WO2010079519A1 - Procédé et système de reconnaissance automatique de lésions dans un ensemble d'images de résonance magnétique d'un sein - Google Patents

Procédé et système de reconnaissance automatique de lésions dans un ensemble d'images de résonance magnétique d'un sein Download PDF

Info

Publication number
WO2010079519A1
WO2010079519A1 PCT/IT2009/000005 IT2009000005W WO2010079519A1 WO 2010079519 A1 WO2010079519 A1 WO 2010079519A1 IT 2009000005 W IT2009000005 W IT 2009000005W WO 2010079519 A1 WO2010079519 A1 WO 2010079519A1
Authority
WO
WIPO (PCT)
Prior art keywords
voxels
dataset
breast
region
images
Prior art date
Application number
PCT/IT2009/000005
Other languages
English (en)
Other versions
WO2010079519A9 (fr
Inventor
Anna Vignati
Valentina Giannini
Diego Persano
Lia Morra
Alberto Bert
Original Assignee
Im3D S.P.A
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 Im3D S.P.A filed Critical Im3D S.P.A
Priority to EP09787626A priority Critical patent/EP2386102A1/fr
Priority to PCT/IT2009/000005 priority patent/WO2010079519A1/fr
Publication of WO2010079519A1 publication Critical patent/WO2010079519A1/fr
Publication of WO2010079519A9 publication Critical patent/WO2010079519A9/fr

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
    • G06T7/0014—Biomedical image inspection using an image reference approach
    • G06T7/0016—Biomedical image inspection using an image reference approach involving temporal comparison
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/40—Analysis of texture
    • G06T7/41—Analysis of texture based on statistical description of texture
    • G06T7/44—Analysis of texture based on statistical description of texture using image operators, e.g. filters, edge density metrics or local histograms
    • 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
    • G06T2207/10016—Video; Image sequence
    • 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
    • G06T2207/30068—Mammography; Breast

Definitions

  • the present invention relates to a method of identification of objects and/or structures in magnetic resonance images.
  • the invention ' relates to a digital segmentation method of the contents of an image of a breast, as defined in the preamble of claim 1.
  • MRI breast magnetic resonance imaging
  • a dataset comprising a plurality of tomographic images, each of them representing an axial, coronal or sagittal section of the patient's chest is acquired, hi magnetic resonance images, the grey level intensity of each tissue depends on its chemical structure, in a way which reflects the specific acquisition protocol used.
  • DCE-MRI dynamic contrast enhanced MRI
  • a contrast agent bolus is injected intravenously, and a plurality of datasets are acquired at different time instants, in order to extract both anatomical and functional information.
  • the presence of the contrast agent increases the local signal intensity, whereas the distribution of the contrast agent depends on the microcirculatory properties of each tissue.
  • the contrast uptake is characterized by a kinetic as a function of the time, which is typically represented as a curve.
  • the shape of this curve depends on the characteristics of the vascular systems of the lesions, which are known to be related with characteristics of the lesions themselves such as malignancy, prognosis and metastatic spread.
  • Various acquisition protocols can be used to achieve this goal, with different results in terms of intensity distribution and signal-to-noise ratio.
  • segmentation refers to the partitioning of an image into multiple regions in order to locate specific objects and/or features.
  • Object of the present invention is therefore to provide a method of segmentation which is capable of segmenting in an efficient, accurate and fully automatic manner the lesions present in the datasets of images of a breast, thus overcoming the limitations of the prior art.
  • FIG. 1 is a flow chart of the operations performed by a method according to the invention.
  • FIG. 1 - figures from Ia to Ie show altogether flow charts of the steps performed in different operations of figure 1; - figure 2 is a schematic image of a breast section;
  • FIG. 3 is a schematic lateral view of breasts
  • - figure 4 is an example of a breast atlas
  • figure 5 represents typical trends of the contrast uptake in voxels belonging to malignant and benign lesions, as well as the typical average contrast uptake of an entire lesion; and figure 6 is a diagrammatic representation of a processing system for the implementation of the method according to the invention.
  • a method according to the invention starts, at step 200, with the acquisition of magnetic resonance image datasets of a portion of a patient's body which comprises at least one breast, said portion being referred as sampled region.
  • a dataset may comprise one ore more images.
  • a reference coordination system is defined wherein a z axis is the longitudinal axis of the patient, and is oriented from the feet to the head, and the images represent sections of breasts on planes perpendicular to the z axis.
  • Said planes are each defined by a couple of perpendicular axes denoted x and y in the following, wherein the x axis is the lateral axis, and is oriented from the left to the right of the patient, and the y axis is the anterior-posterior axis, and is oriented from the back to the belly of the patient.
  • Step 200 comprises two succeeding steps: at step 200a a first dataset is acquired at a first time instant t ⁇ said first dataset being in the following called pre-contrast frame. Subsequently, a contrast agent is administered to the patient by means of an intravenous injection and after that, at step 200b, a plurality of datasets in succeeding time instants t l5 t 2 , ..., t N is acquired, said datasets being in the following altogether called post-contrast frames.
  • a schematic image of a breast section is represented, such image comprising, in a known manner, a plurality of voxels, some of which have been represented as minuscule squares indicated as a whole by the reference 2.
  • Such voxels 2 have intensity values related to the properties of the corresponding tissue of the sampled volume.
  • the reference 4 indicates the contour of the body of the patient
  • the reference 6 indicates two breasts
  • the reference 8 indicates a ribcage area, which comprises the ribcage and internal organs such as the lungs and the heart
  • reference 10 indicates a region of interest which is the portion of clinical interest, said region of interest 10 including the breasts 6 and axillary regions 12.
  • a preliminary anatomic segmentation is performed in a step 210 in order to approximately locate the breasts 6 and determine their approximate size. More specifically, the approximate location is obtained by locating an infra-breast zone 14 (see figure 2) which is a zone of the sinus placed between the two breasts 6. This zone extends at most between a first breast terminal point 16 and a second breast terminal point 18.
  • a reference point 20 is considered of said infra-breast zone 14, said reference point 20 being a point of the segment extending between the first breast terminal point 16 and the second breast terminal point 18, this reference point 20 being selected as disclosed herein below.
  • said infra-breast zone 14 is a strip that corresponds to the skin placed on the sternum 21. According to the preferred embodiment above disclosed, only a line is considered of said strip, this line passing through the reference points 20 of each image.
  • the voxels 2 belonging to said infra-breast zone 14 are identified by recognizing voxels 2 placed between the breasts 6 and belonging to a surface external with respect to the contour 4 of the body of the patient as herein below disclosed.
  • Figure Ia shows a flow chart of the steps performed to carry out said step 210.
  • step 210 it is first necessary to individuate the contour 4 of the body of the patient, i.e. the interface between the body of the patient and the air external to it.
  • the contour 4 of the body of the patient i.e. the interface between the body of the patient and the air external to it.
  • the contour 4 of the body of the patient is based on the removal of the air external to the body of the patient.
  • a combination of all post-contrast frames is calculated in order to take into account the whole dynamic series of datasets so as to minimize the effect of the acquisition noise due to presence of the air external to the patient's body.
  • a new intensity value which is a predetermined combination function of the intensity values of the respective voxels 2 in each post- contrast frame.
  • the combination of the post-contrast frames is in the following denoted as "intensity projection” and the dataset of images obtained after said combination is called projected frame.
  • the combination function is for instance the calculation of the mean of the intensity values, i.e. the intensity value of each voxel 2 of the projected frame is the mean of the intensity values of the respective voxels 2 in each post-contrast frame, thus obtaining a "mean projected frame”.
  • other types of combination functions can be used, for example the calculation of the minimum, or of the maximum or the median of the intensity values of the voxels 2.
  • a first projected frame is calculated as above disclosed.
  • a working region 22 (see figure 3) is extracted from the first projected frame by removing images that correspond, along the z axis, to portions of the sampled region too far from the breasts 6.
  • the images corresponding to the upper 30% and to the lower 10% along the z axis of the sampled region are removed.
  • transition zones between different structures contained in the images of the working region 22 of the first projected frame are identified by means of an edge detection filtering (e.g. gradient, laplacian, Sobel).
  • an edge detection filtering e.g. gradient, laplacian, Sobel.
  • a smoothing operation on the images is performed before applying the edge filter.
  • Image smoothing allows to attenuate high frequency contents of the image, thus enhancing strong phase transitions.
  • a gaussian smoothing filter is applied, for example with a standard deviation of 2.0mm, and a gradient magnitude filter is subsequently applied.
  • a main edge is identified within the plurality of edges found at step 213, said main edge substantially corresponding to the contour 4 of the body of the patient.
  • external air with respect to the patient's body is identified by means of a region growing procedure starting from voxels 2 placed out of the body of the patient, and therefore belonging to the external air.
  • a confidence connected region growing procedure is used, said procedure starting with the selection of a seed region comprising one or more voxels 2 placed out of the body of the patient.
  • the voxels 2 that are for example selected are the voxels 2 placed in the top left corner of the images.
  • the edge surface not recognized as air and having the biggest area corresponds to the transition between the air and the patient's body.
  • a threshold is determined, for example by means of the Otsu method, and is then applied to the images of the working region 22 of the first projected frame, so as to select voxels 2 having intensity values higher than a threshold value representative of the air.
  • the interface between the selected voxels 2 and the other remaining voxels 2 of the image is the contour 4 of the body of the patient.
  • the line which passes through the reference point 20 and represents the infra- breast zone 14 is calculated as herein below disclosed.
  • Each image of the working region 22 of the first projected frame is scanned along the -y direction, said scanning step beginning from the upper part of the image (in the y direction) in correspondence of the central point of the image along the x direction.
  • the first point belonging to the contour 4 of the body of the patient reached in each image is the reference point 20, which has respective x and y coordinates.
  • the line passing through the reference points 20 of each image represents the infra-breast zone 14.
  • step 216 the scanning of the images of the working region 22 of the first projected frame is repeated, said scanning starting from all the points of the upper part of the images, along the x direction.
  • the point having the maximum y coordinate (considering the y axis starting from the bottom part of each image) is denoted maximum breast point 24 and represents the point of maximum extension of the breasts 6.
  • a registration procedure is performed between the post-contrast frames and the pre-contrast frame in order to correct possible misalignments among the frames composing the dynamic time series.
  • Such misalignments can be caused by patient movements, breathing, cardiac motion and so forth. In this way, an alignment of all the post-contrast frames with the pre-contrast frame is performed.
  • the registration step 220 is performed for all voxels 2 preferably belonging to a working area 26, for all the frames or alternatively for only the images belonging to the working region 22.
  • the working area 26 is an area of each image smaller than the area of the whole image and that comprises the breasts 6. Said working area 26 is defined by fixing, in each image, a starting point having predetermined x and y coordinates. Preferably, said starting point is the reference point 20, i.e. in each image of the pre- and post contrast frames the voxel 2 having x and y coordinates equal to the coordinates of the reference point 20 is considered as the starting point.
  • the working area 26 is created, for example, as a rectangular area delimited by a first line 28, parallel to the x axis, which runs above (in the y direction) the starting point of a first predetermined distance and a second line 30, parallel to the x axis, which runs below said starting point of a second predetermined distance.
  • the term above and below are herein referred to the figure 2 and indicate respectively positions having a y coordinate higher or lower.
  • the working area 26 has a lateral extension, along the x direction, equal to the lateral dimension of the images.
  • the registration step 220 preferably comprises three sub-steps: a translation, a rigid-body transformation and a non-rigid transformation.
  • a predetermined cost function based on an image similarity measure is used to compare the images, said function being the same for all the sub-steps or alternatively being different in each sub-step.
  • image similarity measure in particular the method specified by Mattes et al., "Non-rigid multimodality image registration", Medical Imaging 2001: Image Processing, pp. 1609- 1620, 2001.
  • Non-rigid transformation is the Free-Form Deformation (FD) model based on B-splines, as proposed in Ruecker et al., "Non-rigid registration using free-form deformations: application to breast MR images", IEEE Transactions on Medical Imaging, 18(8):712-721, 1999.
  • FFD Free-Form Deformation
  • the respective cost functions are optimized by means of known optimizers.
  • a gradient descent optimizer is used for the translation and the rigid-body transformation
  • the LBFGSB Lited memory - Broyden, Fletcher, Goldfarb and Shanno- for Bound constrained optimization
  • the images of the pre- and post-contrast frames have an original resolution along the x, y and z directions, hi order to reduce the computing time, the following steps are performed.
  • the images are down-sampled to said predetermined resolution. This is done according to the direction or directions in which the original resolution is higher than the predetermined one. If the pre- or post-contrast frames present original resolution lower than said predetermined resolution, along the x, y and z directions, the registration is performed at the original resolution.
  • said predetermined resolution is the same along the x, y and z directions.
  • the output of the registration step 220 is a first deformation field, one for each post- contrast frame, each first deformation field being a vector field wherein the vectors allow to match homologous points in the pre- and post-contrast frames. If the post-contrast frames were down-sampled before the registration, the respective first deformation fields are up-sampled to the original resolution.
  • the working areas 26 of the original post-contrast frames are warped by applying to each post-contrast frame the respective first deformation field, i.e the voxels 2 are moved so that anatomical structures in the dataset have spatial distribution and relative spatial position corresponding to that of the pre-contrast frame, thus obtaining aligned post-contrast frames.
  • an interpolation is used.
  • the B-spline interpolation is used.
  • said registration step is omitted and all the subsequent steps according to the invention are applied on the post-contrast frames.
  • an accurate anatomical segmentation is performed in order to more precisely extract different anatomical structures.
  • the object of this step is to identify the voxels 2 belonging to the region of interest 10.
  • This object is achieved by adapting a breast atlas to the specific patient, more specifically by performing a registration between said atlas and a starting frame related to the pre- or post-constant aligned frames.
  • said starting frame is the pre- contrast frame.
  • said starting frame is a second projected frame, different from the first projected frame and calculated as above disclosed by using the post-contrast aligned frames.
  • the registration between the atlas and the starting frame is performed by either warping the atlas or by warping the starting frame.
  • Said atlas is a reference dataset which represents a typical breast MRI dataset and is used to align the starting frame to a common reference anatomical space.
  • Said atlas is therefore a set of images wherein each voxel 2 has a known intensity value.
  • the reference dataset can be a simplified, piece- wise constant representation of a real dataset, i.e. each voxel 2 of the atlas has a predetermined intensity value depending on the anatomical structure to which it belongs, where said intensity value is related to the average intensity observed in typical breast MRI datasets.
  • the structures included in the reference dataset are not limited to the structures to be segmented, in order to provide sufficient information to successfully align the starting frame with the atlas.
  • the known intensity values could be obtained by calculating the average of corresponding intensity values of images of real datasets, said real datasets being previously registered with one another.
  • a smoothing filter is applied to said real datasets to remove unnecessary details.
  • the atlas is associated to a probability dataset, i.e. a set of images in which to each voxel 2 is associated at least one probability coefficient indicative of the probability that the voxel 2 belongs to one or more predetermined anatomical structures.
  • Each voxel 2 may have either one discrete coefficient indicating whether the voxel 2 belongs to a predetermined anatomical structure or not, or a plurality of probability coefficients, each coefficient indicating the probability that the voxel 2 belongs to a respective anatomical structure.
  • the probability coefficients are determined with reference to the region of interest 10, i.e. they represent the probability that a voxel 2 belongs to the region of interest 10.
  • FIG 4 an example of an image of a breast atlas is shown, representing a simplified version of a typical breast dataset.
  • Figure Ib shows a flow chart of the steps performed to carry out said step 230.
  • Said step 230 is carried out only for the voxels 2 of the starting frame belonging to a working area, i.e. voxels 2 having x and y coordinates comprised in the area denoted working area 26.
  • a predetermined atlas is selected as a function of the distance between the infra-breast zone 14 and the maximum breast point 24, said atlas being selected in a set of atlases corresponding to different breast dimensions.
  • three atlases are defined, each having associated a respective probability dataset.
  • the starting frame is down-sampled to remove unnecessary information and reduce the computational burden.
  • the starting frame is also smoothed, for example by means of a gaussian or median filter, in order to remove unnecessary details and noise. Alternatively, this step may be omitted.
  • a registration between the atlas and the starting frame is performed, as disclosed with reference to step 220.
  • the starting frame is kept fixed and the atlas is deformed.
  • the atlas is kept fixed and the starting frame is deformed.
  • the registration 233 begins with a translation step in which an infra-breast area of the atlas, whose position is known because it is determined during the creation of the atlas itself, is aligned with the infra-breast area 14 of the starting frame.
  • the infra-breast area comprises voxels 2 having the same coordinates as those of the voxels 2 of the infra-breast area 14 calculated at step 215.
  • the output of this step is a second deformation field.
  • the probability dataset is used in order to segment the region of interest 10 as herein below disclosed.
  • the probability dataset is aligned to the starting frame by warping it with the second deformation field obtained at step 233.
  • this step is omitted.
  • each voxel 2 of the working area 26 of the starting frame is associated a probability value representing the probability that said voxel 2 belongs to the region of interest 10, said probability value being equal to the probability coefficient of the respective voxel 2 of the probability dataset.
  • a voxel 2 is assigned to the region of interest 10 if its associated probability is higher than a predetermined threshold (e.g. 0.5).
  • the probability dataset can be supplied as a priori probabilities to a known probabilistic classifier, such as an Expectation Maximization Model, arranged to classify each voxel 2 of the working area 26 of the starting frame by using features derived from the intensity value of the voxel 2 and the a priori probability that the voxel 2 belongs to the region of interest 10.
  • a known probabilistic classifier such as an Expectation Maximization Model
  • the result of step 234 is a binary mask in which a value of 1 or 0 is associated to each voxel 2 according to the fact that the voxel 2 belongs to the region of interest 10 or not, respectively.
  • At the subsequent step 240 at least one normalization factor is calculated in order to correct effects due to different acquisition modalities, different types and amounts of injected contrast agent and other possible factors like for example the technical characteristics of the scanner, which result in variations of image intensities between scanners, patients or even between different datasets from the same patient.
  • said normalization is based on the intensity values of anatomical structures always present in the field of view, such as the aorta 32 or the mammary arteries 34 (see figure 2). Firstly, the anatomical structures need to be identified and segmented; then, a normalization factor is extracted from the contrast uptake of the voxels 2 belonging to said anatomical structures.
  • the mammary arteries 34 are employed, and the normalization is performed according to the steps shown in figure 1 c.
  • Figure Ic shows a flow chart of the steps performed to carry out said step 240.
  • a combination of the pre- and post-contrast aligned frames is calculated in order to enhance the intensity of said arteries 34, thus obtaining an intermediate frame.
  • the arteries 34 comprise voxels 2 whose intensity values are increased by the presence of the contrast agent.
  • Such intermediate frame could be obtained by assigning to each voxel 2 an intensity value equal to the subtraction between the intensity value of the respective voxel 2 in a predetermined post-contrast aligned frame and in the pre-contrast frame.
  • at each voxel 2 is assigned an intensity value equal to the maximum intensity value of the respective voxel in the post-contrast aligned frames.
  • an infra breast-sternum area 36 is identified for each image of the intermediate frame (see figure 2).
  • This is an area comprised between the infra-breast zone 14 and sternum 21 and that includes said mammary arteries 34.
  • this is a rectangular area delimited by a first line, parallel to the x axis, which passes by the intra-breast zone 14, and a second line, parallel to the x axis, which runs below (along the y direction) said first line of a first predetermined quantity chosen so as to include the mammary arteries 34.
  • the area is further delimited by two other lines, parallel to the y axis and placed at a predetermined distance, along the x axis, as to include only a predetermined central portion of the image, said central portion including the mammary arteries 34.
  • a rectangle of 35mm x 100mm is considered by setting the y coordinate of a reference point, determined as disclosed with respect to the reference point 20 at step 215, as the midpoint of the higher side (along the y direction) of the rectangle itself.
  • the mammary arteries 34 are identified. This is done by performing, in known manner, a vessel recognition procedure so as to identify tubular structures within the voxels 2 belonging to an infra breast-sternum region of the intermediate frame, said infra breast-sternum region being the tri-dimensional set of the infra breast-sternum area 36.
  • a Vessel Enhancing Filter is applied to the voxels 2 of the infra breast-sternum region.
  • other method can be used to perform said identification.
  • the output of said Vessel Enhancing Filter is a measure of the "vesselness" of the local structure of each voxel 2, to which a vessel can be approximated.
  • Voxels 2 whose vesselness value exceeds a predetermined threshold are classified as belonging to the mammary arteries 34.
  • said threshold is derived from the histogram of the result of the vesselness filter for all the voxels 2 belonging to the infra breast-sternum region 36. Examples of such threshold are half the maximum vesselness value, or the vesselness value corresponding to a predetermined percentile, such as the 95 th percentile.
  • step 243 voxels 2 belonging to the mammary arteries 34 are recognized.
  • the normalization factor is calculated based on the voxels 2 having the same coordinates as those of the voxels 2 belonging to the mammary arteries 34, either in the intermediate frame or in the second projected frame.
  • Such normalization factor is, for example, the mean of the intensity values of the voxels 2.
  • a plurality of normalization factors is calculated, each factor being calculated with reference to a different post-contrast aligned frame.
  • step 250 the post-contrast aligned frames are processed in order to segment regions 40 corresponding to potential lesions.
  • references 500a and 500b indicate curves of the steady enhancement type, in particular 500a is a straight line while 500b is a curve.
  • Reference 502 indicates a curve having a plateau of signal intensity, and reference 504 indicates a curve having a washout of signal intensity.
  • Curves 500a and 500b are indicative of benign lesions, curve 502 is indicative of possible malignancy, and curve 504 strongly suggests malignancy.
  • curves 500a, 500b, 502 and 504 are referred only to individual voxels 2 or sets of contiguous voxels 2 (typically formed by a few voxels) belonging to a single tissue having uniform vascular characteristics, and thus having uniform kinetics of the contrast uptake, whereas an average intensity curve calculated over an entire lesion (typically not having homogeneous vascular characteristics) is not necessarily discriminative.
  • curve 506 is an example of said average intensity curve; as can be noted, curve 506 is quite different from the curves 500a, 500b, 502 and 504.
  • Figure Id shows a flow chart of the steps performed to carry out said step 250.
  • At step 251 at least one working dataset is calculated in order to exploit the information contained in the whole dynamic sequence acquired.
  • Such working dataset comprises voxels 2 to which are associated their respective intensities in the pre-contrast and post-contrast aligned frames.
  • voxels 2 At each voxel 2 is associated a signal intensity curve representative of the intensity variation as a function of the time.
  • a combination of said post-contrast aligned frames for example a maximum, median or mean intensity projection, is calculated and the working dataset comprises voxels 2 to which are associated their respective intensities in the intensity projection after subtracting their respective intensities in the pre-contrast frame.
  • the subtraction is done in order to neglect the contribution of portions of the sampled region that do not show a contrast uptake after contrast injection.
  • the anatomical mask obtained at step 234 is applied to said working dataset, so that the following steps are executed only for the voxels 2 belonging to the region of interest 10.
  • Step 252 not only reduces the computational burden, but also avoids generating false positives (such as the heart, vessels or other regions comprised in the ribcage area 8) which are placed out of the region of interest 10.
  • a denoising filter such as median, an anisotropic diffusion or a gaussian filter is applied to said working dataset. Alternatively, this step is omitted.
  • the intensity of each voxel 2 of the region of interest 10 of said working dataset is divided by one of the normalization factors obtained at step 240 thus obtaining a normalized working dataset.
  • the post-contrast aligned frames are normalized before calculating the working dataset.
  • each voxel 2 of the normalized working dataset is classified in one of two classes, which are potential lesion, or other.
  • the classification must be performed in such a way so as to detect most of the voxels 2 showing a contrast uptake, and hence achieve a high sensitivity, while excluding the voxels 2 belonging to vessels. Since lesions are often connected to the supplying vessels, it is important that vessels are excluded at this step, as they might be segmented together with the lesions; over-segmentation could limit the performance of the subsequent steps and reduce the diagnostic quality of the segmentation.
  • n variables are identified which characterize potential lesions, hereinafter termed "features”.
  • a ⁇ -dimensional vector is associated to each voxel 2 of the normalized working dataset, such n-dimensional vector containing the values of these features for the voxel 2 itself, hi the following, said vectors of variables will be denoted "voxel feature vectors”.
  • the features are extracted from the normalized working dataset and are indicative of the intensity (i.e. of the contrast uptake) and of the shape of the anatomical structure to which each voxel 2 belongs.
  • Examples of features related to the intensity are the intensity value of the voxels 2 itself or of the neighbouring voxels 2, for example the first 26 voxels.
  • Examples of features indicative of the shape are the vesselness values obtained by applying a vessel recognition procedure as above disclosed.
  • the voxel feature vectors belonging to each voxel 2 of the region of interest 10 are then supplied to a known classifier which establishes to which class the corresponding voxel 2 belongs, as described hereinafter.
  • a known classifier which establishes to which class the corresponding voxel 2 belongs, as described hereinafter.
  • the classifier can operate directly on the voxel feature vectors of the voxels 2.
  • the classifier can be based on the combination of rules and hypotheses derived from a priori knowledge of the properties of lesions and normal tissues with rules drawn from the analysis of the voxel feature vectors themselves.
  • An example of a classifier which operates on the basis of the voxel feature vectors distinguishes the voxels 2 on the basis of a specific threshold for each feature.
  • fixed thresholds T 1 and T 2 can be associated respectively to the intensity and to the vesselness values of the voxels 2 of the normalized working dataset: voxels 2 with intensity value higher than the first threshold T 1 and vesselness value lower than the second threshold T 2 are classified as belonging to potential lesions.
  • said classifier is constituted by a neural network with a specific structure and with parameters derived from the features of all voxels 2 through known methods of unsupervised training.
  • classifiers of the k-means and fuzzy c-means type include classifiers of the k-means and fuzzy c-means type.
  • classifier which integrates a priori knowledge regarding the characteristics of breast lesions is a classifier of the fuzzy type, which classifies each voxel 2 of the region of interest 10 according to predetermined rules.
  • the output of the classifier is a final dataset in which each voxel 2 of the normalized working dataset is either classified as belonging to a potential lesion or not. Said final dataset is further processed in order to extract regions 40 representing potential lesions, constituted by connected voxels 2 classified as belonging to a potential lesion.
  • the regions 40 may include not only malignant and benign lesions, but also vessels, lymph nodes and different kinds of false positives (such as motion artifacts and noise). Therefore, it is necessary to exclude from said detected regions 40 all the structures which are not of clinical interest, i.e. to reduce the false positives.
  • a vector of m features derived from the dimension, the kinetic, the position and the morphology of the region- 40 is associated to each region 40.
  • the aim is to classify each region 40 in two or more classes according to its clinical interest.
  • two classes are defined: lesions and non-lesions.
  • Non-lesions can be further discriminated according to their nature (such as vessels, lymph nodes and other false positives).
  • the features included in the region feature vector differ from those included in the voxel feature vectors as they characterize the global properties of a breast lesion as a whole, rather than the individual properties of each voxel 2 belonging to said breast lesion. This two-steps classification allows to exploit both high and low-level features of breast lesions in order to reduce the number of false positives.
  • An example of morphological feature is a vesselness measure of the region 40.
  • a vesselness value is calculated for each voxel 2 as above disclosed, and then a single value for the feature is extracted combining the vesselness values of all voxels 2 of the region 40, for example by calculating the mean.
  • Kinetic features characterize time-signal intensity curves in order to identify trends different from those depicted in figure 5.
  • An examples of such features is the standard deviation of the mean of the intensity values of the voxels 2 having coordinates corresponding to those of the voxels 2 of each region 40 at each acquisition time (of the post-contrast aligned frames), or the ratio between the mean of the intensity values of said voxels 2 at a predetermined time instant t N and the corresponding value calculated at the preceding time instant t N _,.
  • a classifier is applied to said region feature vectors in order to distinguish the various classes of regions 40.
  • the classification criteria are determined for example from a priori knowledge of the characteristics of lesions, or by a comparison between the region feature vectors and corresponding vectors associated to known lesions extracted from a set of real exams.
  • An example of said classifier classifies the regions 40 applying a predetermined threshold to each feature of the region feature vector, where the thresholds can be fixed based on a priori knowledge of the characteristics of the lesions, or extracted from the vectors of the set of real exams.
  • An example of threshold derived from a priori knowledge is a fixed threshold on the volume of the regions 40, as very small regions are probably non-lesions and furthermore lesions of very small size are not considered clinically significant.
  • classifiers are linear classifiers, such as the Fisher's linear discriminant or non-linear classifiers, such as artificial neural networks or support vector machines.
  • said classifiers can be applied in cascade to the previously mentioned threshold-based classifier.
  • the method according to the invention is performed by a system of the type depicted in figure 5, which comprises a workstation 500, of known type, having an elaborating subsystem 510, a display 520, a keyboard 530, a mouse 540 and a device for connection to a local network (network bus) 550.
  • the elaborating system can be of a distributed type (not shown) having an elaborating subsystem and input/output peripheral drives, local or remote.
  • the workstation 500 or the distributed system are arranged to elaborate groups or modules of programs stored on a disk 560 or accessible through a network, to display the method described, and to display the results obtained. Said solution here mentioned are considered well known in the art and will not be further described because they are not relevant for the understanding and carrying out of the present invention.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • Quality & Reliability (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)
  • Apparatus For Radiation Diagnosis (AREA)

Abstract

On décrit un procédé d'identification de lésions potentielles d'un sein à partir de jeux de données d'images tomographiques d'une région thoracique d'une patiente, les jeux de données comportant une pluralité de voxels (2) présentant chacun une valeur d'intensité, les images comprenant une région (10) d'intérêt qui comprend au moins un sein (6). Le procédé comporte les étapes consistant à : acquérir un ensemble d'images après administration d'un agent de contraste à la patiente ; normaliser (254) l'intensité des voxels (2) appartenant à la région (10) d'intérêt des images acquises en fonction d'au moins un facteur de normalisation ; classifier (255) chacun des voxels normalisés (2) sur la base d'un critère de classification, de manière à identifier des régions (40) représentant des lésions potentielles. Le procédé est caractérisé en ce que le facteur de normalisation est basé sur des voxels (2) de normalisation correspondant à une structure anatomique (34), lesdits voxels (2) de normalisation présentant des valeurs d'intensité renforcées du fait de l'administration de l'agent de contraste.
PCT/IT2009/000005 2009-01-09 2009-01-09 Procédé et système de reconnaissance automatique de lésions dans un ensemble d'images de résonance magnétique d'un sein WO2010079519A1 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP09787626A EP2386102A1 (fr) 2009-01-09 2009-01-09 Procédé et système de reconnaissance automatique de lésions dans un ensemble d'images de résonance magnétique d'un sein
PCT/IT2009/000005 WO2010079519A1 (fr) 2009-01-09 2009-01-09 Procédé et système de reconnaissance automatique de lésions dans un ensemble d'images de résonance magnétique d'un sein

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/IT2009/000005 WO2010079519A1 (fr) 2009-01-09 2009-01-09 Procédé et système de reconnaissance automatique de lésions dans un ensemble d'images de résonance magnétique d'un sein

Publications (2)

Publication Number Publication Date
WO2010079519A1 true WO2010079519A1 (fr) 2010-07-15
WO2010079519A9 WO2010079519A9 (fr) 2010-10-14

Family

ID=41402147

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IT2009/000005 WO2010079519A1 (fr) 2009-01-09 2009-01-09 Procédé et système de reconnaissance automatique de lésions dans un ensemble d'images de résonance magnétique d'un sein

Country Status (2)

Country Link
EP (1) EP2386102A1 (fr)
WO (1) WO2010079519A1 (fr)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013086026A1 (fr) * 2011-12-05 2013-06-13 The Johns Hopkins University Système et procédé de détection automatique d'anomalies tissulaires
CN104809717A (zh) * 2014-01-28 2015-07-29 上海西门子医疗器械有限公司 非刚体医学图像兴趣点标识、定位方法和装置、医疗设备
WO2018005939A1 (fr) * 2016-07-01 2018-01-04 The Board Of Regents Of The University Of Texas System Procédés, appareils et systèmes destinés à créer des représentations tridimensionnelles présentant des caractéristiques géométriques et superficielles de lésions cérébrales
WO2019145896A1 (fr) * 2018-01-25 2019-08-01 Per Hall Compositions et méthodes de surveillance du traitement de troubles du sein
EP3940629A1 (fr) 2020-07-13 2022-01-19 Koninklijke Philips N.V. Correction d'intensité d'image dans l'imagerie par résonance magnétique

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109410181B (zh) * 2018-09-30 2020-08-28 神州数码医疗科技股份有限公司 一种心脏图像分割方法及装置

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2007062135A2 (fr) * 2005-11-23 2007-05-31 Junji Shiraishi Procede assiste par ordinateur permettant la detection des modifications d'intervalles dans des scintigrammes successifs des os du corps entier ainsi que progiciels et systemes associes
WO2007059615A1 (fr) * 2005-11-23 2007-05-31 The Medipattern Corporation Procede et systeme d'analyse quantitative et qualitative assistee par ordinateur d'images medicales

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2007062135A2 (fr) * 2005-11-23 2007-05-31 Junji Shiraishi Procede assiste par ordinateur permettant la detection des modifications d'intervalles dans des scintigrammes successifs des os du corps entier ainsi que progiciels et systemes associes
WO2007059615A1 (fr) * 2005-11-23 2007-05-31 The Medipattern Corporation Procede et systeme d'analyse quantitative et qualitative assistee par ordinateur d'images medicales

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013086026A1 (fr) * 2011-12-05 2013-06-13 The Johns Hopkins University Système et procédé de détection automatique d'anomalies tissulaires
US9607392B2 (en) 2011-12-05 2017-03-28 The Johns Hopkins University System and method of automatically detecting tissue abnormalities
CN104809717A (zh) * 2014-01-28 2015-07-29 上海西门子医疗器械有限公司 非刚体医学图像兴趣点标识、定位方法和装置、医疗设备
CN104809717B (zh) * 2014-01-28 2018-03-23 上海西门子医疗器械有限公司 非刚体医学图像兴趣点标识、定位方法和装置、医疗设备
WO2018005939A1 (fr) * 2016-07-01 2018-01-04 The Board Of Regents Of The University Of Texas System Procédés, appareils et systèmes destinés à créer des représentations tridimensionnelles présentant des caractéristiques géométriques et superficielles de lésions cérébrales
US11727574B2 (en) 2016-07-01 2023-08-15 The Board Of Regents Of The University Of Texas System Methods, apparatuses, and systems for creating 3-dimensional representations exhibiting geometric and surface characteristics of brain lesions
WO2019145896A1 (fr) * 2018-01-25 2019-08-01 Per Hall Compositions et méthodes de surveillance du traitement de troubles du sein
EP3940629A1 (fr) 2020-07-13 2022-01-19 Koninklijke Philips N.V. Correction d'intensité d'image dans l'imagerie par résonance magnétique
WO2022013023A1 (fr) 2020-07-13 2022-01-20 Koninklijke Philips N.V. Correction d'intensité d'image dans une imagerie par résonance magnétique

Also Published As

Publication number Publication date
WO2010079519A9 (fr) 2010-10-14
EP2386102A1 (fr) 2011-11-16

Similar Documents

Publication Publication Date Title
US11344273B2 (en) Methods and systems for extracting blood vessel
US11094067B2 (en) Method and system for image processing
El-Baz et al. Computer‐aided diagnosis systems for lung cancer: challenges and methodologies
EP1851722B8 (fr) Dispositif et procede de traitement d'image
Peters et al. Optimizing boundary detection via simulated search with applications to multi-modal heart segmentation
Montagnat et al. Anisotropic filtering for model-based segmentation of 4D cylindrical echocardiographic images
WO2007044508A2 (fr) Systeme et procede de detection de reperes du corps entier et de segmentation et quantification des variations dans des images numeriques
Marias et al. A registration framework for the comparison of mammogram sequences
Casiraghi et al. Automatic abdominal organ segmentation from CT images
KR101258814B1 (ko) 동적조영증강 유방 ｍｒ 영상에서 조직별 밝기값 보정 및 종양 강성 제약을 적용한 비강체 정합 방법 및 시스템
US9990719B2 (en) Method and system for generating multiparametric nosological images
WO2010079519A1 (fr) Procédé et système de reconnaissance automatique de lésions dans un ensemble d'images de résonance magnétique d'un sein
Kaftan et al. Fuzzy pulmonary vessel segmentation in contrast enhanced CT data
Akram et al. An automated system for liver ct enhancement and segmentation
US8244008B2 (en) Methods involving optimizing and mapping images
Campadelli et al. Automatic liver segmentation from abdominal CT scans
Dabass et al. Effectiveness of region growing based segmentation technique for various medical images-a study
Vignati et al. A fully automatic lesion detection method for DCE-MRI fat-suppressed breast images
Janudhivya et al. A new approach for lung cancer cell detection using Mumford-shah algorithm
Carreira et al. Automatic segmentation of lung fields on chest radiographic images
Jabbar A study of preprocessing and segmentation techniques on cardiac medical images
Hoogi et al. Imaging informatics: An overview
Zhu et al. Liver contour extraction using snake and initial boundary auto-generation
Prevost et al. Registration of free-breathing 3D+ t abdominal perfusion CT images via co-segmentation
Mokhomo Automatic detection and segmentation of brain lesions from 3D MR and CT images

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 09787626

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

WWE Wipo information: entry into national phase

Ref document number: 2009787626

Country of ref document: EP