EP1194880A1 - Detection assistee par ordinateur de masses et strategies de microcalcifications groupees - Google Patents

Detection assistee par ordinateur de masses et strategies de microcalcifications groupees

Info

Publication number
EP1194880A1
EP1194880A1 EP00944852A EP00944852A EP1194880A1 EP 1194880 A1 EP1194880 A1 EP 1194880A1 EP 00944852 A EP00944852 A EP 00944852A EP 00944852 A EP00944852 A EP 00944852A EP 1194880 A1 EP1194880 A1 EP 1194880A1
Authority
EP
European Patent Office
Prior art keywords
image
detections
detection
microcalcification
delta
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
EP00944852A
Other languages
German (de)
English (en)
Inventor
Steven K. Rogers
Kenneth W. Bauer
Randy P. Broussard
Michael J. Collins
Martin P. Desimio
Jeffrey W. Hoffmeister
Richard A. Mitchell
Thomas F. Rathbun
John E. Rosenstengel
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Qualia Computing Inc
Original Assignee
Qualia Computing Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Qualia Computing Inc filed Critical Qualia Computing Inc
Publication of EP1194880A1 publication Critical patent/EP1194880A1/fr
Pending legal-status Critical Current

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

Definitions

  • the present invention relates to a method and system for computer aided analysis of detections on digital images and , more particularly, to a method and system for assisting radiologists in reading medical images.
  • CAD Computer aided detection
  • the first approach is known as bilateral subtraction.
  • the left and right breast images are subtracted to produce a difference image. Since cancer is typically unilateral and since bodies are relatively symmetric, in many cases the difference image emphasizes cancerous regions.
  • the second approach is similar to bilateral subtraction. However, in this case, a temporal subtraction is performed wherein the same view of a body part taken at an earlier time is subtracted from the current image. This temporal subtraction is intended to highlight regions that have changed during the time interval between imaging sessions.
  • the present invention provides a computer aided detection method and system to assist radiologists in the reading of medical images.
  • a set of digital mammogram images is obtained, such as from mammograrns obtained from a patient's screening visit.
  • a rectangular analysis region containing breast tissue is segmented from each digital mammogram image, and a binary mask corresponding to the breast tissue is created.
  • Clustered microcalcifications are detected in a microcalcification detection step. Point features from each microcalcification are applied to a classifier trained to pass microcalcifications representative of malignant clusters. Those microcalcifications remaining after the classifier are represented by their centroid location and subsequently grouped into clusters. Features are computed from the clusters and each is subsequently assigned a class label and a quality score. Densities are detected in a density detection step. In this step, a subsampled image is applied to a bank of DoG filters to produce an image with enhanced local bright spots. This image is then thresholded and the thresholded image is input to a screening classifier. Detections passing the screening classifier as suspicious are input to a region growing step and density features are computed from the region grown area. Detections are input to a density classification step, the output of which includes computed features, a class label, and two quality scores for each detection.
  • Detections including both microcalcifications and densities, are further analyzed in a post processing stage.
  • the detections across all the images in the case are considered collectively by case based ranking.
  • the collection of detections across detection category are assigned to one of 29 image context categories based on associated quality scores. When certain image context categories of detections are observed on an image, all the detections on that image are removed.
  • the remaining detections across all the images in the case are considered within microcalcification and density detection categories.
  • "normalcy" features are computed using the quality scores from all the detections in the case from that category.
  • the present method and system is in contrast to subtraction-based methods proposed by the prior art in that such prior art methods are inherently limited to two input images, whereas the present invention performs analysis of detections extending across a set of images for a patient, such as a set of four images forming a case for the patient.
  • the approach of the present invention provides significant improvement in accomplishing the objectives of obtaining both increased system sensitivity and reduction of false positive detections.
  • Fig. 1 is a flow diagram illustrating the automated method for the detection of masses and clustered microcalcifications in digital mammograms.
  • Fig. 2 is flow diagram of a general classifier architecture.
  • Fig. 3 is a flow diagram of the preprocessing operations.
  • Figs. 4-5 is a flow diagram of the clustered microcalcification detector.
  • Figs. 6-7 is a flow diagram of the density detector.
  • Fig. 8 is a flow diagram of the Difference of Gaussians filterbank.
  • Fig. 9 is a flow diagram of the region growing method.
  • Fig. 10 is a flow diagram detailing the adaptive weighting portion of the region growing method.
  • Figs. 11-12 is a flow diagram illustrating the clustered microcalcifications classifier.
  • Figs. 13-14 is a flow diagram illustrating the density classifier.
  • Fig. 15 is a flow diagram of the post processing stage.
  • a flow diagram generally illustrating a sequence of steps performed in order to detect the locations of clusters of microcalcifications and densities within a set of digital mammograms obtained from the screening views associated with a patient's screening mammography session.
  • a set of digital mammograms is obtained using hardware such as digital mammography systems, or by digitizing mammography films using laser or charge-coupled device (CCD) digitizers.
  • CCD charge-coupled device
  • a rectangular analysis region containing breast tissue is segmented from the digital mammogram image to decrease the time required for processing the mammogram images.
  • a binary mask corresponding to the breast tissue is created for use in later processing steps.
  • the breast mask is used to limit detections to areas of the image containing breast tissue.
  • the preprocessing step 2000 also produces a subsampled version of the input image.
  • the subsampled image is input to a subsequent processing stage. Since the subsampled image has significantly fewer pixels to process, the time required to analyze the subsampled image is much less than for the original input image.
  • the preprocessing stage also creates a binary pectoral mask corresponding to tissue of the pectoral muscle.
  • the pectoral mask may be used to suppress detections in the pectoral region. Clustered microcalcifications are detected in a microcalcification detection step 3000.
  • the resultant image is thresholded to detect microcalcifications. Detections near the breast edge, as well as those not in the breast mask, are removed and "point" features are computed from individual microcalcification detections. The point features from each microcalcification are applied to a classifier trained to pass microcalcifications representative of malignant clusters. Those microcalcifications remaining after the classifier are represented by their centroid location and subsequently grouped into clusters. "Cluster" features are computed from the clusters of microcalcifications in step 4000.
  • DoG difference of Gaussians
  • Detected clusters are assigned class labels, ' 1 ' or '0' for either suspicious or non- suspicious, in a classification step 5000.
  • the output of the microcalcification classification stage includes the computed features, the class label, and a quality score for each detected region.
  • Densities are detected in a density detection step 6000.
  • the input to this step is the subsampled image from the preprocessing stage. After first processing the image to reduce false detections associated with the skin edge boundary, locally bright densities are emphasized by trend correction. The trend corrected image is then applied to a bank of DoG filters.
  • the filter bank outputs are nonlinearly combined to produce an image with further enhanced densities.
  • This image is thresholded in a peak detection step and the resultant groups of pixels are grouped into connected objects and the breast mask is applied. These objects and the image pixels from the original images and the filter bank output images are used to create "density screening" features. These density screening features are input to a density screening classifier used to label the detected objects as suspicious or non-suspicious. Objects passing the screening classifier as suspicious are input to a region growing step which estimates the shape of the detected locally bright spot.
  • step 7000 density features are computed from the region grown area using the subsampled input image and the filter bank outputs.
  • An expert filter step removes certain detections by comparing density features from the region grown areas to thresholds. Detections with density feature values beyond the thresholds are removed from further consideration. Subsequent to the expert filter stage, detections are assigned class labels, ' 1 ' or '0' for either suspicious or non-suspicious, in the density classification step, 8000.
  • the output of the density classification stage includes the computed features, the class label, and two quality scores for each detected region, corresponding to the categories of "mass" and "spiculated mass". The word "density" applies to both types of masses.
  • Class labels and quality scores for the detected microcalcification and density regions are further analyzed in a post processing step, 9000.
  • the detections are considered in two fundamentally different ways than in the preceding steps; across all images in the case and across microcalcification and density detection categories. Only detections with class labels of ' 1 ' at the end of the post processing steps are displayed at the output of the system.
  • Case based ranking limits the number of microcalcification and density detections with class labels of ' 1 ' to predetermined maximum numbers per case and per image using the quality scores associated with each detection.
  • a second post processing step within each image, the collection of detections across microcalcification and density detection categories are assigned to one of 29 image context combinations based on associated quality scores. When certain image context combinations of detections are observed on an image, all the detections on that image have their class labels set to '0'.
  • a third post processing step all detections across all the images in the case, with class labels of ' 1 ' or '0', are considered within microcalcification and density detection categories.
  • "normalcy" features are computed using the quality scores from all the detections in the case from that category. These normalcy features are applied to a classification stage designed to assign the entire case as "Normal” or “Not Normal”.
  • the classifier assigns the case as "Normal”
  • all detections of the corresponding category, microcalcification or density have their class labels set to '0' across all the images in that case.
  • the detections with class labels of ' 1 ' are displayed overlaid on the set of digital images as "Indications".
  • CLASSIFIER DESIGN This invention uses pattern recognition subsystems throughout the processing stages. This section provides an overview to the areas of pattern recognition and classifier design as they apply to this invention.
  • Pattern recognition is the process of making decisions based on measurements.
  • suspicious regions or detections are located by a detector, and ultimately accepted or rejected for display.
  • the detector is designed to locate many regions that may be malignant. This leads to the detection of many regions that are normal.
  • the purpose of the pattern recognition subsystems is to remove as many of the false detections as possible while simultaneously retaining as many of the truly malignant detections as possible.
  • a detection corresponding to a cancerous region is referred to as a "true positive" detection, or simply, a TP.
  • a detection corresponding to a normal region is known as a "false positive" detection, or an FP.
  • the first step in the process is to characterize the detections. Toward this end, multiple measurements are computed from each of the detected regions. Each measurement is referred to as a feature. The collection of measurements for that detected region is referred to as its feature vector, wherein each element of the vector represents a feature value
  • the second step in the process is to use the information in the feature vectors to assign the detection a quality score or possibly a class label.
  • the feature vectors from a detector are provided as inputs to a classifier, which computes a number for a given feature vector input.
  • One method to specify a classifier is by a set of discriminant functions.
  • a discriminant function computes a single value as a function of an input feature vector.
  • the classifier shown in Fig. 2 uses one discriminant function per class.
  • the output value of a discriminant function is referred to as a discriminant.
  • the discriminants are subtracted and compared to a threshold value, ⁇ . If the difference of discriminants is greater than the threshold, the class label is set to ' 1 ', otherwise the class label is set to '0'.
  • the method for designing the discriminant functions is a statistical approach providing quadratic discriminant functions. Under certain general conditions, quadratic discriminant functions can provide the minimum average probability of error.
  • Design of quadratic discriminant functions begins with labeled training data. Labeled training data were generated in the following manner. A collection of mammo graphic images were examined by certified radiologists to confirm regions of cancer based on pathological analysis of biopsied cells from that region. Truth data were generated by the radiologists marking truth boxes over the corresponding digital image regions associated with the cancer. Categories of cancer corresponding to microcalcification, mass, and spiculated mass are identified.
  • the classes are ultimately cancer and not cancer.
  • the mean and covariance matrices are computed from the labeled feature vectors of the training data.
  • the quadratic classifier discriminant functions are given by
  • X is the feature vector
  • M,- is the class i mean
  • ⁇ is the class i covariance matrix.
  • the quantity ? is a bias term which can be modified to fine tune the decision process.
  • the discriminant function, G t (x), is a distance function; therefore small discriminants indicate a high probability that the observation belongs to class i.
  • the difference of discriminants calculated as Delta, G 0 (X) - G,(X) (2) where the difference of discriminants is referred to as a delta discriminant.
  • Delta, is greater than zero implies that the feature vector X is closer to the cancer category than the not cancer category. Therefore, if Delta, > 0 for any i ⁇ 0 , then the observation is considered to be a cancer.
  • the greater the difference of discriminants the higher the quality of the detection may be considered.
  • a digital mammogram image 201 is first low pass filtered in step 2010 and subsequently decimated by a factor of 16 in the horizontal and vertical dimensions in step 2020.
  • the above operations provide a reduced resolution image, 2070, suitable for subsequent steps that do not require full spatial resolution.
  • the reduced resolution image is referred to as a subsampled image.
  • the subsampled image is input to the autocropper, 2030.
  • the autocropper produces a binary mask, 2040, known as a Breast
  • the pectoral mask is a binary mask identifying the location of pectoral muscle tissue in the image, and methods for forming pectoral masks are known in the art. It is desirable to inhibit these detections in the pectoral muscle because cancer only infrequently occurs in the pectoral muscle. Furthermore, this portion of the digital mammogram is usually brighter than the other regions, which can cause false positive detections. Thus, inhibiting detections in the pectoral muscle improves the performance of the system of the present invention.
  • Figs. 4 and 5 where Fig. 5 is a continuation of Fig. 4, there is seen therein a flow diagram illustrating in detail the clustered microcalcification detection and clustered microcalcification feature computation systems 3000 and 4000 of the invention.
  • the digital image 201 and the crop rectangle 2050 are input to a DoG filtering step 3010 to accentuate subimage areas containing locally bright spots of the size range of microcalcifications, thus creating local maxima in the filtered image.
  • the DoG-filtered image is then thresholded in step 3020 to segment the maxima that represent detections of microcalcifications.
  • the coordinate representations of the detected maxima are compared with the binary breast mask 2040 and pectoral mask 2060 to remove false detections outside the breast area and within but near the breast edge in step 3030.
  • the remaining coordinate representations in the analysis area are subjected to a point feature computation step 3040 wherein measurements from individual microcalcifications are calculated.
  • These point features are input to a quadratic classifier step 3050 designed to retain up to a predetermined number of detections most characteristic of individual microcalcifications associated with cancerous lesions.
  • the retained detections are clustered in step 3060.
  • Microcalcification cluster features are computed for the clusters in a feature computation step 4000, producing microcalcification cluster feature vectors and locations of clustered microcalcifications, 4100.
  • the digital image is first filtered with a DoG kernel to enhance microcalcifications.
  • the DoG kernel is defined in an N by N square array of values. In the preferred embodiment, the kernel is 77 by 77.
  • Filtering is accomplished by convolving the input image with the DoG kernel. In an alternative embodiment, filtering is accomplished by first obtaining the fast Fourier transforms (FFTs) of the input image and the DoG kernel, then multiplying the FFTs together, and taking the inverse FFT as the result.
  • FFTs fast Fourier transforms
  • the DoG kernel is applied here as an optimal detector of individual microcalcifications as described in U.S. Patent No. 5,999,639.
  • the DoG kernel is mathematically expressed as
  • ⁇ j and ⁇ 2 are standard deviations of individual Gaussian kernels and ci and c 2 are normalizing constants for Gaussian kernels.
  • Optimal values for ⁇ j are detailed in U.S. Patent No. 5,999,639.
  • ⁇ l. ⁇ i
  • microcalcifications tend to be among the brightest object in DoG-filtered images, they may exist within regions of high average gray levels and thus be difficult to reliably segment.
  • the thresholding process used in one embodiment of the invention that generally addressees these concerns involves the local sloping threshold detailed in U.S. Patent No. 5,999,639.
  • d(x,y) Each pixel is compared to the threshold value T(x,y).
  • T(x,y) The result of thresholding is a locally thresholded image, l_s(x,y). Pixels in l_s(x,y) are set to T where values of d(x,y) are greater than T(x,y), and set to '0' elsewhere. This creates objects in the thresholded image. A predetermined number of detections are retained based on the difference between the brightest image pixel at locations of ' l 's in the object and the threshold value at that location.
  • detections are converted to single-pixel representations by computing the centroid or center of gravity of the objects found by the thresholding process. Detections are thus represented as single pixels with a value of logical ' 1 ' while the remaining pixels have a value of logical '0' . Objects with centroids less than a predetermined distance from the breast edge are rejected to reduce the false positive detections occurring due to the discontinuity in brightness values between the breast and background.
  • Calcifications associated with malignancies have certain characteristics in common; these traits may be exploited during the distinction of regions likely to be from microcalcifications typical of cancerous regions from those microcalcifications or false detections that are not of concern.
  • the separation of individual microcalcification detections is accomplished by first computing point features from each detection, then applying the subsequent point feature vectors to a point classifier.
  • a preferred embodiment of the classifier is a quadratic classifier.
  • Point features are extracted for each microcalcification at location (x,y) as described.
  • the three point features computed for each of the microcalcifications in a preferred embodiment are: 1.
  • the response of the target size 13 (t 13)
  • DoG kernel ( N x N) centered at (x,y), where in a preferred embodiment, N 77.
  • the nonlinear combination of point feature 3 is accomplished with a multilayer perceptron (MLP) artificial neural network, consisting of 5 input nodes and 2 output nodes. The first layer nodes use the hyperbolic tangent activation function and the output node uses a linear function.
  • MLP multilayer perceptron
  • the MLP is trained using the backpropagation learning rule.
  • the desired response of the first output node is one for microcalcification detections inside a cluster truth box and zero otherwise.
  • the desired response of the second output node is one for microcalcification detections outside a cluster truth box and zero otherwise.
  • the third point feature is calculated as follows. First, the measurements 3a, 3b, and 3c from above are computed and stored in a vector. Prior to application to the MLP, the input vector is normalized to zero mean and unit variance to provide robustness across a range of input data variations. The normalized input vector is applied to the MLP and values obtained at each output node. The value of the second output node is subtracted from the value of the first output node to provide the third point feature.
  • the microcalcification point classifier uses a two class quadratic classifier. The two classes are normal and malignant. The point features for each detection are tested with the classifier providing normal and malignant discriminant values. If the difference of discriminants is negative then the detection is more likely to be normal than malignant. Therefore, that point is rejected from further analysis. However, if the difference of discriminants is positive the detection is retained for further analysis and the associated point features and difference of discriminants are stored for further processing.
  • Remaining detections are rank ordered, largest to smallest, by their difference of discriminants. If their rank is greater than a predetermined number (50 in a preferred embodiment) they are rejected from further analysis. This operation retains only the microcalcification detections that are most likely malignant. These remaining detections are applied to the subsequent clustering stage.
  • a set of microcalcification detections must satisfy two requirements. First, every microcalcification in a cluster must be within a specified distance of another microcalcification in that cluster. Second, there must be at least a minimum number of microcalcifications in a cluster. Therefore, when the set of detections from the point classifier are clustered, some detections may be dropped. For example, a lone microcalcification separated from another microcalcification by more than a predetermined maximum distance will not survive the clustering step.
  • Linear density computed as the number of detected microcalcifications divided by the maximum interpoint distance (PT).
  • Average Gabor value the average value resulting from applying the set of pixels from the digital image corresponding to the pixel locations of the detected cluster through a bank of Gabor filters. Averaging is across the maximum value over all N orientations of filters in the Gabor filter bank.
  • Gabor filters are defined at 24 orientation angles, between 0 degrees and 180 degrees, in an M by M array.
  • the period T is found as the kernel size divided by twice the target size.
  • the kernel size corresponds to a neighborhood of pixels with size 32 by 32 and the target size is 7.0 pixels.
  • the width of the Gaussian function equals 1.0.
  • the phase of the sinusoid equals ⁇ /2 to better respond to locally bright spots (PT).
  • Average intensity ratio - this feature is computed by region growing each detection in the cluster using the point locations as seed pixels. The average intensity of image pixels inside each region grown area is divided by the average intensity of image pixels outside the region grown area within a rectangular neighborhood of predetermined size. The intensity ratios for all detections in the cluster are summed then divided by the number of points in the cluster to obtain the average intensity ratio feature (PT).
  • PT average intensity ratio feature
  • Average second central moment - this feature is computed by region growing each detection in the cluster using the point locations as seed pixels. A rectangular neighborhood of predetermined size centered at the point location in the input image is used normalized by the energy in that neighborhood. Then, a second central moment of the region grown portions of the energy normalized image are computed. The second central moments are averaged to obtain this feature.
  • the particular second central moment used in this embodiment is defined as
  • microcalcification cluster features are applied to the microcalcification cluster classifier, step 5000.
  • Density detection is the process of locating regions of mammographic images associated with malignant masses. Density detection is considered more difficult than microcalcification detection for several reasons. Microcalcifications of concern although small, range in size over a relatively small interval, are typically very bright, and will have high contrast within most breast tissue. Densities of concern however are quite varied in size ( 0.5 to 2.5 cm) , and may have low contrast. Furthermore, microcalcifications have distinct boundaries with breast tissue; whereas the boundaries between densities and background tissue may be ill- defined.
  • Fig. 7 is a continuation of Fig. 6, there is seen therein a flow diagram illustrating in detail the density detection and density feature computation systems 6000 and 7000 of the invention.
  • the subsampled digital image 2070 and the breast mask 2040 are input to a preprocessing step 6010.
  • the boundary between the breast and background is smoothed to reduce the number of false detections which occur due solely to the large contrast between breast tissue and the background of the film.
  • the output image from processing the breast tissue is trend corrected whereby the local value of image intensity is estimated in a neighborhood, then subtracted from that neighborhood. This provides increased contrast and therefore allows for detection of lesions with low contrast in the original image.
  • the resultant image is then input to a MaxDoG filterbank in step 6030, the output of which is an image with accentuated locally bright and approximately circular regions with diameters 0.5 to 2.5 cm. Peaks in the MaxDoG image are detected in step 6040. Peaks correspond to the pixel locations locally bright regions in the MaxDoG image. These locations are grouped into connected objects in step 6050 which become the initial detections of the density detector. Screening Features from these connected objects are computed in step
  • the Screening Classifier's primary purpose is to reduce the number of initial detections prior to more detailed subsequent analysis while retaining malignant regions for further consideration. Detections passing the screening classifier are input to a region growing step 6080.
  • Region growing provides more accurate estimate of the shape of the lesion. Region growing is a crucial step of the invention. Many methods have been described for region growing in mammography implying that it is a difficult problem.
  • the region growing method described in this invention is very general and advances the state of the art in two primary areas: adaptive weighting at each iteration of the growing, and classifier based selection of the best region grown mask, designed using a database of hand truthed lesions.
  • the region grown objects are analyzed in the Density Expert Filter Feature Computation step 6090.
  • features are computed for the purpose of separating suspicious and not suspicious regions. These density features are analyzed in an Expert Filter step 6100 to remove those objects most obviously not associated with cancer.
  • the detections remaining after the expert filter step are input to the Density Feature Computation step, 7000.
  • the output of the density detector includes density feature vectors and locations of detected densities, 7100.
  • the subsampled digital image 2070 and the breast mask 2040 are input to a preprocessing step 6010.
  • the image is first prepared for subsequent processing to reduce the number of false detections due discontinuities associated with the breast and mammographic film border edges. Large discontinuities cause ringing when filters operate on an image. Ringing can cause false detections and may obscure lesions, causing them to not be detected.
  • the density preprocessing step 6010 reduces these undesirable effects.
  • Density preprocessing is used to reduce ringing on the edge of the breast due to the sharp discontinuity in intensity between the background and the tissue. Preprocessing begins with the formation of two masks, the dilated breast mask and the eroded breast mask. Trend correction is then applied to the region between the eroded and dilated breast masks. This operation maintains the detail in the image and eliminates the local background brightness. The local mean is calculated at the edge of the dilated mask and this value is added to the local trend corrected region. This produces an image that maintains detailed information in the transition between breast and background while reducing the discontinuity in brightness between breast and background.
  • a trend correction step 6020 is applied to the image.
  • the trend correction step uses a standard Bezier surface fitting algorithm. Since the Bezier surface is a bi-cubic, parametric patch, 16 control points are used to characterize a surface. These 16 control points are organized in a 4 x 4 grid, and the corresponding parametric surface is used as an approximation to the background image intensities in the local portion of the image. Each pixel intensity within the boundary of the 4 x 4 grid has subtracted from it an approximation to the background intensity value resulting in a density preprocessed image for input to the next step.
  • the density preprocessed image is applied to the MaxDoG filterbank, 6030.
  • the MaxDoG filter bank shown in Fig. 8, first convolves the input image with a bank of DoG filters.
  • Each DoG kernel is individually convolved with the input image, producing five DoG filtered images, each image most sensitive to a particular target size. Then, the maximum response at each pixel location is selected across all of the DoG filtered images to create the MaxDoG output image.
  • DoG kernel is a rough approximation to how a cancer density appears on a mammogram
  • the response of convolving such a kernel with an image indicates regions of the image that resemble the kernel, and therefore resemble a cancer density.
  • DoG kernels respond strongest when the density is approximately the same size as the DoG; hence, a bank of five differently sized DoGs is used. In this way, densities of different sizes spanning a diagnostically relevant size interval are detected.
  • the MaxDoG image is input to a peak detector step, 6040.
  • peaks in the MaxDoG image that may be cancer densities are located for further processing. The steps are as follows. First, for each MaxDoG image pixel location, the peak detector examines all pixels which are three pixels away from that location.
  • the current pixel's intensity value is greater than 90% of the pixels a distance of three pixels away, the current pixel is set to the logical value of ' 1 ' on a local maxima image at the output of the peak picking stage. Otherwise the output image from peak picking is set to the logical value of '0'.
  • the next step, 6050 groups connected pixels with values of T in the local maxima image into logical "objects". This is accomplished with a neighborhood algorithm as follows.
  • the local maxima image is searched in raster order (from the top left of the image to the bottom right of the image, row by row) for pixels indicated to be local maxima.
  • For each local maxima pixel found the neighboring pixels above it and to the left of it are examined. If neither neighbor pixel is part of an object, then a new object is created with the local maxima pixel as its first member. If one of the neighbors is already part of an object, the local maxima pixel is placed in that object. If both neighbors are part of an object, then their objects are combined and the local maxima pixel is placed in the resulting object.
  • the algorithm has generated an ordered list of N objects on the local maxima image. Objects are identified by the set of pixel locations associated with each object index 1 to N.
  • the next stage of the density detector eliminates many of the regions initially detected by the density detector which are not likely to be cancerous regions. Quantitatively, the number of detections is reduced from often over 100 to some predetermined limit, here set to 20. Several steps contribute to this reduction. The first is to remove all detections located in the pectoral muscle, using the pectoral mask, 2060. Next, only a predetermined number of detections with the highest responses in the MaxDoG image are kept.
  • the final step in this stage involves calculating density screening features and classifying the detections, keeping no more than a second predetermined number of detections that pass the quadratic screening classifier.
  • Inputs to the screening classifier are the surviving detected objects and associated set of density screening features.
  • the designator "PT" in a feature's description indicates that the power transform is applied to the measurement.
  • the density screening features in the preferred embodiment are:
  • the density screening classifier uses two discriminant functions, one each for normal and abnormal. After the discriminants are computed, the difference is found. This difference is compared to a predetermined threshold value. If the difference in discriminants is less than the threshold, the detected object is removed from further consideration. Those regions that have a difference of density screening discriminants greater than the threshold are ranked by the difference of discriminants, such that the largest difference of discriminant has the smallest rank value, i.e. 1. Any detected object that has a rank greater than yet another predetermined upper limit is eliminated from further consideration. The remaining detected objects are passed to the next stage for subsequent processing.
  • Each remaining detected object is now input to a region growing step; 6080.
  • the shapes of each detection are more closely estimated.
  • the improved shape estimation is used in the computation of further density features, 7000, for the next stage of classification, step 8000.
  • Shape estimation by region growing is now described. REGION GROWING At this stage of the density detector, region growing is used to estimate the boundary of the detections. Ideally, all detections correspond to suspicious lesions. However, in practice, the steps up to this point detect regions not associated with cancer.
  • a useful method to reject unwanted detections is to compute features based on the shape of a locally bright region.
  • shape estimation is a difficult problem.
  • Region growing is still an active area of research in the medical image processing community. Shape estimation is difficult to automatically accomplish in medical imagery due to the subtlety of lesions, the complexity of their boundaries, and the properties of the background image surrounding the lesion. In many cases the surrounding tissue contains bright structures not related to the lesion.
  • One drawback of a conventional percentage-based floodfill is that it can prematurely terminate if a bright, elongated structure is present in the image chip. This occurs if the floodfill grows into the structure, because the tendency of the floodfill is to follow the bright, elongated structure all the way to the bounds of the image chip.
  • the region growing method described in this invention has two unique aspects.
  • iterative dynamic weighting of the image is used to reduce the effects of background structure on the estimated shape.
  • Iterative dynamic weighting is the process of computing a mixture image as the sum of an input image and its weighted version. The weighting is a function of the current shape estimate.
  • a classifier is used to select the shape most representative of the lesion. The problem of selecting the best shape estimate from a set of estimates created during the iterative process is solved by the use of a "region grow mask selection classifier". This classifier identifies the best shape estimate for the region based on region grow discriminant functions.
  • the set of candidate masks is formed using a floodfill algorithm with iterative dynamic Gaussian weighting to insure that the grown regions are compact and reasonably well formed.
  • the output is a set of selected region grown masks and region grow discriminant values associated with each detection.
  • region growing is applied to detected objects.
  • the detected objects passing the density screening classifier are input to the region growing step, 6080.
  • a block diagram showing an overview of the operations is provided in Fig. 9.
  • Inputs to the region growing step are the subsampled image, a mask of the detected object, and a predetermined maximum allowed region grow neighborhood size.
  • the output is a set of the 'best' Boolean images associated with the lesion.
  • Boolean images represent the shape estimates for the lesion. These Boolean images are known as region grown masks. The number of masks to ultimately select is predetermined as N, where the selection of the best N masks is based on the discriminant values associated with each mask.
  • the method for region growing described in this invention uses two iteration loops.
  • An outer iteration is used to establish a monotonically decreasing threshold.
  • multiple mask estimates are created through the iterative dynamic weighting loop.
  • the discriminant values associated with the region grown masks are also provided as an output for subsequent processing.
  • a pixel location referred to as a seed pixel, is found as the brightest point in the subsampled image in the area of the detection This is the beginning of the process; the first current region grown mask is set to ' 1 ' at the location of the seed pixel.
  • An image chip is then formed by selecting a square neighborhood of twice the maximum region grow neighborhood size centered about the seed pixel.
  • a percentage-based floodfill is successively applied to the image chip to create Boolean masks of the region.
  • Floodfill is the operation of creating a next version of a mask by testing all pixels connected to the current mask, then adding to the current mask the pixels passing the test.
  • a pixel under consideration must 1) have an intensity within x% percent of the seed pixel's intensity and 2) be connected to another pixel already located in the region.
  • the floodfill percentage, x is sequentially increased from predetermined lower to upper limits in predetermined step sizes. In the preferred embodiment, the floodfill percent ranges from 1% to 40% in steps of 1%.
  • One criteria criterion for terminating the floodfill is if the floodfill percentage reaches the upper limit.
  • a second criteria criterion for terminating the floodfill is if the flood filled region for a given percent exceeds the maximum region size. Exceeding the maximum size here is defined as any portion of the mask contacting the edge of the square neighborhood. In practice, the second terminating criteria is vastly more common. Increasing the floodfill percentage allows the region to grow.
  • Controlling the growth is accomplished with the following additional iterative set of steps for each percentage. These additional steps involve creating and applying a circularly symmetric Gaussian weighting kernel dynamically to the image chip pixel values.
  • the Gaussian weighting kernel is found as
  • Gaussian WeightingKernel exp ⁇ (-l/2 ⁇ r 2 ) [( ⁇ '- ⁇ ⁇ ) 2 +(v- ⁇ y ) 2 )] ⁇ (7)
  • ⁇ x and ⁇ y are the x and coordinates of the centroid of the current mask and ⁇ r is three times the maximum distance from the mask's centroid to the edge of the current mask.
  • This window is multiplied with the input image chip to form a weighted image. Then, a mixture image is created for the subsequent floodfill operation as
  • ⁇ 0.5.
  • Fig. 10 details the dynamic Gaussian weighting and iterations within a specific floodfill percent value.
  • a mixture image is formed.
  • the Gaussian weighting kernel is created from the current mask.
  • the weighting is then applied to the original image to form a weighted image, and a mixture image is created from the weighted image and the original image. Region growing is performed in the mixture image.
  • This subset of masks passing the classifier are applied to a pod chopper to remove extraneous tendrils from the core growth.
  • a single erode is performed on the region grown mask to break the connections between the pods and the core.
  • objects are formed from the eroded mask and only the largest object is retained.
  • a mask of the largest object is dilated once within the bounds of the original mask to form the pod chopped version of the region grown mask.
  • a classifier is used to select the best set of region grown masks, therefore, features are required.
  • the features computed from the region grown mask are as follows wherein the Bauer-Fischer ratios referenced below are described in detail in the section for density features.
  • the designator "PT" in a feature's description indicates that the power transform is applied to the measurement.
  • the two areas used to create the feature are the region of the last weighting iteration of the previous floodfill percent and the non- overlapping portion of the region of the first weighting iteration of the current floodfill percent.
  • the two areas used to create the feature are the region of the last weighting iteration of the previous floodfill percent and the non- overlapping portion of the region of the first weighting iteration of the current floodfill percent.
  • the region grown mask classifier determines the set of 'best' shape estimates from the prior region growing step.
  • the region grown mask classifier uses a quadratic classifier to determine the best region grow masks for a detected object. Many region grown masks are generated up to this point and features from those grown regions are input to the classifier. For each mask, the classifier computes discriminants associated with 'best' and 'not best' categories. The discriminants are subtracted, providing region grown mask quality scores. The differences are then rank ordered with the largest difference receiving the rank value of 1. The top N masks are selected based on the ranking.
  • Density features are calculated for each detection from the region grow stage. These features are passed to the density classifier, where only the detections most likely to be cancer densities are kept. Many of the features are contrasts of one type or another, such as intensity contrasts or texture contrasts. The two measures of contrast used in the density detector features are the Bauer-Fischer ratio
  • the subscripts 1 and 2 represent two regions on an input image corresponding, respectively, to an area within a detection and a second area outside the detection. Measurements are computed in each region. In some cases measurements may simply be pixel values from the input image.
  • the area within a detection is defined as all of the points that are part of the region grown mask.
  • the area outside the detection is defined as an inverse mask of the region grown mask, bounded by a box twice as large as the region grown mask (with a minimum of a 1 cm increase in each direction).
  • Some features additionally erode the inside and outside masks so as to not so heavily depend on an exact distinction between inside and outside. In the case of eroded mask areas the inside mask is eroded once; the outside mask is eroded until it is half of its original area. Neither mask is eroded to become smaller than 30 pixels.
  • the limitation of the Fisher ratio is that the relationship between the mean values is lost; there is no information regarding whether ⁇ i is greater or less than ⁇ 2 .
  • the Bauer-Fisher ratio incorporates the sign(.) term to add the relative information to the Fisher ratio.
  • delta features are created by taking the maximum, minimum, mean or standard deviation of a feature value across all N masks passing the region grow classifier. Typically, false positives have fewer "good” masks returned from the region grow than cancer densities, and delta features exploit this difference.
  • an expert filter step, 618 is applied prior to computation of the density features.
  • the expert filter performs bounds checks on expert filter features from all surviving detections. If expert filter features do not fall within the bounds, a region is removed from further consideration. Noting that "best" refers to the region grown mask ranked number 1 from the region grown mask selection classifier, the expert filter features are as follows: 1) circularity of the best region grown mask.
  • expert filter feature 3 is found by first using the E5L5 texture image as the measurement. The feature is found as the standard deviation of the measurement. Features are computed for bjects passing the expert filter. The designator "PT" in a feature's description indicates that the power transform is applied to the measurement. The following eighteen density features computed for subsequent input to the density classifier:
  • I mean of the MaxDoG image under the original detected object (PT). 2) mean ratio of image intensities inside and outside the best region grown mask (PT).
  • I I mean across the top four region grown masks of the mean ratio of the Laws R5S5 image inside and outside the best region grown mask. 12) mean across the top four region grown masks of the Bauer-Fischer ratio of intensity inside and outside the eroded region grown mask (PT).
  • the within-image microcalcification cluster feature vectors, 4100 are provided as inputs to the microcalcification classifier as shown in Figs. 11 and 12 where Fig. 12 is a continuation of Fig 11.
  • These microcalcification cluster feature vectors are applied to two quadratic discriminant functions, G c (x) and G t (x), steps 5010 and 5020.
  • G c (x) is the discriminant function for truthed clusters of microcalcification
  • G t (x) is the discriminant function associated with tissue regions outside truth boxes. Tissue outside a truth box is considered normal.
  • the difference of discriminants, G t (x) - G c (x) is computed in step 5030, to produce a "delta discriminant".
  • step 5040 the delta discriminant is compared to a threshold value, in this embodiment the threshold is 0. If the delta discriminant is less than the threshold, the detection is considered normal and receives the class label of '0' in step 5050. Otherwise, the detection is considered suspicious and receives the class label of ' 1 ' in step 5060.
  • the delta discriminant and the classifier labels are stored in step 5070 for subsequent analysis in the post processing phase 9000.
  • the within-image density feature vectors from the density detector are input to the density classifier as shown in Figs. 13 and 14 where Fig. 14 is a continuation of Fig 13.
  • These density feature vectors are applied to three quadratic discriminant functions, G m (x), G s (x), and G n (x).
  • G m (x), 8010 is the discriminant function for elliptical masses.
  • G n (x), 8030 is the discriminant function for tissue outside the truth box. Tissue outside the truth box is considered normal.
  • G s (x), 8020 is the discriminant function for spiculated masses. Two delta discriminants are then formed. In steps 8040 and 8050, the mass and spiculated mass discriminants are subtracted from the normal discriminant.
  • step 8060 the two delta discriminants are compared to a threshold.
  • the threshold is 0. If both of the delta discriminants are less than the threshold, the detection is considered normal and receives the class label of '0' in step 8070. However, if either of the detections is greater than the threshold, the detection is considered suspicious and therefore receives the class label of ' 1 ' in step 8080.
  • the two density delta discriminants and the class labels are stored in step 8090 for subsequent analysis in the post processing phase 9000.
  • the post processing portion, 9000, of the invention performs three primary functions.
  • the steps are Case Based Ranking, steps 9020 and 9030; Image Context step 9040; and Normalcy Classification, steps 9050 and 9060.
  • Fig. 15 shows the post processing phase of the invention.
  • Inputs to post processing are the within- case microcalcification delta discriminants and the within-case density delta discriminants with class labels of ' 1 '.
  • Within case refers to the set of images input from the same patient from one mammographic screening exam.
  • the class label of a detection may change from a ' 1 ' to a '0'. Only those detections with class labels of ' 1 ' at the output of the post processing phase are displayed to the end user.
  • the first step in post processing is Radial Inhibition, 9010. If more than one density detection has a label of 1 from the density classifier within some local region, only the most cancer like region is retained for consideration.
  • the regions that are affected by this module are those that occur within a one centimeter radius, measured from center to center, of each other.
  • the criteria by which one region is considered more cancerous than another is the sum of difference of discriminants from the feature classifier. The region with the largesr sum of difference of discriminants is considered to be more cancer like and is therefore retained as a potential cancerous region. The region with the smaller sum of discriminants is removed from further consideration.
  • Case Based Ranking In Case Based Ranking (CBR), 910, only the top N detections from the case are retained with no more than M detections retained on any one image. Retain denotes leaving the class label as ' 1 '. Eliminate denotes changing the class label to '0'. The number of detections that retain their classification label depends on how many images are in the case, the distribution across the images, and the category of detections (microcalcification or density).
  • CBR inputs are delta discriminants for density and microcalcification classifiers and class labels for each detection in the case.
  • CBR operates independently on both the microcalcification and density detections.
  • lesion category microcalcification or density
  • the detections are rank ordered by a quality score, largest to smallest.
  • quality scores are delta discriminants in the case of microcalcifications, or by the sum of their delta discriminants in the case of densities.
  • the microcalcification classifier provides a single delta discriminant, normal minus malignant
  • the density classifier provides two delta discriminants: normal minus spiculated, and normal minus mass. In both cases, the larger the quality score, the more likely a detection is to be cancerous.
  • the number of detections to allow per case is determined based on a predetermined number of allowed detections, N, assuming a four film case. When other than four films constitute the case, the number of allowed detections is found as
  • Case based ranking proceeds as follows. First, the number of detections to allow, D, is computed as described above. Then, the delta discriminants are rank ordered, largest to smallest. Detections with rank greater than D have their class labels set to '0'. Once the case detections are ranked to D, a second condition is applied; no more than M detections can have a classification label of ' 1 ' on any one image. If an image has more than M detections with the label equal to ' 1 ', the detections from that image are again rank ordered, largest to smallest by their delta discriminants. Any detection on an image with rank greater than M has its classification label set to '0'. Values for M and N may be set independently based on the category of detection.
  • image based ranking To better appreciate the benefits of case based ranking, first consider the strategy of image based ranking.
  • An example of image based ranking would be to pass the top two ranked detections from each image, the results of which are as shown in Table 2. There are eight surviving detections, two for each image. In this example, one detection is a true positive, for detection index 9. All other detections are false positives.
  • Table 3 shows the benefits of CBR with a limit of six detections per case.
  • ranking the collection of all detections within a case and limiting to the six best, an additional cancer is found with fewer false positives than for the image based ranking method.
  • the cancer on Image 2 is the third "best" detection in the case.
  • Case based ranking allows this detection to be retained, while image based ranking does not.
  • Table 1 Example set of density detections and sum of delta discriminant scores.
  • Image Context Processing Whereas case based ranking examines all the same category detections across all the images of a case, image context processing removes selected combinations of detections across and within detection categories for an image. The motivation for this stage is that certain combinations of detections are observed to predominantly occur on normal images of the training database.
  • the density and microcalcification detections are assigned confidence ratings of high or low, based on comparing the delta discriminant values to appropriate thresholds, wherein the thresholds are obtained from training data.
  • Other embodiments may use more than two confidence values.
  • the number, type, and confidence of the set of detections are recorded. From an analysis of the training data, it was observed that certain combinations of detections occur predominantly on normal images. Therefore, whenever these certain combinations occur, all detections from that image are removed. This step greatly increases the number of unmarked images without unduly lowering sensitivity.
  • Image context quality scores are computed for each detection as follows. For density detections, the delta discriminants are averaged. For microcalcifications, the delta discriminant is used as is. The quality scores are compared to predetermined thresholds to assign confidence ratings to each detection.
  • the predetermined thresholds for each detection category are obtained from an analysis of a training database such that a majority of true positive detections receive a high confidence rating and a majority of the false positive detections receive a low confidence rating.
  • a single low confidence density detection occurred 6 times on images with a malignant mass, 0 times on images with a malignant cluster of microcalcifications, 1 time on an image with both lesion categories, and 324 times on normal images.
  • the combination is one of the following: i) One low confidence density indication (d); ⁇ ) Three low confidence density indications (ddd); iii) One low confidence microcalcification indication and one low confidence density indication (cd); then the images are most often normal. In such instances all detections on the images may be dropped without unduly decreasing sensitivity.
  • Normalcy Classification Normalcy classification refers to the process of examining the discriminants from each category of detection across all the images in the case. When normalcy conditions are satisfied, the set of density and/or microcalcification detections are removed from all images in the case.
  • the "normalcy" condition is assigned by a classifier using quality scores derived from all the detections across a case as input features.
  • the quality scores input to the Normalcy stage are the density and microcalcification delta discriminant values.
  • a further input is the assigned class label for each detection. Separate normalcy classifiers are used for each lesion category.
  • microcalcification normalcy classifier requires two "microcalcification normalcy" features. They are:
  • Microcalcification Normalcy Feature 1 The delta discriminant values from the microcalcification classifier are sorted by image for all the images in the case. The maximum delta discriminant value for each image is stored. Then, the stored, per image delta discriminants are searched across all the images in the case. The maximum from the across image search is used as the first microcalcification normalcy feature.
  • Microcalcification Normalcy Feature 2 The maximum delta discriminant value for each image is stored. Then, the within-image maximums are averaged across all images in the case to provide the second normalcy feature.
  • These two microcalcification normalcy features are used as inputs to the microcalcification normalcy classifier.
  • One embodiment may be a statistical quadratic classifier. The weights of this classifier are designed using a training database of cases of two categories: normal cases and cases containing microcalcifications associated with malignancies. Two discriminant functions are obtained. One discriminant indicates a "normal” case in the microcalcification sense, and the other a "not normal” case in the microcalcification sense. The "not normal” discriminant is subtracted from the "normal” discriminant producing a microcalcification normalcy delta discriminant, or normalcy value. When this normalcy value is greater than a predetermined threshold, all microcalcification detections are removed from the case.
  • the sum of the delta discriminants are computed for every detection on each image. Even detections with class labels of '0' are considered in this step. Noting that each detection includes two delta discriminants (mass and spiculated mass), a sum of delta discriminants is computed for each detection, and the detection on each image with the maximum sum of delta discriminants is identified.
  • Density normalcy features are computed from the delta discriminants.
  • the designator "PT" in a feature's description indicates that the power transform is applied to the measurement.
  • the following normalcy features are computed using the identified per image detections as follows:
  • Density Normalcy Feature 1 Compute the mean of the mass and spiculated mass delta discriminants for the identified detection on each image. Then, searching across all images in the case, select the minimum of the mean delta discriminants as the first feature (PT).
  • Density Normalcy Feature 2 For the identified detection on each image, select the maximum of the two delta discriminants, mass or spiculated mass. Then, computing the mean of the selected delta discriminants as the second feature (PT).
  • Density Normalcy Feature 3 For the identified detection on each image, select the minimum of the two delta discriminants, mass or spiculated mass. Then, computing the mean of the selected delta discriminants as the third feature.
  • One embodiment may be a statistical quadratic classifier.
  • the weights of this classifier are designed using a training database of cases of two categories: normal cases and cases containing malignant densities.
  • Two discriminant functions are obtained.
  • One discriminant indicates a "normal” case in the density sense, and the other a "not normal” case in the density sense.
  • the "not normal” discriminant is subtracted from the "normal” discriminant producing a density normalcy delta discriminant, or normalcy value. When this normalcy value is greater than a predetermined threshold, all density detections are removed from the case.
  • the output of the normalcy stage is a set of detections and associated class labels.
  • the detections with class labels of ' 1 ' are displayed overlaid on the medical images as requiring further interpretation by a radiologist.
  • the present invention provides an improved method and system for identifying detections on digital images which meet criteria indicative of cancerous lesions, and particularly provides a method and system which may be advantageously used to identify cancerous lesions on mammograms obtained from a patient's screening visit.
  • the present invention presents processes for classifying microcalcifications and detections, and for post processing the detections in such a manner that the number of false positive detections is reduced while providing the further advantage of increasing the sensitivity of the method and system, resulting in increased accuracy in identifying cancerous lesions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Quality & Reliability (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

L'invention concerne un procédé et un système de détection conçus pour assister des radiologistes dans la lecture d'images médicales (1000). Lesdits système et procédé sont particulièrement destinés au domaine de la mammographie y compris de la détection de microcalcifications groupées et de densités (3000). Cette invention a trait à un détecteur de microcalcifications, par le biais duquel des détections individuelles sont classées par ordre de grandeur et classifiées (5000), et une des caractéristiques de la classification est dérivée au moyen d'un perceptron multicouche.
EP00944852A 1999-06-23 2000-06-23 Detection assistee par ordinateur de masses et strategies de microcalcifications groupees Pending EP1194880A1 (fr)

Applications Claiming Priority (3)

Application Number Priority Date Filing Date Title
US14063699P 1999-06-23 1999-06-23
US140636P 1999-06-23
PCT/US2000/017410 WO2000079474A1 (fr) 1999-06-23 2000-06-23 Detection assistee par ordinateur de masses et strategies de microcalcifications groupees

Publications (1)

Publication Number Publication Date
EP1194880A1 true EP1194880A1 (fr) 2002-04-10

Family

ID=22492142

Family Applications (1)

Application Number Title Priority Date Filing Date
EP00944852A Pending EP1194880A1 (fr) 1999-06-23 2000-06-23 Detection assistee par ordinateur de masses et strategies de microcalcifications groupees

Country Status (5)

Country Link
EP (1) EP1194880A1 (fr)
JP (1) JP2003502129A (fr)
AU (1) AU5888500A (fr)
CA (1) CA2375314A1 (fr)
WO (1) WO2000079474A1 (fr)

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7783089B2 (en) * 2002-04-15 2010-08-24 General Electric Company Method and apparatus for providing mammographic image metrics to a clinician
GB2395263A (en) * 2002-11-12 2004-05-19 Qinetiq Ltd Image analysis
GB2398379A (en) * 2003-02-11 2004-08-18 Qinetiq Ltd Automated digital image analysis
FR2863749B1 (fr) 2003-12-10 2006-04-07 Ge Med Sys Global Tech Co Llc Procede de traitement d'image radiologique pour la detection de microcalcifications
US8285019B2 (en) 2006-02-10 2012-10-09 Synarc Inc. Breast tissue density measure
GB0602739D0 (en) * 2006-02-10 2006-03-22 Ccbr As Breast tissue density measure
US8218850B2 (en) 2006-02-10 2012-07-10 Synarc Inc. Breast tissue density measure
FR2897461A1 (fr) 2006-02-16 2007-08-17 Gen Electric Dispositif de rayonnement x et procede de traitement d'images
FR2904882B1 (fr) 2006-08-11 2008-11-14 Gen Electric Procede de traitement d'images radiologiques pour une detection d'opacites
JP5801379B2 (ja) * 2010-04-30 2015-10-28 ヴィユーコンプ インクVucomp, Inc. 確率密度関数推定器
WO2012085818A1 (fr) * 2010-12-23 2012-06-28 Koninklijke Philips Electronics N.V. Score de calcium de mammographie
WO2023139970A1 (fr) * 2022-01-19 2023-07-27 富士フイルム株式会社 Dispositif de traitement d'image, procédé de traitement d'image et programme

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5799100A (en) * 1996-06-03 1998-08-25 University Of South Florida Computer-assisted method and apparatus for analysis of x-ray images using wavelet transforms

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See references of WO0079474A1 *

Also Published As

Publication number Publication date
CA2375314A1 (fr) 2000-12-28
JP2003502129A (ja) 2003-01-21
WO2000079474A1 (fr) 2000-12-28
AU5888500A (en) 2001-01-09

Similar Documents

Publication Publication Date Title
US6763128B1 (en) Method for analyzing detections in a set of digital images using case based normalcy classification
AU741337B2 (en) Method and system for automated detection of clustered microcalcifications from digital mammograms
AU705713B2 (en) Method and system for the detection of lesions in medical images
US5999639A (en) Method and system for automated detection of clustered microcalcifications from digital mammograms
US6137898A (en) Gabor filtering for improved microcalcification detection in digital mammograms
Biswas et al. Mammogram classification using gray-level co-occurrence matrix for diagnosis of breast cancer
Singh et al. An approach for classification of malignant and benign microcalcification clusters
Unni et al. Tumour detection in double threshold segmented mammograms using optimized GLCM features fed SVM
WO2000079474A1 (fr) Detection assistee par ordinateur de masses et strategies de microcalcifications groupees
Antony et al. A new approach to determine the classification of mammographic image using K-means clustering algorithm
Sarada et al. Spatial Intuitionistic Fuzzy C-means with Calcifications enhancement based on Nonsubsampled Shearlet Transform to detect Masses and Microcalcifications from MLO Mammograms
Salazar-Licea et al. Automatic segmentation of mammograms using a Scale-Invariant Feature Transform and K-means clustering algorithm
Valliappan et al. A theoretical methodology and prototype implementation for detection segmentation classification of digital mammogram tumor by machine learning and problem solving approach
Constantinidis et al. Evaluating classification strategies for detection of circumscribed masses in digital mammograms
Novitasari et al. Automatic detection of breast cancer in mammographic image using the histogram oriented gradient (HOG) descriptor and deep rule based (DRB) classifier method
Babu et al. Quantitative analysis of digitized mammograms using nonsubsampled contourlets and evolutionary extreme learning machine
Sahba et al. Mean shift based algorithm for mammographic breast mass detection
Styblinski et al. Circuit performance variability reduction: principles, problems, and practical solutions
Shili et al. An efficient model based on spatial fuzzy clustering and region growing for the automated detection of masses in mammograms
Conant‐Pablos et al. Feature Selection for the Classification of Microcalcifications in Digital Mammograms Using Genetic Algorithms, Sequential Search and Class Separability
Patrocinio et al. Classifier scheme for clustered microcalcifications in digitized mammograms by using Artificial Neural Networks
Choraś Shape and texture feature extraction for retrieval mammogram in databases
Licea et al. Location of mammograms ROI’s and reduction of false-positive
Rajkumar et al. Detection and Classification of Abnormal Mammogram Images Using Lazy Classifiers
AU7737801A (en) Method and system for automated detection of clustered microcalcifications from digital mammograms

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE

17P Request for examination filed

Effective date: 20020111

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LI LU MC NL PT SE

AX Request for extension of the european patent

Free format text: AL;LT;LV;MK;RO;SI