRI technologist can simply identify the files of those patients agreeing to participate in the study, delete any identifying information and copy the files onto CD.
- Each MRI file contains roughly 35 MB of information, permitting up to 18 images to be placed on a single CD. At a cost of $2 per CD, this is a minor expense for the study, and considerably less tedious than redigitizing film versions of the individual scans.
- Electronic capture also facilitates anonymity by permitting identifying information to be removed from the data before it is conveyed to an analysis center, such as Neurolmaging Research Inc.
- a complete MRI of the brain is acquired on a 1.5 Tesla Signa Advantage scanner (GE Medical Systems, Milwaukee, WI).
- the multispectral MRI dataset for each brain slice consists of proton density weighted (PDw) (TR/TE e ff 3000/17 msec), TI weighted (Tlw) (TR/TE e ff 550/27 msec) and T2 weighted (T2w) (TR/TE e f 3000/102.2 msec) acquisitions with an acquisition matrix of 256x192, a field of view of 22 cm and a slice thickness of 5mm.
- PDw proton density weighted
- Tlw TR/TE e ff 550/27 msec
- T2 weighted TR/TE e f 3000/102.2 msec
- images obtained in the axial plane are analyzed, where, for example, the first of nine consecutive slices are identified by the concurrent presence of the lateral ventricle, thalamus, basal ganglia (head of the caudate and lenticular nuclei) and the genu of the co ⁇ us callosum.
- the selected volume for example, just superior to the eye orbit through to the apex of the brain, allows maximum exposure to white matter while minimizing the effects of extraneous tissue.
- the brain area is automatically identified from the images.
- This unsupervised masking of the intracranial tissue is empirically developed. It uses a logarithmic transformation of the gray values in the PDw slice to effect a better separation of the dark background pixels from the brighter pixels of the imaged head. Once each pixel in the raw PDw image is replaced by its natural logarithm, the image is thresholded by retaining only those pixels that constitute the top 30% of the intensities in this image, the rest being set to value zero. This step has shown to result in a reliable mask for the whole head.
- the intensities of the PDw image are now shifted lower in value by effectively dividing the image intensities by a factor of 10 and then replacing each pixel by its natural logarithm. These steps serve to move the darkest pixels in the original PDw image to zero so that the natural boundary between the skull and intracranial tissue can be more easily found with a simple threshold.
- the image is then thresholded such that the lowest 30% of the resulting intensities are set to zero.
- a region labeling algorithm is now applied to the result followed by an erosion operation ensuring that any imaged tissue connecting the skull and brain not removed during thresholding will be severed. Once these residual tissue fragments are removed, the final mask is created by applying a dilation operation to the remaining pixels.
- the resulting dilation is converted to a binary image and applied as a multiplicative mask to the original multispectral image set (PDw,T2w,Tlw).
- This process is repeated for each of the nine consecutive slices acquired for each patient, since the brain area will vary with the slice.
- the following optional final step in the unsupervised masking is executed if the previous steps did not result in an adequate brain segmentation.
- the success or failure is judged by a simple measurement of the masked brain area.
- the desired intracranial mask will comprise from about 65% to about 83% of the total image area in a 256x256 pixel image with a 22 cm field of view. Any deviation from this range at the end of the first stage leads to several iterations with small incremental adjustments in threshold values until a mask with area less than about 83% of the whole head for that slice is obtained.
- An optional design goal of the masking technique described above includes a completely automatic and unsupervised procedure to provide masked images to, for example, the VGC module.
- VGC fuzzy c-means
- FCM fuzzy c-means
- FCM is an iterative clustering algorithm that finds a cluster center for each class (Bezdek et al, 1993).
- the optimization procedure in FCM inherently leads to classes that are approximately of the same size, and the resulting segmentation may not be anatomically meaningful (Velthuizen et al, 1995).
- Bensaid et al. (1996) developed a refinement of FCM using a validity measure called Validity Guided Clustering ("VGC”).
- VCC Validity Guided Clustering
- the validity measure is essentially a fuzzy and normalized version of the Wilks's lambda statistic, which is used for example in linear discriminant analysis (Devijver and Kittler, 1982).
- VGC operates on the output of FCM and iteratively tries to improve the validity of the partitioning by splitting and merging clusters. At each iteration, the modified partition is evaluated and is retained if it constitutes an improvement. This method has been applied successfully to brain tumor images (Velthuizen, 1995), and because of its capability to detect small classes, VGC is a suitable clustering tool for application to MRI hyperintensities.
- VGC is applied to the masked image sets and four classes are searched: cerebrospinal fluid (CSF) (1 st class), white matter (WM) (2 fi d class), gray matter (GM) (3 r ⁇ class) and a fourth class tissue type comprised partially of WM HYP.
- CSF cerebrospinal fluid
- WM white matter
- GM gray matter
- a fourth class tissue type comprised partially of WM HYP.
- KGMAP The knowledge rules of KGMAP are based on intensities combined with the spatial relationship of anatomical structures.
- the ) ⁇ > an( I Z2C-77 w ) scores are optionally used in specific rules described below.
- KGMAP analyses begin with the identification of the lateral ventricles, caudate nuclei, lenticular nuclei, and the thalamus in the first image of each patient case. The locations of these structures, once identified, are used to identify their possible presence in subsequent slices. Details of these localization procedures are discussed below.
- Lateral Ventricles Localization of the lateral ventricles begins by applying an iterative region growing operation to an initial seed of 1 pixel positioned at the midpoint of the intracranial tissue, calculated as the point in the image that represents the average x and y coordinates of the tissue. With each growth iteration, horizontal lines are generated, initially 50 pixels in length and stacked in single pixel layer increments in the ⁇ y direction. With the generation of each new line, the length is adjusted in the ⁇ x direction, limited by the most lateral occurrences of CSF that lie medial to the white matter comprising the internal capsules. The extension in the ⁇ y direction continues until white matter pixels are detected (the genu and splenium of the co ⁇ us callosum).
- CSF pixels identified by the VGC segmentation are grouped into regions by a region labeling operation (also known as "blobbing"; Russ, 1995).
- the ventricular CSF comprising the lateral ventricles is identified and isolated as those regions that have a non-empty intersection of these regions with the mask just generated.
- the lateral ventricle is divided both vertically and horizontally, guided by its anatomical shape.
- the vertical division is defined as the midpoint between its maximum and minimum location with respect to the x-axis.
- the horizontal division occurs where its width is minimum.
- the upper half of the lateral ventricle is further divided into three distinct regions defined as superior, central, and inferior to the caudate curvatures, as shown, by way of example, in Figure 3. These curvatures, occurring bilaterally with respect to the vertical division, appear as concave regions in the exterior wall of the lateral ventricle.
- the superior border of each curvature is defined by the point where the x-coordinates of the pixels comprising the lateral ventricle reach their absolute maximum distances from the vertical dividing line.
- the inferior border is defined by the presence of the thalamostriate vein immediately adjacent to the wall of the lateral ventricles and by 4 tn class z-scores indicating pixels with bright appearance on the T2w and PDw images.
- the central region of the caudate curvature is then defined as the area between its superior and inferior borders.
- the caudate nuclei are identified as segmented clusters of GM and 4 m class tissue located immediately adjacent to and confined within the central region of the caudate curvatures defined above. While the medial edge and body of each nucleus are defined solely by their locations within their respective curvatures, the lateral edges occur where their tissues converge with the white matter comprising the internal capsule. To find this edge, pixels comprising the nuclei are grouped with respect to their y-coordinates. Each subgroup of pixels then represents a horizontal line extending from the medial edge outward. Individual subgroups are analyzed by searching for the occurrence of white matter in the vicinity of a (roughly vertical) line joining the maximum x- coordinates of the superior and inferior borders. To account for any missing expected WM within these subgroups, a spline curve is fit to those subgroups in which WM is found, thus fixing the lateral edge of the caudate nuclei .
- Lenticular Nuclei (Putamen and Globus Pallidus)
- the lenticular nuclei are identified using knowledge of their bilateral locations, positioned generally lateral and slightly inferior to the caudate nuclei as viewed in the axial plane. These nuclei are clusters composed of GM and 4 m class tissue that are bordered by WM pixels making up the internal and external capsules. The demarcation of the lenticular nuclei from the surrounding WM is accomplished in a manner that is similar to the method used to depict the lateral edge of the caudate nuclei. Lesions confined within the lenticular nuclei are identified as small clusters of hyperintense tissue with 4 m class pixel z-scores indicating their bright appearance on the T2w and PDw images
- Thalamus Viewed in the axial plane, the thalamus is identified as segmented clusters of GM and 4th class tissue bilaterally positioned inferior to the lower edge of the caudate curvatures and immediately adjacent to the lateral ventricle. Clusters of GM and 4 tn class tissue, identified by VGC segmentation are grouped by region labeling. The thalamus is identified as the intersection of these regions with the confined area of interest.
- Hippocampus The hippocampus is initially identified by KGMAP on the image in which temporal horn and body of the lateral ventricle appear to separate in the axial plane. This generally, but not necessarily, occurs within 2 images below the first image of analyzed by KGMAP. As the analyzes of KGMAP works downward from the first image into the temporal lobe of the brain, the tissue comprising the hippocampus is initially identified as that tissue confined within the space separating the temporal horn and body of the lateral ventricle. In the axial plane, its superior border lies inferior to the remaining body of the lateral ventricle, while its lateral and inferior borders lie medial and superior to the temporal horn of the lateral ventricle.
- KGMAP Periventricular Ring: With the delineation of ventricles and nuclei, KGMAP proceeds to classify lesion tissue in the image beginning with the periventricular ring. These lesions are identified by iteratively applying a mo ⁇ hological dilation operator (Haralick et al, 1987) to the previously defined exterior edge of the lateral ventricle. Each dilation generates a single pixel ring around the previous single layer.
- KGMAP identifies white matter hyperintense lesions ("WM HYP") contained entirely within regions identified as WM.
- WM HYP white matter hyperintense lesions
- the resultant clusters of pixels comprised primarily of 4 tn class tissue, are individually analyzed via 4 tn class z-scores derived from the PDw and T2w images to determine their eligibility as WM HYP lesions.
- Large clusters e.g., 10 or more pixels
- Pixel clusters are removed from further consideration as WM HYP lesion if not surrounded primarily by WM identified either by VGC segmentation or pixels with z-scores indicating WM on the PDw and T2w images.
- pixel clusters representing perivascular spaces, with z-scores in the second class that possibly represent the hypointensities of CSF on the Tlw image are also removed.
- the remaining regions in the image are subjected to a geometrical shape analysis which judges their circularity by forming the quotient of the cluster area divided by the square of the cluster perimeter (Jahne, 1995; Castleman, 1996). This measure of circularity is insensitive to cluster size, and is a number ranging from 1/(4 ) for a perfect circle, becoming substantially smaller as the cluster becomes more elongated. Clusters deemed essentially linear in shape are rejected from further analysis. The prevailing clusters are then identified as tissue comprising WM HYP lesions.
- Total Intracranial Volume This is the volume of all pixels remaining after the skull is removed from the image. At present, the values I have, refer only to the volumes derived from the 15 most dorsal 5 mm axial planes.
- Total Brain This is the volume of brain tissue (e.g., white and/or gray matter) after all CSF pixels are subtracted.
- Sulcal CSF This is the volume of pixels identified as CSF after the lateral ventricle CSF pixels are subtracted.
- Volumetric measures are calculated as the total number of pixels comprising a brain region multiplied by the slice thickness and by the squared value of the field of view divided by the image size.
- the data analysis is optionally performed essentially as shown above.
- Optional changes include the addition of regional volume information, which soon will encompass the entire cranial vault, and the inclusion of several covariates in the analyses such as age, gender, race education and genotype as these data are available with the cases.
- An analytic strategy is, for example, completed in two phases.
- the first phase optionally examines whether there are significant diagnostic group (e.g., AD and Non-AD) differences in terms of the lesion density or regional volumes in the circumscribed brain areas. This is optionally accomplished in two ways.
- a MANOVA comparing regional or HYP volumes in the various brain areas across the diagnostic groups is optionally computed.
- descriptive analyses of the MRI images optionally involve discriminant function analysis.
- two groups are optionally chosen to compare (e.g., AD and Non-AD) and attempt to develop a discriminant function equation.
- the first step optionally include age, gender, and/or years of education in an attempt to control for the potential confounding effects of these relationships.
- those brain structures shown to exhibit significant group differences in the mean-level analysis is optionally entered in a step-wise fashion.
- Subjects optionally include a consecutive series of patients receiving MRI studies at local MRI centers. Criteria for inclusion is, for example, age (e.g., over 30), consent to participate, and/or a clear post-MRI study clinical diagnosis. Thus, for example, women and minorities are optionally included. As subjects are optionally medical patients, at least some have some type of illness causing their presentation at the MRI clinic.
- CERAD Cerebral Deformation Detection of a standardized MRI evaluation of Alzheimer's disease.
- Fazekas F., Kapeller, P., Schmidt, R., Offenbacher, H., Payer, F., & Fazekas, G. (1996) The relation of cerebral magnetic resonance signal hyperintensities to Alzheimer's disease. J Neurol. Sci.
- an automated, or computer-implemented, method of identifying suspected lesions in a brain is provided, by way of illustration, in Figure 9.
- Step 9 an automated, or computer-implemented, method of identifying suspected lesions in a brain is provided, by way of illustration, in Figure 9.
- a processor or scanner provides a magnetic resonance image (MRI) of a patient's head, including a plurality of slices of the patient's head, which MRI comprises a multispectral data set that can be displayed as an image of varying pixel or voxel intensities.
- the processor identifies a brain area within each slice to provide a plurality of masked images of intracranial tissue.
- Step S 120 the processor applies a segmentation technique to at least one of the masked images to classify the varying pixel intensities into separate groupings, which potentially correspond to different tissue types.
- the processor refines the initial segmentation into the separate groupings of at least the first masked image obtained from Step SI 20 using one or more knowledge rules that combine pixel intensities with spatial relationships of anatomical structures to locate one or more anatomical regions of the brain.
- the processor identifies, if present, the one or more anatomical regions of the brain located in Step S130 in other masked images obtained from Step SI 20.
- Step SI 50 the processor further refines the resulting knowledge rule-refined images from Steps SI 30 and SI 40 to locate suspected lesions in the brain.
- a method for determining volumetric measurements and/or detecting lesion tissue is provided.
- Such determinations and/or detections occur, for instance, in subcortical regions of a brain.
- subcortical regions include, but are not limited to, the hippocampus, the thalamus, the hypothalamus, the caudate nuclei, and the lenticular nuclei, which in turn includes the globus pallidus and putamen.
- the determinations and/or detections according to the instant invention are equally applicable to non-subcortical regions of the brain including, but not limited to, the brain stem, and/or the cerebellum.
- CSF cerebrospinal fluid
- Step S200 a user performs a series of standard, consecutive imaging scans of a patient's head.
- the standard imaging scans include standard magnetic resonance imaging ("MRI") scans.
- MRI scans include standard, traditional MRI scans and/or standard, functional MRI scans.
- the scans are axial, coronal, or sagittal.
- the scans are of standard thickness, e.g., 3mm or 5mm.
- 3mm or 5mm One of ordinary skill in the art will readily recognize that alternative thicknesses are acceptable.
- a series of thicker scans will yield more coarse volumetric measurements than a series of thinner scans.
- each scan is conceptually divided into plurality of sections, for example, quarters. If the scans are so divided, one of ordinary skill in the art will readily appreciate that the following steps apply for each section.
- Each MRI scan in the series includes one or more standard modalities or standard data acquisitions.
- standard data acquisitions include a proton density weighted (“PDw”) data acquisition, a spin-lattice relaxation time (“Tls”) data acquisition, and/or a spin-spin relaxation time (“T2w”) data acquisition.
- PDw proton density weighted
- Tls spin-lattice relaxation time
- T2w spin-spin relaxation time
- Alternate standard acquisitions are acceptable.
- One of ordinary skill in the art will readily appreciate the trade-offs between the myriad factors associated with MRI brain scans including, but not limited to, the type of image to acquire, cost, time, resolution, slice thickness, distance between slices, and signal-to-noise ratios.
- a standard processor effecting the instant invention detects a location of a lateral ventricle, if present, at least in part, by detecting CSF in the lateral ventricle in a first imaging scan.
- the processor sums all of the voxels within and/or including a border of the lateral ventricle.
- Step S220 based, at least in part, on the detected location of the lateral ventricle, the processor detects a location in the first imaging scan a caudate nuclei, if present, which curves into the lateral ventricle.
- the processor optionally determines the location of a superior border of a curvature of the caudate nuclei, at least in part, from one or more extreme minimum/maximum points on the lateral ventricle.
- the processor optionally determines a location of an inferior border of the caudate nuclei by following the curvature thereof.
- the processor optionally determines a location of a exterior border of the caudate nuclei, at least in part, by detecting an internal capsule, which includes very dark, white matter.
- the processor optionally determines the location of the exterior border of the caudate nuclei, at least in part, from a standard spline fit or inte ⁇ olated curve of the known borders. For example, the processor moves the standard spline fit until the processor detects deep white matter, thereby generating a smart "best fit" border for the caudate nuclei.
- the processor sums all of the voxels within and/or including a border of the caudate nuclei.
- Step S230 the processor detects a location in the first imaging scan of a thalamus, if present, based, at least in part, on the locations of the detected lateral ventricle and the detected caudate nuclei.
- the processor identifies a superior border of the thalamus, at least in part, from the inferior border of the caudate nuclei and identifies a medial border of the thalamus, at least in part, from a lateral edge of the lateral ventricle.
- the processor sums all of the voxels within and/or including a border of the thalamus.
- the processor detects a location in the first imaging scan of a hippocampus, if present.
- the processor identifies the hippocampus as being inferior to the thalamus and proximate to the lateral ventricle.
- the processor sums all of the voxels within and/or including a border of the hippocampus.
- the processor detects a location in the first imaging scan of a lenticular nuclei based, at least in part, from the locations of the detected caudate nuclei and the detected thalamus. For instance, the processor identifies the lenticular nuclei as being inferior to the superior border of the caudate nuclei, superior to the inferior border of the thalamus, and/or medial to an insolor cortex, which is entrapping CSF. Optionally, the processor further identifies a border of the lenticular nuclei from a standard spline fit.
- the processor moves the spline fit defining a presumptive border of the lenticular nuclei until it reaches an external capsule of the brain.
- the processor sums all of the voxels within and/or including a border of the lenticular nuclei.
- Step S260 the processor detects a location in the first imaging scan of one or more lesions, if present.
- the processor generates from the first imaging scan a corresponding virtual or actual image free of the detected subcortical regions, CSF in a sulci, and/or gray matter, for example, around the sulci, thereby retaining or defining a border of white matter in the generated image.
- the processor optionally generates an intensity gradient on the remaining generated image to identify possible or suspected lesions, if present. For instance, the processor generates the intensity gradient using standard Z-scores.
- the processor repeats Steps 210 through 250.
- the processor maintains a running sum of voxels for each detected cerebral region to a previous sum obtained from the preceding scan(s).
- a volumetric measurement of each detected cerebral region at least in part, by multiplying the sum of voxels for each region by a volume of a single voxel.
- total intracranial volume, ventricular volume, total CSF volume, sulcal CSF volume, white matter volume, and/or gray matter volume are measured using the instant invention.
- Step S270 the processor uses the detected locations of each brain region in a preceding imaging scan, at least in part, as an approximate or suggested search area for the same brain region in the subsequent imaging scan. For example, the processor performs a standard blobbing or voxel-clustering technique to identify such approximate or suggested search areas for a cerebral region. The processor, for example, then dilates and/or constricts the suggested search area of each detected cerebral region to detect volume changes from a preceding imaging scan to a subsequent imaging scan.
- at least a portion of the above-mentioned method, and preferably the entire method is run on a plurality of occasions separated in time to monitor disease progression and/or success of therapy.
- the separations in time include regular or irregular time intervals.
- volumetric changes in regions of interest detected over time may show atrophy of a region in a patient.
- atrophy of the hippocampus as evidenced by output of the instant invention, may be indicative of progression of Alzheimer's disease.
- Figure 11 is an illustration of a main central processing unit for implementing the computer processing in accordance with a computer implemented embodiment of the present invention.
- the procedures described herein are presented in terms of program procedures executed on, for example, a computer or network of computers.
- a computer system designated by reference numeral 900 has a computer 902 having disk drives 904 and 906.
- Disk drive indications 904 and 906 are merely symbolic of a number of disk drives which might be accommodated by the computer system. Typically, these would include a floppy disk drive 904, a hard disk drive (not shown externally) and a CD ROM indicated by slot 906.
- the number and type of drives varies, typically with different computer configurations.
- Disk drives 904 and 906 are in fact optional, and for space considerations, are easily omitted from the computer system used in conjunction with the production process/apparatus described herein.
- the computer system also has an optional display 908 upon which information is displayed.
- a keyboard 910 and a mouse 902 are provided as input devices to interface with the central processing unit 902. Then again, for enhanced portability, the keyboard 910 is either a limited function keyboard or omitted in its entirety.
- mouse 912 optionally is a touch pad control device, or a track ball device, or even omitted in its entirety as well.
- the computer system also optionally includes at least one infrared transmitter and/or infrared received for either transmitting and/or receiving infrared signals, as described below.
- Fig. 12 illustrates a block diagram of the internal hardware of the computer system 900 of Fig. 11.
- a bus 914 serves as the main information highway interconnecting the other components of the computer system 900.
- CPU 916 is the central processing unit of the system, performing calculations and logic operations required to execute a program.
- Read only memory (ROM) 918 and random access memory (RAM) 920 constitute the main memory of the computer.
- Disk controller 922 interfaces one or more disk drives to the system bus 914. These disk drives are, for example, floppy disk drives such as 904, or CD ROM or DVD (digital video disks) drive such as 906, or internal or external hard drives 924. As indicated previously, these various disk drives and disk controllers are optional devices.
- a display interface 926 interfaces display 908 and permits information from the bus 914 to be displayed on the display 908. Again as indicated, display 908 is also an optional accessory. For example, display 908 could be substituted or omitted. Communications with external devices, for example, the components of the apparatus described herein, occurs utilizing communication port 928. For example, optical fibers and/or electrical cables and/or conductors and/or optical communication (e.g., infrared, and the like) and/or wireless communication (e.g., radio frequency (RF), and the like) can be used as the transport medium between the external devices and communication port 928.
- Peripheral interface 930 interfaces the keyboard 910 and the mouse 912, permitting input data to be transmitted to the bus 914.
- the computer also optionally includes an infrared transmitter and/or infrared receiver.
- Infrared transmitters are optionally utilized when the computer system is used in conjunction with one or more of the processing components/stations that transmits/receives data via infrared signal transmission.
- the computer system optionally uses a low power radio transmitter and/or a low power radio receiver.
- the low power radio transmitter transmits the signal for reception by components of the production process, and receives signals from the components via the low power radio receiver.
- the low power radio transmitter and/or receiver are standard devices in industry.
- Figure 13 is an illustration of an exemplary memory medium 932 which can be used with disk drives illustrated in Figures 11 and 12.
- memory media such as floppy disks, or a CD ROM, or a digital video disk will contain, for example, a multi-byte locale for a single byte language and the program information for controlling the computer to enable the computer to perform the functions described herein.
- ROM 918 and/or RAM 920 illustrated in Figure 12 can also be used to store the program information that is used to instruct the central processing unit 916 to perform the operations associated with the production process.
- Computer system 900 is illustrated having a single processor, a single hard disk drive and a single local memory, the system 900 is optionally suitably equipped with any multitude or combination of processors or storage devices.
- Computer system 900 is, in point of fact, able to be replaced by, or combined with, any suitable processing system operative in accordance with the principles of the present invention, including sophisticated calculators, and hand-held, laptop/notebook, mini, mainframe and super computers, as well as processing system network combinations of the same.
- the hardware configuration is, for example, arranged according to the multiple instruction multiple data (MIMD) multiprocessor format for additional computing efficiency.
- MIMD multiple instruction multiple data
- the details of this form of computer architecture are disclosed in greater detail in, for example, U.S. Patent No. 5,163,131; Boxer, A., Where Buses Cannot Go, IEEE Spectrum, February 1995, pp. 41-45; and Barroso, L.A. et al., RPM: A Rapid Prototyping Engine for Multiprocessor Systems, IEEE Computer February 1995, pp. 26-34, all of which are inco ⁇ orated herein by reference.
- processors may be replaced by or combined with any other suitable processing circuits, including programmable logic devices, such as PALs (programmable array logic) and PLAs (programmable logic arrays).
- PALs programmable array logic
- PLAs programmable logic arrays
- DSPs digital signal processors
- FPGAs field programmable gate arrays
- ASICs application specific integrated circuits
- VLSIs very large scale integrated circuits
- FIG 14 is an illustration of the architecture of the combined internet, POTS, and ADSL architecture for use in the present invention in accordance with another embodiment.
- the voice part of the spectrum (the lowest 4 kHz) is optionally separated from the rest by a passive filter, called a POTS splitter 258, 260.
- the rest of the available bandwidth - from about 10 kHz to 1 MHz - carries data at rates up to 6 bits per second for every hertz of bandwidth from data equipment 262, 264, 294.
- the ADSL equipment 256 then has access to a number of destinations including significantly the Internet 268, and other destinations 270, 272.
- ADSL makes use of advanced modulation techniques, of which the best known is the discrete multitone (DMT) technology.
- DMT discrete multitone
- ADSL transmits data asymmetrically - at different rates upstream toward the central office 252 and downstream toward the subscriber 250.
- Cable television providers are providing analogous Internet service to PC users over their TV cable systems by means of special cable modems.
- modems are capable of transmitting up to 30 Mb/s over hybrid fiber/coax systems, which use fiber to bring signals to a neighborhood and coax to distribute it to individual subscribers.
- Cable modems come in many forms. Most create a downstream data stream above 50 MHz (and more likely 550 MHz) and carve an upstream channel out of the 5-50-MHz band, which is currently unused. Using 64-state quadrature amplitude modulation (64 QAM), a downstream channel can realistically transmit about 30 Mb/s (the oft-quoted lower speed of 10 Mb/s refers to PC rates associated with Ethernet connections). Upstream rates differ considerably from vendor to vendor, but good hybrid fiber/coax systems can deliver upstream speeds of a few megabits per second. Thus, like ADSL, cable modems transmit much more information downstream than upstream.
- 64 QAM quadrature amplitude modulation
- the internet architecture 220 and ADSL architecture 254, 256 may also be combined with, for example, user networds 222, 224, and 228.
- users may access or use or participate in the administration, management computer assisted program in computer 240 via various different access methods.
- the various databases 230, 232, 234, 236 and/or 238 are accessible via access to and/or by computer system 240, and/or via internet/local area network 220. These databases may optionally include objective criteria for evaluating the corporate governance characteristics for ranking the co ⁇ oration. For example, environmental data is generally publicly available which indicates a co ⁇ oration's compliance history, outstanding violations or potential violations, and the like.
- standard legal and/or regulatory and/or administrative and/or business databases may be consulted to obtain additional information on co ⁇ orate governance techniques, potential for government intervention, shareholder participation and/or customer loyalty. All this data may then be collected and analyzed to determine the overall attributes of the co ⁇ orate, shareholder, government, and customer agents, for input into the simulation. Alternatively, the individual data may be used and input into the simulation, and the simulation may digest or process the data individually or collectively as part of the simulation.
- workstation 240 optionally includes modules 242, 246, 248, and 250 for individually handling the operations/simulation of the different agents.
- one module or a different number of modules may be used for processing the agent relationships, processes, and or interactions.
- the scope of the instant invention includes any suitable internet (lower case), i.e., any set of networks interconnected with devices, such as routers, that forward messages or fragments of messages between networks or intranets.
- the Internet is one of the largest examples of an internet.
- the elements of the service provider network shown for illustrative pu ⁇ oses in Figures 12 and 14 as being located in geographic proximity to one another in a substantially centralized processing environment, may alternatively be arranged in a standard distributed processing environment so as to leverage resources, e.g., servers and storage devices, located at two or more sites.
- resources e.g., servers and storage devices
- the above-mentioned computer network may include a virtual private network (VPN), thereby taking advantage of existing PSTN infrastructure while providing a secure and private environment for information exchange regarding resource usage.
- VPN virtual private network
- data sent from the VPN is encrypted, thereby enhancing the privacy of customers.
- the instant invention effectively uses the Internet as part of a private secure network. That is, the "tunnel" is the particular path that a given company message or file might travel through the Internet.
- the above-described computer network may alternatively include an extranet, wherein customers may securely exchange large volumes of resource usage data using a standard data exchange format, for example, Electronic Data Interchange.
- an extranet may enable customers to share news of common interest, for example, aggregated resource usage, exclusively with partner companies.
- standard text-only browsers such as Lynx
- Lynx may be used for UNIX shell and VMS users. Users of such text-only browsers may download comma-delimited ASCII files of, for example, their usage data.
Abstract
Description
Claims
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
AU46752/00A AU4675200A (en) | 1999-04-29 | 2000-05-01 | Method and system for knowledge guided hyperintensity detection and volumetric measurement |
Applications Claiming Priority (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US13159099P | 1999-04-29 | 1999-04-29 | |
US60/131,590 | 1999-04-29 | ||
US19882500P | 2000-04-21 | 2000-04-21 | |
US60/198,825 | 2000-04-21 |
Publications (3)
Publication Number | Publication Date |
---|---|
WO2000065985A2 true WO2000065985A2 (en) | 2000-11-09 |
WO2000065985A3 WO2000065985A3 (en) | 2001-02-01 |
WO2000065985A9 WO2000065985A9 (en) | 2002-07-11 |
Family
ID=26829611
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US2000/011457 WO2000065985A2 (en) | 1999-04-29 | 2000-05-01 | Method and system for knowledge guided hyperintensity detection and volumetric measurement |
Country Status (2)
Country | Link |
---|---|
AU (1) | AU4675200A (en) |
WO (1) | WO2000065985A2 (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2002093450A1 (en) * | 2001-05-16 | 2002-11-21 | Cellavision Ab | Information processing for distinguishing an object |
WO2005057498A1 (en) * | 2003-12-12 | 2005-06-23 | Agency For Science, Technology And Research | Method and apparatus for identifying pathology in brain images |
DE102005058095A1 (en) * | 2005-12-05 | 2007-06-06 | Forschungszentrum Jülich GmbH | Method for the topographic representation of changes in a studied brain |
WO2009088370A1 (en) | 2008-01-10 | 2009-07-16 | Agency For Science, Technology And Research | Discriminating infarcts from artifacts in mri scan data |
CN103077298A (en) * | 2012-10-24 | 2013-05-01 | 西安电子科技大学 | Image voxel and priori brain atlas division fused brain network construction method |
US9466105B2 (en) | 2015-01-05 | 2016-10-11 | National Central University | Magnetic resonance imaging white matter hyperintensities region recognizing method and system |
CN106204600A (en) * | 2016-07-07 | 2016-12-07 | 广东技术师范学院 | Cerebral tumor image partition method based on multisequencing MR image related information |
US9867566B2 (en) | 2014-11-21 | 2018-01-16 | The Trustees Of Columbia University In The City Of New York | Technologies for white matter hyperintensity quantification |
CN109903278A (en) * | 2019-02-25 | 2019-06-18 | 南京工程学院 | Ultrasonic tumor of breast form quantization characteristic extracting method based on shape histogram |
CN111968130A (en) * | 2020-07-23 | 2020-11-20 | 沈阳东软智能医疗科技研究院有限公司 | Brain angiography image processing method, apparatus, medium, and electronic device |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5163131A (en) | 1989-09-08 | 1992-11-10 | Auspex Systems, Inc. | Parallel i/o network file server architecture |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5812691A (en) * | 1995-02-24 | 1998-09-22 | Udupa; Jayaram K. | Extraction of fuzzy object information in multidimensional images for quantifying MS lesions of the brain |
GB9512012D0 (en) * | 1995-06-13 | 1995-08-09 | British Tech Group | Apparatus for image enhancement and related method |
-
2000
- 2000-05-01 WO PCT/US2000/011457 patent/WO2000065985A2/en active Application Filing
- 2000-05-01 AU AU46752/00A patent/AU4675200A/en not_active Abandoned
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5163131A (en) | 1989-09-08 | 1992-11-10 | Auspex Systems, Inc. | Parallel i/o network file server architecture |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2002093450A1 (en) * | 2001-05-16 | 2002-11-21 | Cellavision Ab | Information processing for distinguishing an object |
US7889895B2 (en) | 2003-12-12 | 2011-02-15 | Agency For Science, Technology And Research | Method and apparatus for identifying pathology in brain images |
WO2005057498A1 (en) * | 2003-12-12 | 2005-06-23 | Agency For Science, Technology And Research | Method and apparatus for identifying pathology in brain images |
DE102005058095A1 (en) * | 2005-12-05 | 2007-06-06 | Forschungszentrum Jülich GmbH | Method for the topographic representation of changes in a studied brain |
US8160317B2 (en) | 2005-12-05 | 2012-04-17 | Forschungszentrum Juelich Gmbh | Method for topographical presentation of alterations in an examined brain |
EP2232443A1 (en) * | 2008-01-10 | 2010-09-29 | Agency for Science, Technology And Research | Discriminating infarcts from artifacts in mri scan data |
US20100290689A1 (en) * | 2008-01-10 | 2010-11-18 | Varsha Gupta | Discriminating infarcts from artifacts in mri scan data |
JP2011509141A (en) * | 2008-01-10 | 2011-03-24 | エージェンシー フォー サイエンス,テクノロジー アンド リサーチ | Discrimination between infarctions and artifacts in MRI scan data |
WO2009088370A1 (en) | 2008-01-10 | 2009-07-16 | Agency For Science, Technology And Research | Discriminating infarcts from artifacts in mri scan data |
EP2232443A4 (en) * | 2008-01-10 | 2012-07-04 | Agency Science Tech & Res | Discriminating infarcts from artifacts in mri scan data |
CN103077298A (en) * | 2012-10-24 | 2013-05-01 | 西安电子科技大学 | Image voxel and priori brain atlas division fused brain network construction method |
CN103077298B (en) * | 2012-10-24 | 2015-09-30 | 西安电子科技大学 | The brain network construction method that fused images voxel and priori brain map divide |
US9867566B2 (en) | 2014-11-21 | 2018-01-16 | The Trustees Of Columbia University In The City Of New York | Technologies for white matter hyperintensity quantification |
US9466105B2 (en) | 2015-01-05 | 2016-10-11 | National Central University | Magnetic resonance imaging white matter hyperintensities region recognizing method and system |
CN106204600A (en) * | 2016-07-07 | 2016-12-07 | 广东技术师范学院 | Cerebral tumor image partition method based on multisequencing MR image related information |
CN109903278A (en) * | 2019-02-25 | 2019-06-18 | 南京工程学院 | Ultrasonic tumor of breast form quantization characteristic extracting method based on shape histogram |
CN111968130A (en) * | 2020-07-23 | 2020-11-20 | 沈阳东软智能医疗科技研究院有限公司 | Brain angiography image processing method, apparatus, medium, and electronic device |
CN111968130B (en) * | 2020-07-23 | 2023-08-04 | 沈阳东软智能医疗科技研究院有限公司 | Brain contrast image processing method, device, medium and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
WO2000065985A9 (en) | 2002-07-11 |
WO2000065985A3 (en) | 2001-02-01 |
AU4675200A (en) | 2000-11-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US6430430B1 (en) | Method and system for knowledge guided hyperintensity detection and volumetric measurement | |
Mortazavi et al. | Segmentation of multiple sclerosis lesions in MR images: a review | |
US4945478A (en) | Noninvasive medical imaging system and method for the identification and 3-D display of atherosclerosis and the like | |
Udupa et al. | Multiple sclerosis lesion quantification using fuzzy-connectedness principles | |
Klifa et al. | Quantification of breast tissue index from MR data using fuzzy clustering | |
US9036878B2 (en) | Method for delineation of tissue lesions | |
Clas et al. | A semi-automatic algorithm for determining the demyelination load in metachromatic leukodystrophy | |
US20090129671A1 (en) | Method and apparatus for image segmentation | |
Hu et al. | Segmentation of brain from computed tomography head images | |
CN113610808A (en) | Individual brain atlas individualization method, system and equipment based on individual brain connection atlas | |
Sun et al. | Intracranial hemorrhage detection by 3D voxel segmentation on brain CT images | |
WO2000065985A2 (en) | Method and system for knowledge guided hyperintensity detection and volumetric measurement | |
Ramasamy et al. | Segmentation of brain tumor using deep learning methods: A review | |
Ghorbanzadeh et al. | An Investigation into the Performance of Adaptive Neuro-Fuzzy Inference System for Brain Tumor Delineation Using ExpectationMaximization Cluster Method; a Feasibility Study | |
JPH11507565A (en) | Image improvement apparatus and method | |
Supriyanti et al. | Coronal slices segmentation of mri images using active contour method on initial identification of alzheimer severity level based on clinical dementia rating (CDR) | |
Campadelli et al. | Automated morphometric analysis in peripheral neuropathies | |
Bijar et al. | Increasing the contrast of the brain MR FLAIR images using fuzzy membership functions and structural similarity indices in order to segment MS lesions | |
Forbes et al. | Model-based region-of-interest selection in dynamic breast MRI | |
Cherradi et al. | Fully automatic method for 3D T1-weighted brain magnetic resonance images segmentation | |
Roy et al. | Tumor delineation from 3-d mr brain images | |
CN115240014B (en) | Medical image classification system based on residual error neural network | |
Anbumozhi | Performance analysis of brain tumor detection based on fuzzy logic and neural network classifier | |
Mansingh et al. | PREMORBID BRAIN VOLUME AGAINST ALZHEIMER'S DISEASE USING MULTIMODAL BIG MEDICAL DATA ANALYSIS | |
Kruggel | Automatical adaption of anatomical masks to the neocortex |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A2 Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY CA CH CN CR CZ DE DK DM DZ EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX NO NZ PL PT RO RU SD SE SG SI SK SL TJ TM TR TT TZ UA UG UZ VN YU ZA ZW |
|
AL | Designated countries for regional patents |
Kind code of ref document: A2 Designated state(s): GH GM KE LS MW SD SL SZ TZ UG ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE BF BJ CF CG CI CM GA GN GW ML MR NE SN TD TG |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application | ||
DFPE | Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101) | ||
AK | Designated states |
Kind code of ref document: A3 Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY CA CH CN CR CZ DE DK DM DZ EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX NO NZ PL PT RO RU SD SE SG SI SK SL TJ TM TR TT TZ UA UG UZ VN YU ZA ZW |
|
AL | Designated countries for regional patents |
Kind code of ref document: A3 Designated state(s): GH GM KE LS MW SD SL SZ TZ UG ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE BF BJ CF CG CI CM GA GN GW ML MR NE SN TD TG |
|
REG | Reference to national code |
Ref country code: DE Ref legal event code: 8642 |
|
AK | Designated states |
Kind code of ref document: C2 Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY CA CH CN CR CZ DE DK DM DZ EE ES FI GB GD GE GH GM HR HU ID IL IN IS JP KE KG KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX NO NZ PL PT RO RU SD SE SG SI SK SL TJ TM TR TT TZ UA UG UZ VN YU ZA ZW |
|
AL | Designated countries for regional patents |
Kind code of ref document: C2 Designated state(s): GH GM KE LS MW SD SL SZ TZ UG ZW AM AZ BY KG KZ MD RU TJ TM AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE BF BJ CF CG CI CM GA GN GW ML MR NE SN TD TG |
|
COP | Corrected version of pamphlet |
Free format text: PAGES 1/8-8/8, DRAWINGS, REPLACED BY NEW PAGES 1/10-10/10; DUE TO LATE TRANSMITTAL BY THE RECEIVINGOFFICE |
|
122 | Ep: pct application non-entry in european phase | ||
NENP | Non-entry into the national phase in: |
Ref country code: JP |