Systems and methods for processing brain scan information to automatically identify abnormalities

Abstract

A method for processing medical information includes receiving an image slice of the brain including a segmented region, forming a first histogram of intensity values ​​for the image slice, and forming a second histogram of intensity values ​​for the image slice. A difference histogram is then generated based on the first and second histograms, and a presence of an abnormality in the image slice is determined based on the difference histogram. The first histogram corresponds to a first segmented region in a first portion of the image slice, and the second histogram corresponds to a second segmented region in a second portion of the image slice that is complementary to the first portion. The first and second segmented regions are, for example, segmented and labeled ASPECTS regions.

Description

[Technical field]

[0001] FIELD OF THE DISCLOSURE

[0001] One or more embodiments described herein relate to processing information to automatically identify abnormalities in brain scans. [Background technology]

[0002]

[0002] Medical professionals are continually challenged to find new ways to provide essential treatment to patients suffering from brain-related conditions. The ability to quickly and accurately detect these conditions can guide the course of treatment with the hope of saving more lives, especially for those who have suffered serious conditions such as ischemic stroke.

[0003]

[0003] When a person is suspected of having certain types of brain abnormalities, the clinical course of action is usually to obtain computerized tomography (CT) images of the brain. The CT images are then evaluated by a radiologist or other practitioner, possibly with the generation of an Alberta Stroke Program Early CT Score (ASPECTS). This score indicates the severity of symptoms the patient has suffered. ASPECTS scores range from 0 to 10, with scores of 6 or higher indicating appropriate treatment for procedures such as mechanical thrombectomy.

[0004]

[0004] As shown, ASPECTS scores are generated based on visual assessment. These assessments are often inaccurate for various reasons, such as misinterpretation of the images by radiologists, tilting or moving of the patient's head during scanning, and uncertainty about the extent of anatomical regions and subtle contrast changes. Manual assessment of brain scan images is subject to significant delays. These and other reasons may prevent patients from receiving qualified treatment and lead to complications that, in other situations, could be prevented if an expert system was used to automatically perform more effective analysis of brain scan assessment in real time, at least from the time of image acquisition. Summary of the Invention [Problem to be solved by the invention]

[0005]

[0005] The embodiments described in this specification provide systems and methods for automatically evaluating brain scans to identify the location of abnormalities in patients suspected of having abnormalities, including, but not limited to, ischemic lesions caused by stroke. [Means for solving the problem]

[0006]

[0006] According to one or more embodiments, a method for processing medical information includes receiving an image slice of a brain including a segmented region, forming a first histogram of intensity values ​​for the image slice, forming a second histogram of intensity values ​​for the image slice, and determining an abnormality for the image slice based on the first histogram and the second histogram, where the first histogram corresponds to a first segmented region in a first portion of the image slice and the second histogram corresponds to a second segmented region in a second portion of the image slice that is complementary to the first portion. The first segmented region and the second segmented region are complementary ASPECTS regions.

[0007] The method includes generating at least one difference histogram based on the first histogram and the second histogram, and determining an anomaly for the image slice is based on the at least one difference histogram. Determining the anomaly includes identifying the at least one difference histogram having values ​​within a range and determining the anomaly based on values ​​of the at least one difference histogram within the range.

[0008]

[0008] Determining the anomaly includes identifying at least one difference histogram having one or more values ​​that exceed a predetermined reference value and determining the anomaly based on the one or more values ​​of the at least one difference histogram that exceed the predetermined reference value. Generating the at least one difference histogram includes subtracting intensity values ​​of the first histogram from intensity values ​​of the second histogram.

[0009]

[0009] Generating at least one difference histogram includes generating a first difference histogram based on a difference between the first histogram and a first reference histogram, and generating a second difference histogram based on a difference between the second histogram and the second reference histogram. The first reference histogram is indicative of brain tissue free of pathology in the first segmented region, and the second reference histogram is indicative of brain tissue free of pathology in the second segmented region. Determining an anomaly includes generating a feature vector based on the at least one difference histogram, inputting the feature vector into a classifier model, and predicting the anomaly based on an output of the classifier model.

[0010] The method further includes determining whether the first histogram has a concentration of intensity values ​​within a predetermined range and identifying the image slice as having an old anomaly when the first histogram has a concentration of intensity values ​​within the predetermined range. The method includes generating a difference histogram based on exclusion of intensity values ​​within the predetermined range of the first histogram corresponding to the old anomaly.

[0011]

[0011] In accordance with one or more embodiments, a system for processing medical information includes a histogram generator configured to generate a first histogram of intensity values ​​for an image slice and a second histogram of intensity values ​​for the image slice, the image slice including a segmented region of the brain, and a decision engine configured to determine an abnormality in the image slice based on the first histogram and the second histogram, wherein the first histogram corresponds to a first segmented region in a first portion of the image slice and the second histogram corresponds to a second segmented region in a second portion of the image slice that is complementary to the first portion.

[0012]

[0012] The system has difference logic configured to generate at least one difference histogram based on the first histogram and the second histogram, and the decision engine is configured to determine anomalies in the image slice based on the at least one difference histogram.

[0013]

[0013] The decision engine is configured to determine the anomaly by identifying at least one difference histogram having a value within a range and determining the anomaly based on the value of the at least one difference histogram within the range. The decision engine is configured to determine the anomaly by identifying at least one difference histogram having one or more values ​​that exceed a predetermined reference value and determining the anomaly based on the value of the at least one difference histogram that exceeds the predetermined reference value.

[0014]

[0014] The difference logic is configured to generate at least one difference histogram by subtracting intensity values ​​of the first histogram from intensity values ​​of the second histogram. The difference logic is configured to generate a first difference histogram based on a difference between the first histogram and a first reference histogram, and to generate a second difference histogram based on a difference between the second histogram and a second reference histogram. The first reference histogram is indicative of brain tissue free of lesions in the first segmented region, and the second reference histogram is indicative of brain tissue free of lesions in the second segmented region.

[0015]

[0015] The decision engine is configured to determine anomalies by generating a feature vector based on the at least one difference histogram, inputting the feature vector to a classifier model, and predicting anomalies based on an output of the classifier model. The difference logic is configured to generate the at least one difference histogram by subtracting intensity values ​​of the first histogram from intensity values ​​of the second histogram.

[0016] The system includes identification logic configured to determine whether the first histogram has a concentration of intensity values ​​within a predetermined range and to identify the image slice as having an old anomaly when the first histogram has a concentration of intensity values ​​within the predetermined range. The difference logic is configured to generate at least one difference histogram based on exclusion of intensity values ​​within the predetermined range of the first histogram corresponding to the old anomaly.

[0017]

[0017] Further objects and features of the present invention will become more readily apparent from the following detailed description and the appended claims when taken in conjunction with the drawings in which: Several illustrative embodiments have been illustrated and described, with like reference numerals identifying like parts in each of the drawings. [Brief description of the drawings]

[0018] [Figure 1]

[0018] FIG. 1 illustrates an embodiment of a medical image analyzer. [Figure 2A]

[0019] FIG. 1 illustrates an embodiment of a method for analyzing medical images. [Figure 2B] FIG. 1 illustrates an embodiment of a method for analyzing medical images. [Figure 3A]

[0020] FIG. 2 shows examples of segmented and labeled image slices. [Figure 3B] FIG. 2 shows examples of segmented and labeled image slices. [Figure 4]

[0021] FIG. 1 shows an example of a lesion in one region of an image slice. [Diagram 5]

[0022] FIG. 2 illustrates an embodiment for generating one type of difference histogram. [Figure 6A]

[0023] FIG. 13 illustrates an embodiment for generating another type of difference histogram. [Figure 6B] FIG. 13 shows another depiction of a difference histogram. [Figure 7]

[0024] FIG. 13 illustrates an example of a difference histogram generated for multiple regions. [Figure 8]

[0025] FIG. 1 illustrates an embodiment for identifying old and new lesions. [Figure 9]

[0026] FIG. 13 shows examples of histogram curves for old and new lesions. [Figure 10]

[0027] FIG. 1 illustrates an embodiment of a classifier model for analyzing brain scan images. [Figure 11A]

[0028] FIG. 1 illustrates an embodiment of a method for filtering brain scans. [Figure 11B] FIG. 1 illustrates an embodiment of a method for filtering brain scans. [Figure 12]

[0029] FIG. 13 illustrates an example of the variability of image parameters used during training of the classifier model. [Figure 13]

[0030] FIG. 13 illustrates an example of the variability of image parameters used during training of the classifier model. [Figure 14]

[0031] FIG. 13 illustrates an example of the variability of image parameters used during training of the classifier model. [Figure 15]

[0032] FIG. 1 illustrates an embodiment of a medical image analyzer. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019]

[0033] It should be understood that the figures are only schematic and are not drawn to scale, and that the same reference numerals are used throughout the figures to indicate the same or similar parts.

[0020]

[0034] The description and drawings illustrate the principles of the various exemplary embodiments. Thus, it will be appreciated that those skilled in the art can create various configurations that embody the principles of the present invention and fall within its scope, even if not expressly described or shown herein. Furthermore, all examples described herein are expressly intended to be primarily for pedagogical purposes to aid the reader in understanding the principles of the invention and concepts contributed by the inventor(s) to the technology they promote, and should be construed as such specifically described examples and conditions, but not as limiting. Additionally, the term "or" as used herein refers to a non-exclusive or (i.e., and / or) unless otherwise indicated (e.g., "otherwise" or "or alternatively"). Furthermore, the various exemplary embodiments described herein are not necessarily mutually exclusive, as some exemplary embodiments may be combined with one or more other exemplary embodiments to form new exemplary embodiments. Descriptive terms such as "first," "second," "third," and the like are not intended to limit the order of the elements discussed, but are used to distinguish one element from the next, and are generally interchangeable. Values ​​such as maximum or minimum are predefined and are set to different values ​​based on the application.

[0021]

[0035] 1 shows an embodiment of a medical image analyzer 1 for automatically identifying abnormalities in a patient's brain scan. The system works with image slices received from an imaging system, such as a CT scanner 2, to identify the abnormalities. For example, for each patient, the system receives at least one image slice that targets a specific location of the brain where an abnormality is suspected. When an ASPECTS region is subjected to analysis by the system, two image slices are received that together represent all ten ASPECTS regions on a per-hemisphere basis.

[0022]

[0036] 1, the medical image analyzer includes a histogram generator 10, difference logic 20, and a decision engine 30. The histogram generator 10 receives image slice(s) from a CT scanner 2 output from segmentation and labeling logic 3. The segmentation and labeling logic analyzes the image slice(s) to identify the location of ASPECTS regions and then overlays contours and labels corresponding to those regions for input to the histogram generator. A more comprehensive discussion of how segmentation and labeling of ASPECTS regions is performed in accordance with one or more embodiments is provided below.

[0023]

[0037] For each received image slice, the histogram generator 10 generates a first histogram of intensity values ​​for a first segmented region of the image slice and a second histogram of intensity values ​​for a second segmented region. The first and second segmented regions are, for example, complementary ASPECTS regions for different lateral parts of the brain (e.g., left and right hemispheres). In one embodiment, the intensity values ​​correspond to a full range of Hounsfield Units (HU), or in some embodiments, to a limited range of HU units, as will be described. The histogram generator 10 generates the first and second histograms for only one complementary pair of ASPECTS regions, or in another embodiment, generates the first and second histograms for multiple complementary pairs of ASPECTS regions, including up to all ten regions.

[0024]

[0038] The difference logic 20 is configured to generate a difference histogram based on the first and second histograms for a complementary pair of regions, or for each complementary pair of regions when multiple pairs of regions are regions of interest. This is accomplished, for example, by subtracting the HU values ​​in the first histogram from the HU values ​​from the second histogram. The first and second histograms will appear significantly different from each other when one of the regions in the pair contains an abnormality. This is because tissue affected by the abnormality produces intensity values ​​that are significantly different from healthy brain tissue. The difference will be reflected in the difference histogram.

[0025]

[0039] The decision engine 30 is configured to determine anomalies in one or more regions based on the difference histogram. This can be performed in a variety of ways. In one embodiment, the decision engine 30 compares values ​​in the difference histogram with one or more predefined criteria values, ranges, or patterns, and generates a decision as to whether an anomaly exists in at least one of the corresponding regions based on the comparison. In another embodiment, the decision engine 30 implements a classifier model to predict, at least within a certain probability, the presence and location of an anomaly. An embodiment of the operations performed by the decision engine will be described in more detail with reference to an embodiment of the method. Once a decision is made as to whether an image slice(s) contains an anomaly (e.g., within seconds of receiving the segmented and labeled image slices), medical personnel can more quickly and accurately determine a course of treatment that could previously be performed using manual techniques.

[0026]

[0040] The medical image analyzer 1 is modified in various ways to generate further embodiments. For example, in one embodiment, the difference logic 20 compares the first histogram to a first reference histogram and the second histogram to a second reference histogram. The first reference histogram has intensity values ​​indicative of all normal brain tissue (e.g., healthy tissue or at least tissue that is free of lesions or of a particular type of lesion) in the corresponding brain region. The second reference histogram has intensity values ​​indicative of normal and healthy brain tissue (e.g., healthy tissue or at least tissue that is free of lesions or of a particular type of lesion) in the corresponding complementary region, e.g., the same region on the opposite hemisphere. The difference logic 20 then generates a first difference histogram based on a comparison of the first histogram to a first reference histogram (e.g., subtracting the intensity values ​​in the first histogram from the intensity values ​​in the first reference histogram) and generates a second difference histogram based on a comparison of the second histogram to a second reference histogram (e.g., subtracting the intensity values ​​in the second histogram from the intensity values ​​in the second reference histogram).

[0027]

[0041] The decision engine 30 then generates two decisions: a first decision indicating whether the region corresponding to the first difference histogram has an abnormality, and a second decision separately indicating whether the region corresponding to the second difference histogram has an abnormality. This embodiment of the system can more quickly or much more accurately locate, for example, old stroke lesions or new stroke lesions within each region that is subject to evaluation based on one of the first and second histograms.

[0028]

[0042] In another embodiment, the system expands the region of interest from 2D (ASPECT area) to 3D (volume) including voxels in different (e.g. adjacent) image slices that also belong to the ASPECT area. This embodiment ensures that the histogram is generated by lesions that are more obvious around (below or above) the 2D ASPECT area. The system and method embodiment is then applied to these 3D image slices in a similar manner.

[0029]

[0043] In one embodiment, the medical image analyzer includes discrimination logic 40 for identifying and excluding old lesions via difference histogram data input to the decision engine 30. Identifying and excluding old lesions allows the difference histogram(s) to include only data corresponding to new lesions, thereby allowing radiologists and physicians to target treatment in a more effective manner. These and other operations related to discriminating between new and old lesions performed by the discrimination logic are discussed in more detail below.

[0030]

[0044] 2A and 2B show operations involved in an embodiment of a method for analyzing medical images to automatically identify anomalies in brain scans. The method may be performed by any of the system embodiments described herein or by a different system. For purposes of explanation, the method will be described as being performed by the system of FIG.

[0031]

[0045] 2A, the method includes receiving, at 205, one or more brain scan images of a patient to be evaluated, e.g., a patient suspected of having a brain abnormality, such as an ischemic stroke or other brain condition. The images are generated by a scanning system, which may be, for example, a computerized tomography scanner. In this case, the brain scan images are non-contrast CT (NCCT) images. The CT scanner may be local (e.g., in the same hospital or clinical environment) or located remotely from the processing logic of the system implementing the method. In this latter case, embodiments of the system and method are used, for example, in an outsourcing situation, where images are received from a network.

[0032]

[0046] The brain scan images correspond to a particular area of ​​the brain. Given an axis passing longitudinally through the patient's body (e.g., from head to toe), the one or more images are from one or more axial slices of the brain that include a target location. The target location (or slice) is, for example, a location where one or more types of brain abnormalities are expected to occur given a suspected medical condition of the patient. In one embodiment, two images are received that correspond to axial slices at different parts of the brain, e.g., a superior slice and an inferior slice.

[0033]

[0047] When a patient is suspected of having an ischemic stroke, a slice of the brain scan image includes, for example, a region including the middle cerebral artery (MCA). As will be described in more detail, such a stroke will cause an obstruction (e.g., an ischemic lesion) in the image at the location of the MCA, at least for the hemisphere of the brain in which the stroke occurred. For purposes of illustration, the method will be described based on receiving two image slices of the brain corresponding to an inferior slice and a superior slice, although different image slices are received and evaluated according to the embodiments described herein. The image slices are stored in the system's memory for processing.

[0034]

[0048] At 210, gantry tilt correction is performed on the images. Gantry tilt refers to aspects of a helical scanning CT system equipped with a multi-row detector that operates at a gantry tilt angle. The tilt angle can cause distortions in the image slices that can prevent artifacts and other features of clinical importance from being clearly or accurately seen. Therefore, the image slices are pre-processed to reduce or filter out any distortions caused by the gantry tilt angle. This can be accomplished, for example, by reformatting each of the image slices to create a rectangular volume in the absence of gantry tilt.

[0035]

[0049] At 215, the image slices are subjected to intensity normalization, which, for example, improves the process of assigning intensity values ​​to pixels in the image slices in subsequent operations.

[0036]

[0050] In one embodiment, intensity normalization of the brain scan images is performed by adding a predefined offset value (e.g., an offset value of 4,000) to the HU values ​​in the image slice being evaluated. This type of normalization centers the HU values ​​over a smaller range, which helps accentuate differences between regions within the slice that allow for more accurate identification of abnormalities.

[0037]

[0051] At 220, a segmentation operation is performed to identify different regions of interest in the image slices. This operation includes segmenting each of the image slices into predefined regions, such as regions defined by the ASPECTS protocol. When segmentation is performed according to the ASPECTS protocol, a total of 20 regions of interest are defined to be included in the two image slices, i.e., 10 regions on the left hemisphere of the brain and 10 regions on the right hemisphere of the brain. The 10 regions in the left hemisphere are of the same type as the regions on the right hemisphere, and therefore the 10 regions on the left hemisphere are considered to be complementary to the respective regions on the right hemisphere, thereby forming complementary region pairs.

[0038]

[0052] Table 1 identifies 10 regions in each hemisphere of the brain that correspond to the ASPECTS protocol. The ASPECTS protocol and its corresponding regions are discussed at https: / / linkinghub.elsevier.com / retrieve / pii / S0140673600022376, the contents of which are incorporated by reference herein for all purposes. Seven of the 10 regions appear in the superior image slice (FIG. 3A) shown on a hemisphere-by-hemisphere basis, and the remaining three regions appear in the inferior image slice (FIG. 3B). Together, all 10 regions are considered candidate regions for lesions or other abnormalities affecting the brain.

[0039] [Table 1]

[0040]

[0053] In one embodiment, the segmentation of the image slices is performed automatically in 3D using a model-based approach. The model processes the images to identify (extract) and then generates overlay graphics that highlight the contours of each of the 10 regions of interest. The model then labels each of the segmented regions (see, e.g., operation 225) as shown in Figures 3A and 3B. The contours are matched to each other, for example, using multi-planar reformatting.

[0041]

[0054] To train a model for ASPECTS region segmentation, multiple datasets from different individual brains are used to create independent ground truth segmentations. The datasets include a combination of scans containing abnormalities (e.g., lesions from stroke victims) and scans of control patients without lesions. In the plane, pixel spacing varies between images in the dataset. For example, in a realistic application, pixel spacing in the plane varies within a range of 0.38mm x 0.38mm and 0.58mm x 0.58mm, and kVp ranges between 100 and 140. The scans have a slice thickness of a predefined value (or a predefined range of values). An example slice thickness for the dataset is 3mm.

[0042]

[0055] Ground truth annotations for training the model-based region segmentation are obtained by performing two operations, each of which harvests previously available information on NCCT region annotations. Both operations use a dedicated annotation application that serves, for example, the following workflow: First, the application performs a fully automatic fitting of the initially trained model to the scan. Then, it performs an automatic multiplanar reformatting on the lower ASPECTS slices. The multiplanar reformatting operation is based on locating the centroids and principal components of vertices in a mesh that corresponds to the cerebral cortical areas M1-M3. The mesh corresponds, for example, to a pre-labeled version of the brain scan in Fig. 3C.

[0043]

[0056] Next, interactive refinement and confirmation of the inferior ASPECTS slice is performed, followed by removing (e.g., making invisible) the region boundaries in the different slices other than the inferior slice. Then, the ASPECTS region boundaries in the inferior ASPECTS slice are corrected (e.g., interactively), and then a multiplanar reformatting is automatically performed for the superior ASPECTS slice. This operation is followed by interactive refinement and confirmation of the superior ASPECTS slice, followed by removing or otherwise making invisible the region boundaries in the different slices other than the inferior slice. Finally, the correction of the ASPECTS region boundaries in the superior ASPECTS slice is performed in the same manner (e.g., interactively).

[0044]

[0057] The automatic segmentation of image slices is performed, for example, according to the techniques described in WO2020 / 109006, the contents of which are incorporated herein by reference.

[0045]

[0058] At 225, a feature extraction operation is performed by the model to automatically label each of the contour regions in the upper and lower image slices, e.g., as shown in Table 1. The feature extraction operation is performed as follows: First, a fitting of region boundaries in the image slices is performed, e.g., according to a predefined format. The positions of the fitted mesh vertices of the cortical areas M1-M3 (lower slices) or M4-M6 (upper slices) are extracted. The cortical areas are topologically defined, e.g., as 1 cm thick stripes. The center of the observation plane (e.g., the focal point) is calculated as the centroid of the set of extracted vertices.

[0046]

[0059] Next, the extracted vertices are stacked into a matrix A, and the three principal axes are calculated, for example, using the singular value decomposition (SVD) technique, where A=USV′, where U and V are orthogonal matrices, and S is a diagonal matrix. The normal to the viewing plane is extracted from the principal axes as the right-most vector of V. Then, the left-right (LR) vector is extracted as the average between the corresponding left and right vertices, since the areas on both sides are topologically symmetric by design. Finally, the AP vector is then calculated based on the cross product of the right-most vector of V and the LR vector using 3D algebra.

[0047]

[0060] 3A and 3B show examples of upper and lower image slices, respectively, that have been automatically annotated to include overlaid contours and labels for each of the regions. FIG. 3A shows the labels for the seven ASPECTS regions in the upper image slice, and FIG. 3B shows the labels for the remaining three ASPECTS regions in the lower image slice. Not only the lower and upper image slices (with the graphically added contours and labels) but also the corrected ASPECTS regions are stored as a mesh. (In one embodiment, a mesh can be considered as a graphically contoured and labeled image slice, e.g., as shown in FIGS. 3A and 3B.) As the number of training data sets increases, the model for performing automatic image segmentation becomes more accurate.

[0048]

[0061] At 230, intensity (HU) values ​​are generated for pixels in each of the labeled regions in the image slices. The intensity values ​​are, for example, values ​​in Hounsfield Units (HU), for example, allocated within at least a predefined HU range, for example, an unsigned short value range. The HU values ​​provide an indication of tissue density, represented in a color or gray scale, for example, in various shades between white and black, inclusive. Less dense tissues have a darker shade (or intensity), while more dense tissues are represented in a lighter shade (or intensity). Thus, the HU values ​​effectively correspond to grayscale intensity values ​​in a CT image that provide an indication of tissue density. In one embodiment, the HU values ​​for each region in each image slice are organized and stored in a table. For example, a first table includes HU values ​​for regions in an upper image slice on a hemisphere-by-hemisphere basis, and a second table includes HU values ​​for regions in a lower image slice on a hemisphere-by-hemisphere basis.

[0049]

[0062] At 235, intensity histograms are generated based on the HU values ​​generated for the upper and lower image slices. The intensity histograms are generated on a bilateral basis, e.g., for each image slice, one set of histograms is generated for regions in the left lateral portion of the brain, and a complementary set of histograms is generated for regions in the right lateral portion of the brain. In this sense, the histograms generated for each image slice are referred to as bilateral histograms, which are generated as follows:

[0050]

[0063] In one embodiment, intensity histograms are generated based on HU values ​​in the left and right lateral portions of the superior image slice. The set of histograms generated for the left lateral portion includes seven histograms, one for each of the seven labeled ASPECTS regions in the superior image slice. The histogram generated for a first of the seven regions in the left lateral portion of the brain provides an indication of the HU values ​​of the pixels in the first region. For example, a first number of pixels in the first region have a first HU value, a second number of pixels in the first region have a second HU value, etc. These numbers form a distribution of HU values ​​in the histogram that is used as a basis for determining the nature of the tissue in that region and, if any, lesions or other abnormalities. Histograms for the remaining six regions in the left lateral portion of the superior image slice are then generated in a similar manner.

[0051]

[0064] Once all of the histograms for the regions in the left lateral portion have been generated, histograms are generated in a similar manner for each of the seven regions in the right lateral portion of the upper image slice. Thus, operation 235 produces a total of 14 histograms, seven for each region in the left lateral portion and seven complementary histograms for each region in the right lateral portion of the upper image slice.

[0052]

[0065] The range of HU values ​​is the entire range of HU values ​​or a predetermined subset of the entire range of values ​​that may be, for example, related to a particular type(s) of brain abnormality of interest. The subset of values ​​may, for example, correspond to a predetermined number of bins into which the entire range of HU values ​​is divided. For example, in one implementation, the intensity histogram has values ​​that are limited to a range between HU value 10 and HU value 60 spread over 25 bins with each bin having a size of 2 HU. This range may be suitable for some applications, for example, a lower HU boundary value of 10 excludes most of the CSF, while an upper HU boundary value of 60 includes entirely gray matter but excludes calcifications and hemorrhagic foci. The histogram data is set based on different ranges of HU values ​​in another embodiment.

[0053]

[0066] The set of histograms generated for the right lateral portion includes three histograms, one for each of the three labeled ASPECTS regions in the sub-image slice. The histogram generated for the first of the three regions in the left lateral portion of the brain provides an indication of the HU values ​​of the pixels in that first region. These counts form a distribution of HU values ​​within the histogram that is used as a basis for determining the nature of the tissue in that region and, if any, lesions or other abnormalities. The histograms for the remaining two regions in the left lateral portion of the sub-image slice are generated in a similar manner.

[0054]

[0067] Once all of the histograms for the regions in the left lateral portion have been generated, histograms are generated in a similar manner for each of the three regions in the right lateral portion of the sub-image slice. Thus, operation 340 produces six histograms in total: three histograms for each region in the left lateral portion of the sub-image slice, and three complementary histograms for each region in the right lateral portion. The same ranges used to generate the histograms for the super-image slice are used to generate the histograms in the sub-image slice.

[0055]

[0068] Based on operation 235, a total of 20 histograms are generated for the upper and lower image slices that provide a comprehensive set of histogram data to be used to identify lesions. The 10 histograms for the left brain portion are complementary to the 10 histograms for the right brain portion across two image slices, and the histograms for each complementary pair will have a different distribution of HU values ​​when one of the regions has a lesion and the other does not.

[0056]

[0069] FIG. 4 shows an example of a superior image slice of a patient with an ischemic lesion in the lenticular nucleus (L) region 410 of the left lateral portion of the brain. The L region in the right lateral portion 420 of the brain (complementary to region 410) is free of the lesion. The lesion includes different types of tissue with different densities from the brain tissue, and thus the lesion tissue will exhibit a different scan intensity than normal brain tissue. As a result, the histogram distribution of HU values ​​in region 410 will be substantially different from the histogram distribution of HU values ​​in region 420. This difference serves as the basis for generating derivative histogram data that identifies the lesion.

[0057]

[0070] At 240, derived histogram data is generated based on the histograms generated at operation 235. The derived histogram data includes, for example, a plurality of difference histograms for each of the ten labeled ASPECTS regions in the upper and lower image slices. Referring to 240A, in one implementation, each difference histogram is generated based on the difference between the histogram of HU values ​​of one region in the left lateral portion of the brain and the histogram of HU values ​​of a complementary region in the right lateral portion of the brain, for example, the histogram for the L region in the left lateral portion is subtracted from the histogram for the L region in the right lateral portion. Thus, ten difference histograms are generated for each of the ten labeled regions in the upper and lower image slices.

[0058]

[0071] The difference histogram provides a substantial (quantitative and qualitative) indication of how different the intensity values ​​are at each complementary pair of locations in the brain scan. For example, when the difference histogram has difference values ​​that fall below a predefined threshold or within a first range (or otherwise establishes a first type of pattern), the difference histogram is used to infer that the complementary regions in the associated image slice are free of abnormalities. When the difference histogram has difference values ​​that are above a predefined threshold or within a second range (or otherwise establishes a second type of pattern), the difference histogram is used to infer that at least one of the complementary regions in the associated image slice is a candidate for containing an abnormality, e.g., a lesion.

[0059]

[0072] Referring to 240B, a first histogram for the left lateral region is compared to a first reference histogram, and a second histogram for the complementary right lateral region is compared to a second reference histogram for that region. The first reference histogram has intensity values ​​indicative of all normal brain tissue (e.g., healthy tissue, or at least tissue that is free of lesions or a particular type of lesion) in the corresponding brain region. The second reference histogram has intensity values ​​indicative of normal and healthy brain tissue (e.g., healthy tissue, or at least tissue that is free of lesions or a particular type of lesion) in the corresponding complementary region, e.g., the same region on the opposing hemisphere.

[0060]

[0073] Thus, at 240B, a first difference histogram is generated based on a comparison of the first histogram to a first reference histogram (e.g., by subtracting the intensity values ​​in the first histogram from the intensity values ​​in the first reference histogram), and a second difference histogram is generated based on a comparison of the second histogram to a second reference histogram (e.g., by subtracting the intensity values ​​in the second histogram from the intensity values ​​in the second reference histogram).

[0061]

[0074] In another embodiment, the method expands the region of interest from 2D (ASPECT area) to 3D (volume) that includes voxels in different (e.g., adjacent) image slices that also belong to the ASPECT area. This embodiment ensures that the histogram is generated with more obvious lesions at the periphery (lower or higher) of the 2D ASPECT area. The system and method embodiment is then applied to these 3D image slices in a similar manner.

[0062]

[0075] FIG. 5 shows an example of a difference histogram (generalized to a curve) 530 generated based on a histogram (generalized to a curve) 510 corresponding to a region in the left lateral portion of an image slice and a histogram (generalized to a curve) 520 corresponding to the same (or complementary) region in the right lateral portion of the image slice. In the example shown in FIG. 5, both complementary regions are free of pathology. Since there is no pathology (e.g., healthy brain tissue) in this region pair, both histograms 510 and 520 have similar values ​​and peaks in substantially the same range(s) of HU values. As a result, the difference histogram 530 has values ​​that are below a threshold or are within a first range. For illustrative purposes, the histogram 510 is labeled R1 and the histogram 520 is labeled R2. Thus, the difference histogram 530 is calculated by the equation: R1-R2, which explains the negative difference values ​​in the histogram 530.

[0063]

[0076] Referring again to FIG. 4, an example of a superior image slice containing an ischemic lesion in the left lateral portion of the brain is shown. In this example, the lesion (indicated by arrow 410) is located in the lentiform (L) region. The lentiform region in the complementary lateral portion of the brain (indicated by arrow 420) is normal in that it does not contain the lesion. As is evident from a comparison of regions in this image slice, the L region in the left lateral portion of the image has a different (e.g., darker) intensity than the L region in the right lateral portion of the image. These differences in intensity create different HU values ​​for these L regions, which will be reflected in the histograms generated for these different L areas.

[0064]

[0077] FIG. 6A shows an example of a difference histogram 570 (generalized to a curve) generated based on a lesion shown in the image slice of FIG. 4. As shown in FIG. 6A, the difference histogram 550 is generated based on a histogram 550 (generalized to a curve) corresponding to an L region in the left lateral portion of the superior image slice and a histogram 560 (generalized to a curve) corresponding to a complementary L region in the right lateral portion of the same image slice. As shown, there is a lesion in the L region corresponding to the histogram 550, but not in the L region 560. As a result, the difference histogram 570 will have values ​​that exceed a threshold or otherwise fall within a second range, for example, that is greater than the first range. An example difference histogram that is not generalized by a curve is shown in FIG. 6B. In one implementation, the shape of the curves and / or the degree of difference in values ​​in the difference histogram form a pattern that indicates whether a lesion is present in one or more of the regions of the image slice.

[0065]

[0078] Furthermore, the method includes an optional extraction operation that involves determining which regions have a certain probability of having a lesion and which do not, the determination being made based on the difference histogram, and those regions deemed to have a higher probability of having a lesion are considered candidate regions for further evaluation.

[0066]

[0079] 7 shows an example of difference histograms 701-710 generated for all 10 ASPECTS regions. In this example, difference histograms 703-706 have values ​​(or patterns) that potentially qualify as candidate regions containing lesions. The remaining difference histograms correspond to a depiction of the absence of lesions in the corresponding ones of the regions.

[0067]

[0080] Thus, the generation of difference histograms for each of the regions in the image slice effectively serves as a classifier used to distinguish between candidate regions containing lesions requiring further evaluation. In one embodiment, discussed in more detail below, a feature vector is generated based on the difference histograms for input to a machine learning model (e.g., a neural network) that acts as a classifier to confirm candidate lesions determined based on the data in the difference histograms. In one embodiment, prior to generation of the difference histograms, each histogram is normalized to the number of voxels to account for differences in the size of the regions between which the histograms are compared.

[0068]

[0081] FIG. 8 illustrates another embodiment of a method for automatically identifying anomalies in a patient's brain scan. This method is similar to the method of FIG. 2, but is supplemented with a filtering operation that includes distinguishing new foci from old foci in one or more of the superior or inferior image slices. This is accomplished, for example, by identifying the presence of two or more foci in a given ASPECTS region or in different ASPECTS regions, identifying at least one new and at least one old focus, discarding data (e.g., HU values) in the resulting histogram(s) that correspond to the old focus and focusing only on the new focus for the purpose of generating a difference histogram and applying a model to determine whether a new focus is present.

[0069]

[0082] Ischemic stroke lesions are classified into three categories: 1) acute, corresponding to strokes that occur in the patient between 0 and 24 hours; 2) subacute, corresponding to strokes older than 24 hours; and 3) old, corresponding to lesions that developed from a previous stroke, which may be months or years old, for example. In practice, the type of lesion (or time span) cannot be determined solely from visual inspection of the non-contrast CT, since further clinical information is required. Since time is of the essence for stroke victims, this further clinical information is not readily available, which may lead to delays in treatment.

[0070]

[0083] For the purpose of providing optional treatment, embodiments of the present method automatically assess and determine in a matter of seconds whether a lesion is more likely to be a new (acute or subacute) lesion than an old lesion, which is particularly beneficial when multiple lesions appear within the same or different ASPECTS regions in an image slice.

[0071]

[0084] Thus, in one embodiment, the method of Fig. 8 includes some additional operations of the method of Fig. 2. These operations are performed, for example, during or after operation 235. As previously discussed, during operation 235, an intensity histogram is generated based on HU values ​​in the left and right lateral portions of the superior image slice, and an intensity histogram is generated based on HU values ​​in the left and right lateral portions of the inferior image slice.

[0072]

[0085] Old lesions will produce HU values ​​with concentrations within a certain range, while new lesions will produce HU values ​​with concentrations within a different range. These differences in intensity values ​​may be due to, for example, calcification and other effects due to aging.

[0073]

[0086] FIG. 9 shows an example of the difference between histogram curve A generated by an old lesion and histogram curve B generated by a new lesion. As shown, histogram curve A has a lower concentration of HU values ​​than histogram curve B. The lower concentration of HU values ​​for curve A is exhibited by old lesions and is therefore used as the basis for identifying curve A as potentially corresponding to old lesions. In contrast, histogram curve B is shifted to the right relative to curve A and therefore has a higher concentration of HU values, which is exhibited by, for example, new lesions. These higher HU values ​​are therefore used as the basis for identifying curve B as potentially corresponding to new lesions.

[0074]

[0087] Thus, referring to Figure 8, the method includes evaluating the range of HU values ​​of the histograms generated for each of the regions (810), identifying potential lesions as old lesions based on determining a concentration of low HU values ​​(820), and then deleting, otherwise invalidating, or excluding (830) histogram values ​​corresponding to the potential old lesions. With these values ​​deleted or invalidated, the method of Figure 2 continues by generating a difference histogram for the corresponding region(s). One or more of the operations of Figure 8 are performed by the identification logic 40 of Figure 1.

[0075] Selection Model

[0088] A sorting model is used to sort the difference histogram generated by the above operation. The sorting model can be implemented in various forms. In one embodiment, the sorting model is a binary sorter that outputs a decision indicating (at least with a certain probability) whether a candidate region has a lesion or not. A model based on the use of a convolutional neural network (CNN) is discussed as one non-limiting example of evaluating the difference histogram to identify and sort whether a candidate region is likely to have a lesion or not.

[0076]

[0089] FIG. 10 illustrates an embodiment of a CNN model used to generate a binary lesion decision. The CNN model includes a first layer 1010, which is a convolutional layer that receives an input vector generated, for example, based on a difference histogram created for each of the 10 ASPECTS regions in an image slice. The first convolutional layer has a first number of kernels and generates a multidimensional vector of a first size. This vector is input to a second layer 1020, which is another convolutional layer with a second number of kernels and generates another multidimensional vector of a second size different from the first size. The vector output from the second convolutional layer is input to a third layer 1030, which is another convolutional layer with a third number of kernels and generates another vector of a third size. The vector output from the third convolutional layer is input to a third layer 1040, which is a fully connected layer with a predetermined number of input and output nodes, each of which represents a class probability (lesion / non-lesion) corresponding to a decision for the region corresponding to the initial input feature vector. As described below, in one embodiment, the output vectors of a convolutional layer are passed through an activation function before being input to a subsequent layer.

[0077]

[0090] 11A and 11B illustrate operations included in one embodiment of a method of using a CNN model to generate a binary lesion determination. The method uses the CNN model of FIG. 10 or a different CNN model of another embodiment. For purposes of explanation, the method is described as using the CNN model of FIG. 10 having a particular configuration. This configuration and / or its example values ​​are modified in other embodiments depending, for example, on the dataset used to train the model. Furthermore, in other embodiments, other numbers of convolutional layers are used and / or the layers are configured differently, for example, to meet the requirements of the intended application.

[0078]

[0091] 11A, the method includes generating 1110 a feature vector 1005 for each difference histogram generated as a candidate having a lesion. The feature vector is a one-dimensional vector that includes a predefined number of entries (e.g., 25 entries) corresponding to the number of intensity bins. For example, as previously described, when the difference histogram is based on a range of 50 HU values, 25 intensity bins (or entries) are generated, with each bin (or stride) spanning two HU values. In another embodiment, the feature vector is generated based on a difference histogram produced from another operation (different from operation 450) and / or a different number of entries or kernels are used.

[0079]

[0092] At 1120, the feature vector 1005 is input to a first convolutional layer 1010 configured to have a first number (e.g., 4) of convolution kernels, each having a kernel size of a first size (e.g., 5). The convolutional kernels effectively function as sliding windows or filters, each of which includes, for example, a matrix of values ​​(in this example, 5 values ​​per window) that are multiplied by values ​​in the feature vector. For example, the first convolutional layer 1010 uses the first window to perform a convolution operation on the values ​​of the feature vector 1005. The windows corresponding to the three remaining kernels (having at least one value different from the first kernel) are used to perform additional convolution operations on the values ​​of the feature vector 1005. When the stride is set to 2, the output of the first convolutional layer is a four-dimensional vector of size 4×13 (zero padded of size 2).

[0080]

[0093] In one embodiment, the internal values ​​(e.g., weights) of the convolutional layers are determined via AI training. The actual setup (e.g., network topology) given the number of kernels and their sizes is based on internal experimental performance. In some embodiments, a variant of a neural network is used to generate the model output.

[0081]

[0094] The kernel is used to effectively test the feature vector by multiplying the feature vector value with the corresponding value in the kernel, providing an indication of the information embedded in the underlying difference histogram. The output of the convolution operation will be different due to different values ​​in the kernel. Furthermore, different feature vectors (based on different difference histograms) will generate different outputs from the first convolution layer for the same kernel. In some cases, the output of the first convolution layer provides an indication as to whether a lesion is present in the region corresponding to the feature vector. However, to make a more accurate determination, one or more further (e.g., at least a second) convolution layer is included.

[0082]

[0095] At 1130, the 4-dimensional vector output of the first convolutional layer is passed through a Leaky Rectified Linear Unit (LeakyReLU) activation function to reduce the size or otherwise process the vector so that it has a predetermined shape. The LeakyReLU activation function is implemented with a predetermined slope, e.g., 0.2 for some applications. The LeakyReLU activation function is expressed as f(x)=1(x<0)(αx)+1(x>=0)(x), where α is a predetermined constant. In one embodiment, operation 1130 is considered optional.

[0083]

[0096] At 1140, the feature vectors output from the first convolutional layer (optionally passed through a LeakyReLU activation function) are input to a second convolutional layer. This convolutional layer has a second number of kernels having a first kernel size. In one embodiment, the second number of kernels is 8 kernels, with a kernel size of 5(×4). The same stride, 2, used in the first convolutional layer is used in the second convolutional layer.

[0084]

[0097] At 1150, the vectors output from the second convolutional layer are optionally passed through a leaky activation function to reduce the output size of the vectors. For example, the vectors are reduced to a size of 8×7.

[0085]

[0098] 11B, the method begins with the feature vector output from the second convolutional layer (optionally passed through a LeakyReLU activation function) being input to a third convolutional layer at 1160. In one embodiment, the third convolutional layer has one kernel with the same kernel size 5(×8).

[0086]

[0099] At 1170, the vector output from the third convolutional layer is optionally passed through another LeakyReLU activation function to reduce the output side of the vector to 4 (×8).

[0087]

[0100] At 1180, the vector output from the third convolutional layer (optionally passed through a LeakyReLU activation function) is input to the fourth layer, which is a fully connected layer with four input nodes 1050 and two output nodes 1051 and 1052 representing the class probabilities corresponding to the decisions (lesion / non-lesion). The class probabilities are normalized, for example using a softmax function, so that the sum of the probabilities equals one.

[0088]

[0101] In 1190, output nodes 1051 and 1052 output their respective probabilities to non-foci and foci decisions. The output node with the larger probability serves as the decision output from the model. Thus, in this manner, the CNN model acts as a classifier that generates data from the output nodes that is used as a basis for indicating a foci or non-foci decision. For example, when the threshold of the probability that a foci exists (generated from output node 1052) is equal to or greater than a predefined percentage (e.g., 50%), the decision generated from the model indicates a foci. Otherwise, the decision indicates a non-foci (because the probability of output node 1051 (non-foci)<the probability of output node 1052 (foci)).

[0089]

[0102] When the difference histogram is generated based on the reference histogram as previously discussed, a difference histogram is generated for each region relative to the reference histogram. In this case, the model is trained using the generated data set relative to the reference histogram for each ASPECTS region, for example, one reference histogram for the region on the left lateral part of the image slice and another reference histogram for the complementary region on the right lateral part of the image slice. The reference histogram represents the non-lesion data for each of those regions.

[0090]

[0103] Regardless of the manner in which the difference histograms are generated, feature vectors corresponding to the difference histograms are generated and input into the model to provide an indication of whether each corresponding ASPECT region is likely to have a lesion. In one embodiment, the difference histograms for all 10 regions are input through the model. This alleviates the need to perform a pre-filtering operation to discard non-candidate regions (e.g., regions with histograms that differ only within a first range, as discussed above), thus resulting in the generation of probability determinations for all 10 regions. In such cases, a comprehensive data set is provided to the medical practitioner for review and use in determining treatment options.

[0091]

[0104] The model is trained based, at least in part, on the manner in which the difference histograms are generated. In one embodiment, the neural network has a total of 269 weights (e.g., trainable parameters) to ensure that overfitting is not an issue. The neural network is trained, e.g., based on PyTorch as a backend, by using its built-in cross-entropy loss function and by using loss weights of 1 and 3 for the “non-focal” and “focal” classes to account for class imbalance in the training dataset. During training, the Adam optimizer is run over a 2×10 -4 In addition, batch normalization is used on the second and third convolutional layers, with a batch size of 16 (histograms).

[0092]

[0105] During training, we introduce random noise (e.g., centered at zero, 3 × 10 -5 Some data augmentation is done by adding to the difference histogram (sampled from a normal distribution with amplitude of 100 Hz) and by reversing the sign of the difference histogram (corresponding to left and right flips on the image) with a 50% probability.

[0093]

[0106] The data used to train the CNN lesion detection classifier includes over 100 non-contrast CT datasets, e.g., 115 datasets were used in real cases. Some of these datasets correspond to datasets used to train models to perform region segmentation. The kVp values ​​are distributed, e.g., from 100 (n=25) to over 120 (n=45) to 140 (n=45).

[0094]

[0107] 12, 13, and 14 show examples of variability in image parameters used during training. In FIG. 12, image dimensions range between (mean±standard deviation) 508.71±13.79 in x-direction, 511.26±21.88 in y-direction, and 64.27±76.18 in z-direction. In FIG. 13, voxel spacing is (mean±standard deviation) 0.46±0.04 mm in x-direction, 0.46±0.04 mm in y-direction, and 3.59±1.13 mm in z-direction. In FIG. 14, CNN selection uses a training set of 33 examples (real examples) or more with and without gantry tilt.

[0095]

[0108] Although some of the system and method embodiments have been described as using two images (e.g., an inferior image slice and an superior image slice) to identify lesions and other brain abnormalities, other embodiments are implemented to identify such abnormalities using only one image slice. In this case, not all 10 ASPECTS regions are taken into account. However, the embodiments are also implemented in a meaningful manner as previously described relative to one image slice to identify ischemic lesions and other brain abnormalities.

[0096]

[0109] FIG. 15 illustrates an embodiment of a medical image analyzer 1500 for implementing embodiments of the methods described herein. Referring to FIG. 15, the medical image analyzer 1500 comprises a controller 1510 and a memory 1520. The controller executes instructions stored in the memory to perform the operations and methods described herein. In this embodiment, the instructions stored in the memory 1520 include a first set of instructions 1521 for implementing a histogram generator, a second set of instructions 1522 for implementing a difference histogram generator, a third set of instructions 1523 for implementing a classifier, and a fourth set of instructions 1524 for implementing a decision generator. Each of these sets of instructions performs operations on features in, for example, the medical image analyzer of FIG. 1.

[0097]

[0110] In another embodiment, the two histograms are fed directly into the classifier instead of the difference histogram, which in this situation is computed implicitly by the training model that is trained using the histogram pair.

[0098]

[0111] In accordance with one or more of the foregoing embodiments, the methods, processes, and / or operations described herein are effected by code or instructions to be executed by a computer, processor, controller, or other signal processing device, either as described herein or in addition to the elements described herein. The algorithms underlying the methods (or the operation of the computer, processor, controller, or other signal processing device) are described in detail, and the code or instructions for implementing the operations of the method embodiments transform the computer, processor, controller, or other signal processing device into a dedicated processor for performing the methods described herein.

[0099]

[0112] Further, another embodiment includes a computer readable medium, e.g., a non-transitory computer readable medium, for storing the above described code or instructions. The computer readable medium is a volatile memory, non-volatile memory, or other storage device that is removably or permanently coupled to a computer, processor, controller, or other signal processing device that can execute the code or instructions to perform operations of the system and method embodiments described herein.

[0100]

[0113] The processors, systems, controllers, other signal generating features, and signal processing features of the embodiments described herein may be implemented in logic including, for example, hardware, software, or both. When implemented at least partially in hardware, the processors, systems, controllers, other signal generating features, and signal processing features may be any one of a variety of integrated circuits including, for example, but not limited to, application specific integrated circuits, field programmable gate arrays, combinations of logic gates, systems on a chip, microprocessors, or other types of processing or control circuitry.

[0101]

[0114] When implemented at least partially in software, the processor, system, controller, and other signal generating and processing features may include, for example, memory or other storage devices for storing code or instructions executable by, for example, a computer, processor, microprocessor, controller, or other signal processing device, such as those described herein or in addition to the elements described herein. Algorithms forming the basis of the methods (or the operation of a computer, processor, microprocessor, controller, or other signal processing device) are described in detail, and code or instructions for implementing the operations of the method embodiments may transform a computer, processor, controller, or other signal processing device into a dedicated processor for performing the methods described herein.

[0102]

[0115] Benefits, advantages, solutions to problems, and any element or elements which cause any benefit, advantage, or solution to occur or become more prominent are not to be construed as critical, required, or essential features or elements of any or all of the claims. The present invention is defined solely by the appended claims, including any amendments made during the pendency of this application and all equivalents of the issued claims.

[0103]

[0116] Although various exemplary embodiments have been described in detail with particular reference to certain exemplary aspects thereof, it should be understood that the invention is capable of other exemplary embodiments and its details can be modified in various obvious respects. As will be apparent to those skilled in the art, variations and modifications can be made while remaining within the spirit and scope of the invention. The embodiments can be combined to form further embodiments. Accordingly, the foregoing disclosure, description, and drawings are for illustrative purposes only and are not intended to limit in any manner the invention as defined by the claims.

Claims

1. receiving an image slice of the brain including the segmented region; forming a first histogram of intensity values ​​for the image slice; forming a second histogram of intensity values ​​for the image slice; and determining an abnormality in the image slice based on the first histogram and the second histogram, wherein the first histogram corresponds to a first segmented region in a first portion of the image slice and the second histogram corresponds to a second segmented region in a second portion of the image slice that is complementary to the first portion.

2. 2. The method of claim 1, further comprising generating at least one difference histogram based on the first histogram and the second histogram, wherein determining the abnormality in the image slice is based on the at least one difference histogram.

3. The step of determining the anomaly comprises: Identifying the at least one difference histogram as having values ​​within a range; and determining the anomaly based on values ​​of the at least one difference histogram within the range.

4. The step of determining the anomaly comprises: Identifying the at least one difference histogram having one or more values ​​that exceed a predetermined reference value; and determining the anomaly based on the one or more values ​​of the at least one difference histogram exceeding the predetermined reference value.

5. The method of claim 2 , wherein generating at least one difference histogram comprises subtracting the intensity values ​​of the first histogram from the intensity values ​​of the second histogram.

6. generating at least one difference histogram; generating a first difference histogram based on a difference between the first histogram and a first reference histogram; and generating a second difference histogram based on a difference between the second histogram and a second reference histogram.

7. the first reference histogram indicates brain tissue that is free of lesions within the first segmented region; The method of claim 6 , wherein the second reference histogram indicates brain tissue that is free of lesions within the second segmented region.

8. determining the anomaly comprises: generating a feature vector based on the at least one difference histogram; inputting the feature vector into a classifier model; and predicting the anomaly based on an output of the classifier model.

9. determining whether the first histogram has a concentration of intensity values ​​within a predetermined range; 3. The method of claim 2, further comprising identifying the image slice as having an old anomaly when the first histogram has the concentration of intensity values ​​within the predetermined range.

10. The method of claim 9 , further comprising generating the difference histogram based on excluding the intensity values ​​within the predetermined range of the first histogram corresponding to the old anomaly.

11. The method of claim 1 , wherein the first segmented region and the second segmented region are complementary ASPECTS regions.

12. a histogram generator that generates a first histogram of intensity values ​​for an image slice and a second histogram of intensity values ​​for the image slice, the image slice including a segmented region of the brain; a decision engine that determines an anomaly in the image slice based on the first histogram and the second histogram; 1. A system for processing medical information, wherein the first histogram corresponds to a first segmented region within a first portion of the image slice and the second histogram corresponds to a second segmented region within a second portion of the image slice that is complementary to the first portion.

13. 13. The system of claim 12, further comprising difference logic that generates at least one difference histogram based on the first histogram and the second histogram, and wherein the decision engine determines an anomaly in the image slice based on the at least one difference histogram.

14. The decision engine Identifying the at least one difference histogram having values ​​within a range; and determining the anomaly based on values ​​of the at least one difference histogram within the range.

15. The decision engine Identifying the at least one difference histogram having one or more values ​​that exceed a predetermined reference value; and determining the anomaly based on the one or more values ​​of the at least one difference histogram exceeding the predetermined reference value.

16. 14. The system of claim 13, wherein the difference logic generates the at least one difference histogram by subtracting the intensity values ​​of the first histogram from the intensity values ​​of the second histogram.

17. The differential logic is generating a first difference histogram based on a difference between the first histogram and a first reference histogram; and generating a second difference histogram based on a difference between the second histogram and a second reference histogram.

18. the first reference histogram showing brain tissue free of lesions within the first segmented region; 18. The system of claim 17, wherein the second reference histogram indicates brain tissue that is free of lesions within the second segmented region.

19. The decision engine generating a feature vector based on the at least one difference histogram; inputting the feature vector into a classifier model; predicting the anomaly based on an output of the classifier model; and The system of claim 13 , wherein the anomaly is determined by:

20. 14. The system of claim 13, wherein the difference logic generates the at least one difference histogram by subtracting the intensity values ​​of the first histogram from the intensity values ​​of the second histogram.

21. determining whether the first histogram has a concentration of intensity values ​​within a predetermined range; 13. The system of claim 12, further comprising identification logic that identifies the image slice as having an old anomaly when the first histogram has the concentration of intensity values ​​within the predetermined range.

22. 22. The system of claim 21, wherein the difference logic generates the at least one difference histogram based on exclusion of the intensity values ​​within the predetermined range of the first histogram corresponding to the older anomaly.

23. A computer program comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method according to any one of claims 1 to 11.

24. A computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform the steps of the method of any one of claims 1 to 11.

Images