WO2016193847A1

WO2016193847A1 - Tumor grading using apparent diffusion co-efficient (adc) maps derived from magnetic resonance (mr) data

Info

Publication number: WO2016193847A1
Application number: PCT/IB2016/052878
Authority: WO
Inventors: Lalit Gupta; Sundararaman VELANDAI KALYANARAMAN
Original assignee: Koninklijke Philips N.V.
Priority date: 2015-06-03
Filing date: 2016-05-18
Publication date: 2016-12-08

Abstract

To perform tumor grading by magnetic resonance (MR) imaging, an apparent diffusion co-efficient (ADC) map (16) is acquired, and a region of interest (ROI) (32) corresponding to a tumor is identified in the ADC map. A heterogeneity map (60) corresponding to the ROI is generated by computing a contrast measure for each pixel of the ROI. A histogram of the computed contrast measures over the heterogeneity map is constructed and, based on a kurtosis metric of the histogram, the tumor is graded. In some embodiments the heterogeneity map is generated based on co-occurrence matrices (44) with different offsets generated for sub-matrices (40) spanning the ROI. In other embodiments the heterogeneity map is generated by computing texture features for sub-matrices spanning in the ROI. The pixels of the heterogeneity map are suitably contrast measures computed using the co-occurrence matrices or texture features.

Description

TUMOR GRADING USING APPARENT DIFFUSION CO-EFFICIENT (ADC) MAPS DERIVED FROM MAGNETIC RESONANCE (MR) DATA

BACKGROUND

The following relates to the medical imaging arts, oncology arts, tumor assessment arts, and so forth.

Magnetic resonance (MR) imaging is a known technique for identifying and monitoring cancerous tumors. MR imaging is non-invasive, or at most minimally invasive if a magnetic contrast agent is administered. Conventionally, an MR image of a tumor is visually assessed by a radiologist or oncologist having the requisite expertise. The MR imaging assessment can provide quantitative information on tumor size, location, and other anatomical characteristics. Tumor (or cancer) staging may be performed using MR imaging, based on MR-observed anatomical features such as tumor location, tumor size, observed tumor spread to nearby tissues, the number of tumors present, and so forth.

Another type of tumor assessment is known as tumor grading. In this technique, a sample of the tumor tissue is drawn by a biopsy or other invasive procedure, and the tumor sample is examined under a microscope in a technique known as tumor histology. Roughly speaking, the tumor grade is an assessment of how well (or poorly) differentiated the tumor cells appear to be under the microscope. Well-differentiated tumor tissue consists of cells that are visually close to those of normal tissue, and which are organized similarly to normal tissue. On the other hand, poorly differentiated or undifferentiated tumor tissue consists of cells that deviate strongly from normal tissue cells, and have reduced or non-existent organization in the tumor tissue. Tumor grade is useful because it has been generally observed that well-differentiated malignant tissue tends not to spread quickly; whereas, poorly differentiated (or undifferentiated) malignant tissue has a greater tendency to spread in the body and possibly metastasize. A low grade tumor is well differentiated and tends to have a better prognosis; whereas, a high grade tumor is poorly differentiated or undifferentiated and tends to spread more rapidly, leading to a poorer prognosis. Various cancer-type specific tumor histology grading schemes have been developed along these lines for cancers such as brain cancer, breast cancer, prostate cancer, and so forth. The tumor staging is distinct from the tumor grading. Neither the tumor stage, nor the tumor grade, is by itself sufficient to diagnose the patient, determine prognosis, or guide treatment. Rather, the oncologist typically uses both imaging-based tumor staging and tumor histology grading, along with other information such as patient age, general health, demographic information, and so forth, in order to diagnose the cancer, assess patient prognosis and formulate and guide the oncology treatment. This requires that the patient undergo both MR imaging and a biopsy or other invasive procedure.

The following contemplates improved apparatuses and methods that overcome the aforementioned limitations and others.

BRIEF SUMMARY

According to one aspect, a tumor grading system comprises a computer programmed to grade a tumor by performing a tumor grading method on a region of interest (ROI) corresponding to the tumor in a magnetic resonance (MR) image. The tumor grading method includes the operations of: generating a heterogeneity map corresponding to the ROI by computing a contrast measure for each pixel of the ROI; constructing a histogram of the computed contrast measures over the heterogeneity map; computing a kurtosis metric of the histogram; and grading the tumor based on the kurtosis metric. The system may further comprise a user interface including at least one user input device and a display device, wherein the tumor grading method includes the further operation of receiving, via the user interface, identification of the ROI corresponding to the tumor in the MR image. The system may further include an MR imaging system, wherein the tumor grading method includes the further operation of causing the MR imaging system to acquire the MR image as an apparent diffusion co-efficient (ADC) map.

According to another aspect, a tumor grading method comprises: acquiring an apparent diffusion co-efficient (ADC) map using a magnetic resonance (MR) imaging system; receiving identification of a region of interest (ROI) in the ADC map corresponding to a tumor via a user interface device; and, using a computer: generating a heterogeneity map corresponding to the ROI by computing a contrast measure for each pixel of the ROI, constructing a histogram of the computed contrast measures over the heterogeneity map, computing a kurtosis metric of the histogram, and displaying, on the user interface, a tumor grade determined based on the kurtosis metric. In some embodiments the heterogeneity map is generated by computing co-occurrence matrices with different offsets for sub-matrices defined in the ROI, and computing pixels of the heterogeneity map as contrast measures computed using the co-occurrence matrices. In other embodiments the heterogeneity map is generated by computing texture features for sub-matrices defined in the ROI and computing pixels of the heterogeneity map as contrast measures computed using the texture features computed for the sub-matrices.

According to another aspect, a non-transitory storage medium stores instructions readable and executable by a computer to perform a tumor grading method including the operations of: generating statistics on pixel gray level co-occurrences in a region of interest (ROI) corresponding to a tumor in an apparent diffusion co-efficient (ADC) map; and assigning a tumor grade to the tumor based on the generated statistics on pixel gray level co-occurrences in the ROI. In some embodiments the operation of generating statistics on pixel gray level co-occurrences in the ROI includes generating a set of co-occurrence matrices with different offsets for each of a plurality of sub-matrices spanning the ROI. In some embodiments the operation of assigning a tumor grade to the tumor based on the generated statistics on pixel gray level co-occurrences in the ROI includes generating a contrast map for the ROI based on the generated statistics on pixel gray level cooccurrences in the ROI, computing a kurtosis metric for a histogram of constrast values in the contrast map, and assigning the tumor grade to the tumor based on the computed kurtosis metric.

One advantage resides in providing non-invasive tumor grading using MR imaging that correlates well with conventional tumor histology grading.

Another advantage resides in providing MR imaging-based tumor grading without the administration of a magnetic contrast agent.

Another advantage resides in providing MR imaging-based tumor grading that is readily tuned for different cancer types by tuning the post-acquisition image processing.

Numerous additional advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

The present application may take form in various components and arrangements of components, and in various process operations and arrangements of process operations. The drawings are only for the purpose of illustrating preferred embodiments and are not to be construed as limiting the present application.

FIGURE 1 diagrammatically shows magnetic resonance (MR) imaging-based tumor grading system.

FIGURE 2 diagrammatically illustrates certain data structures generated and used by the MR imaging-based tumor grading system of FIGURE 1.

FIGURE 3 : upper plot shows a grading scale in which the tumor grade is monotonically non-decreasing with increasing value of the kurtosis metric; lower table presents the tumor grade definitions.

FIGURE 4 shows two examples of apparent diffusion co-efficient (ADC) maps of brain tumors and their corresponding heterogeneity/contrast maps generated by the MR imaging-based tumor grading system of FIGURE 1.

FIGURE 5 compares the output of the MR imaging-based tumor grading system of FIGURE 1 with three other tested MR-based tumor grading techniques as described herein, with tumor grades assigned by tumor histopathology indicated by circled numeral labels. DETAILED DESCRIPTION

Disclosed herein are approaches for grading a tumor based on texture analysis of a magnetic resonance (MR) image of the tumor. In illustrative embodiments, the texture analysis is performed on an MR image comprising an apparent diffusion co-efficient (ADC) map. The ADC map or image is a map of the magnitude of water diffusion in the tissue. Various diffusion weighted imaging (DWI) techniques can be employed to acquire an ADC map. For example, in DWI using the field gradient pulse technique, magnetic field gradient pulses are applied in a first direction to obtain sensitivity to diffusion along that direction, and this is repeated for at least two other directions (e.g. for three orthogonal directions total) to obtain the ADC map.

Local water diffusion is expected to be affected by local cellular structure and organization. Hence, the ADC map might be expected to be characteristic of these tissue features. As tumor grade is a metric of the tumor cellular structure/organization, it might thus be expected that an ADC map of the tumor contains information on the tumor grade.

However, in brain tumor MR imaging experiments (described herein with reference to FIGURE 4), it was found that the mean value, skewness, and kurtosis characteristics of the brain tumor ADC map did not correlate with tumor grade.

It was found, however, that a combination of applying a heterogeneity (or contrast) analysis as disclosed herein to a brain tumor ADC map did provide a strong correlation with brain tumor grade. Without being limited to any particular theory of operation, it is believed that higher heterogeneity corresponds to higher degree of variability within the tumor. Multiple factors could contribute to this increased variability e.g. areas of different tumor grades, necrotic tissue, hemorrhage within the tumor, increased vascularity. Higher prevalence of these factors is associated with an aggressively growing tumor which is poorly differentiated or undifferentiated (i.e. of high tumor grade).

In one approach, the heterogeneity or contrast map can be produced by generating statistics on pixel gray level co-occurrences in a region of interest (ROI) corresponding to the tumor in the ADC map. For example, co-occurrence matrices with different offsets can be computed for sub-matrices defined in the ROI, and pixels of the heterogeneity map are then computed as contrast measures for the co-occurrence matrices. In one approach, each sub-matrix is an N><N sub-matrix associated with, and centered on, a pixel of the ROI (where N=3, or N=5, or N=7, or N=9, or so forth) and the N><N sub-matrices is effectively a kernel that is repeated over two dimensions so as to span a two-dimensional ROI.

In some more specific embodiments tested herein, it was found that a kurtosis metric applied to the contrast values distribution of the heterogeneity or contrast map provides a scalar value that scales in a monotonic non-decreasing fashion with tumor grade. In general, kurtosis is a measure of the extent to which a distribution is concentrated about its mean, that is, a measure of the extent to which a distribution is peaked. Another way of viewing kurtosis is as a measure of the degree of curvature of the peak of the distribution, since a "more peaked" distribution has a higher curvature at its peak than a broader, less peaked distribution. Various kurtosis metrics are known. One kurtosis metric is the ratio of the fourth moment of the distribution (μ⁴) to the square of the second moment (variance σ^Δ) of the distribution, that is, this kurtosis metric is given as A variant of this metric is known as the excess kurtosis (or Pearson kurtosis), and is given by μ⁴

— -Γ- — 3 (where the deduction "-3" sets the excess kurtosis to zero for the normal

(σ²)²

distribution). Other kurtosis metrics that are sometimes used include the fourth moment μ⁴ by itself, or an L-moment. These are illustrative examples: in general, a higher value of a kurtosis metric indicates a more peaked distribution (more heavily concentrated toward its mean) having a higher degree of curvature (that is, more sharply peaked).

With reference to FIGURE 1, an illustrative magnetic resonance (MR) imaging- based tumor grading system is described. An MR scanner 10 is controlled by an MR controller 12 to execute an apparent diffusion co-efficient (ADC) mapping pulse sequence 14 in order to acquire an ADC map 16 of a subject (e.g. a human oncology patient, or a dog, cat, or other veterinary oncology subject). The subject is positioned in the examination region of the MR scanner 10 so that the field of view (and hence the ADC map) encompasses the tumor of interest. The illustrative MR scanner 10 is a Philips Achieva 1.5T MR scanner (available from Koninklijke Philips N.V., Eindhoven, the Netherlands); however, the MR scanner 10 can more generally be any other commercially available MR imaging system, or may be a custom-built MR scanner, that is capable of acquiring the ADC image or map 16. As is known in the art, the ADC image or map 16 is typically acquired without the use of any exogenous contrast agent being administered to the subject, as the ADC map 16 characterizes endogenous or intrinsic water diffusion.

The MR controller and other associated electronic data processing and/or control components are embodied in the illustrative embodiment by a computer 20, which includes a display device 22 (e.g. an LCD display device) and at least one user input device 24 (e.g. an illustrative keyboard, and/or a mouse, trackball, or other pointing device, and/or a touch- sensitive overlay of the display, or so forth). The computer 20 provides user interfacing to enable a radiologist or other trained medical professional to operate the MR scanner 10 to execute the pulse sequence 14 to acquire the ADC map 16 (and, typically, also enables the radiologist to execute myriad other pulse sequences to acquire other types of MR images or data, e.g. Tl -weighted images, proton images, magnetic angiography images, or so forth). The illustrative computer 20 also provides a user interface 30, employing the display device 22 and the at least one user input device 24, via which a user can navigate through the ADC images 16 in order to select the slices that pass through a (suspected) malignant tumor of interest and to identify within each slice a region of interest (ROI) 32 corresponding to the tumor in the ADC image 16. It will also be appreciated that the disclosed image acquisition and processing operations may be physically embodied as a non-transitory storage medium that stores instructions readable and executable by a computer (e.g. computer 20) to perform a tumor grading method as disclosed herein. The non-transitory storage medium may, for example, include one or more of a hard drive or other magnetic storage medium, an EPROM, flash memory, or other electronic storage medium, an optical disk or other optical storage medium, various combination(s) thereof, or so forth.

Numerous medical imaging navigation interfaces may be employed as the user interface 30 - in a typical implementation, a stack of acquired ADC images 16 is displayed, and the user uses a mouse or other pointing device to select relevant slices, which is shown on the display in a separate window along with the ADC images 16. The user (preferably a radiologist, oncologist, or other medical professional trained to interpret MR images with respect to tumors or other oncological features) may then "flip through" slices until all slices that cover the tumor are chosen. The, the mouse or other user input device is employed to draw a contour around the tumor in each chosen slice, thereby identifying the ROI 32 corresponding to the tumor. Additionally or alternatively, automated or semi-automated segmentation and/or contouring may be provided as part of the user interface 30 to provide automated assistance in identifying the ROI 32 - for example, an edge detector-based contour delineation algorithm may be employed, in which the user defines an approximate encircling contour which is then automatically adjusted to align with a high intensity gradient corresponding to an "edge" in the image.

The illustrative computer 20 provides both MR control and image reconstruction functionality (that is, the computer 20 is programmed to implement the MR controller 12 to control the MR scanner 10 and to reconstruct acquired MR imaging data to generate the ADC image 16, or another chosen type of MR image) and post-acquisition image processing functionality such as the image navigation/ROI selection component 30 and subsequent image and data processing components such as described herein. It will be appreciated that the illustrative computer 20 may be replaced by two or more computers or other electronic data processing device(s). For example, in an alternative implementation, the MR controller 12 is implemented as a control computer and a separate image reconstruction computer with high computing capacity, and the post-image acquisition processing is either also performed by the image reconstruction computer, or is peformed by a separate (third) computer dedicated to post-acquisition data processing tasks (for example, implemented as an image navigation workstation).

With continuing reference to FIGURE 1, the input ADC maps 16 typically consists of stack of two-dimensional images, although acquisition of three-dimensional MR data is also contemplated, and the illustrative ROI 32 is a two-dimensional ROI defined in each slice. However, a three-dimensional ROI is also contemplated. The ADC map 16 and the ROI 32 is optionally used for various operations in addition to the tumor grading process. For example, FIGURE 1 illustrates the tumor as represented by the ROI 32 and the surrounding anatomical context as represented by the ADC map 16 being used as input for staging of the tumor or cancer in a staging operation 34. Typically, a tumor is staged based on its size, location respective to critical tissues, and other factors. The overall cancer may be similarly staged based on the size/location of the principal tumor or tumors, along with other information such as the total number of tumors (obtainable from the ADC map 16 analyzed by the radiologist using the user interface 30), the patient's age, gender, general health, or other patient information, and so forth. Although the illustrative staging process 34 uses the ADC image 16, in other embodiments the MR scanner 10 may execute another pulse sequence under control of the controller 12 in order to acquire another type of image (e.g. a proton image, Tl -weighted image, an image acquired using a contrast agent, or so forth) which is then used for staging the tumor and/or overall cancer.

With continuing reference to FIGURE 1, the cancer grading process is performed in the illustrative example in accord with the blocks shown in the righthand column of FIGURE 1, based on a statistical analysis of co-occurrences of gray levels in the ROI 32. A "co-occurrence" of a gray scale level refers to two pixels of an image (e.g. of ROI 32) separated by a chosen offset having designated gray scale levels. For example, the offset may be chosen as "immediately horizontally to the right", the designated gray scale level may be "2" and "2" (the same in this case), and the number of co-occurrences of two pixels with gray scale level "2" immediately next to each other horizontally is then counted. For a given offset, this may be repeated for each possible combination of gray scale level i and gray scale level j, thus producing a two-dimensional "co-occurrence matrix" having axes for i and j, respectively. If the gray scale levels range from 0, ... ,255 then the co-occurrence matrix would have dimensions 256x256. In practice, if the number of gray scale levels is large then the co-occurrence matrix may be sparse - in illustrative FIGURE 1 this sparseness is suppressed by reducing the number of gray scale levels via a re-scaling process 38. For example, by rescaling levels 0...7 as a first level, levels 8... 15 as a second level, up to levels 248...255 as an eighth level, the number of gray levels is reduced from 256 to 8, and the resulting co-occurrence matrix is of dimensions 8x8 with substantially reduced sparseness compared with the 256x256 co-occurrence matrix of the ROI without such gray level re-scaling.

With continuing reference to FIGURE 1 and with further reference to FIGURE 2, an approach for statistically analyzing co-occurrences of gray scale levels in order to generate a heterogeneity or contrast image or map for the ROI 32 is described. In FIGURE 2, the ROI 32 (after the optional re-scaling operation 38, if applied) is shown. At each pixel of the ROI 32, a sub-matrix 40 (also called a "kernel" or an "image kernel" herein) is defined. In some embodiments the kernel 40 is an Ν^χΝ matrix centered on the ROI pixel P. To allow centering, N should be an odd integer, e.g. N=3, or N=5, or N=7, or N=9, et cetera. As shown in FIGURE 2, a co-occurrence matrix 42 is computed for the kernel 40 and for an offset d, which in the illustrative example of FIGURE 2 is directed to the upper right and has a offset magnitude d=l (i.e. one pixel). The co-occurrence matrix 42 has dimensions i, j each running over all possible gray scale levels (again, after the re-scaling 38 if such re-scaling is performed). This is optionally repeated for a plurality of different offsets d, in different directions and/or of different magnitudes, and the resulting plurality of (2D) co-occurrence matrices are concatenated along a third direction (designated by index k herein) to generate a 3D co-occurrence matrix 44 for the pixel P of the 2D ROI 32. With reference back to FIGURE 1, in an operation 50 this processing (generating a plurality of 2D co-occurrence matrices at pixel P with different offsets d) is repeated for each pixel of the 2D ROI 32 to generate a representation 52 of the ROI in which each pixel is represented by a 3D co-occurrence matrix of the form of the 3D co-occurrence matrix 44 shown in FIGURE 2.

While the illustrative example of FIGURE 2 defines a sub-matrix at each pixel of the ROI 32, other approaches are contemplated for generating a set of co-occurrence matrices with different offsets for each of a plurality of sub-matrices spanning the ROI 32. For example, in the illustrative example the N><N kernels 40 defined around two immediately adjacent pixels of the ROI 32 will overlap. In another approach, the kernels are chosen to not overlap, but to span the ROI 32. For example, if the ROI is square with 2048x2048 pixels, then a kernel of dimensions 8x8 can be used, with the 2048x2048 pixel ROI divided into 256x256 non-overlapping kernels each of 8x8 pixels. A 3D co-occurrence matrix analogous to the 3D co-occurrence matrix 44 can be computed for each of the 8x8 kernels as just described. In this case, the resulting 2D ROI representation 52 would be made up of 256x256 "pixels" each represented by a 3D co-occurrence matrix analogous to the 3D co-occurrence matrix 44.

With continuing reference to FIGURES 1 and 2, in an operation 54 each "pixel" of the 2D ROI representation 52 is converted to a contrast or texture value using a suitable approach. In an illustrative example shown in FIGURE 2, the contrast for each 3D co-occurrence matrix 44 is computed as:

where indices i, V and j, j' index gray scale levels of the ROI, index k, k' indexes the co-occurrence matrices, and d_k = \d_k \ denotes the offset construct the k co-occurrence matrix. The denomin

normalization so that \i— j\² is scaled by a value that is strictly less than unity, while scaling by l/d_k emphasizes the significance of "closer" co-occurrences over more distant co-occurrences. This computation is repeated for each "pixel" of the representation 52, so that the output of the operation 54 is a heterogeneity or contrast map 60 for the ROI 32. The heterogeneity or contrast map 60 has the same number of pixels as the number of "pixels" of the 2D ROI representation 52 - this number of pixels may be the same as the number of pixels in the ROI 32 (as in the illustrative embodiment), or may be less than the number of pixels in the ROI 32 (as in the additional example in which a 2048x2048 ROI is processed in operation 50 using a 2D array of 8x8 non-overlapping kernels to generate the representation 52 as a 2D map of 256x256 "pixels").

It is to be appreciated that the foregoing is an illustrative example of generating the heterogeneity or contrast map 60 for the ROI 32 By way of illustrative example, in another approach texture features are computed for the image kernel, and a contrast measure for the kernel is computed based on the texture features for the image kernel. Some suitable texture features, defined in terms of co-occurrence statistics (probabilities) are given in Haralick et al, "Textural Features for Image Classification", IEEE Trans. On Systems, Man and Cybernetics vol. SMC-3 No. 6 pages 610-21 (1973). In this approach, the contrast measure may be computed as a weighed combination of the texture features with higher contrast texture features being assigned relatively higher weights in the combination than lower contrast texture features.

With continuing reference to FIGURE 1, the heterogeneity/contrast map 60 for the

2D ROI 32 is processed to compute a grade for the tumor as follows. In an operation 62, a contrast histogram is computed for the contrast measures of the heterogeneity/contrast map 60. Each contrast bin of the contrast histogram stores a count of the number of pixels of the heterogeneity/contrast map 60 having contrast measure values falling in that contrast bin. The contrast bins of the contrast histogram are optionally normalized, e.g. by the total number of pixels in the heterogeneity/contrast map 60. In an operation 64, a kurtosis metric is computed for the distribution represented by the contrast histogram. As previously described, some suitable illustrative kurtosis metrics include: the excess kurtosis of the histogram; the ratio where μ is the fourth moment of the histogram and σ^Δ is the variance of the histogram; the fourth moment μ⁴ of the histogram; or an L-moment of the histogram. The kurtosis metric is a single scalar value representing the kurtosis of the distribution of contrast measures over the heterogeneity/contrast map 60 generated for the ROI 32. In general, a higher value of a kurtosis metric indicates a more peaked contrast histogram (more heavily concentrated toward its mean contrast value) having a higher degree of curvature (that is, more sharply peaked). Without being limited to any particular theory of operation, the correlation of the kurtosis metric computed in operation 64 with the tumor grade can be conceptually understood as follows. If the tumor is poorly differentiated or undifferentiated (that is, high tumor grade), then the contrast typically varies to a greater extent across the tumor (i.e. across the ROI 32). This leads to a contrast distribution (histogram) with one high peak (sharp spike) and small peaks around it, resulting in a high kurtosis value because of one high peak. On the other hand, if the tumor is well-differentiated (that is, low tumor grade), then the contrast should be roughly uniform due to the presence of well-defined tissue. This leads to a broader, rounder contrast distribution with no sharp peak, resulting in low kurtosis.

In another embodiment (alternative implementation), the x-axis of the contrast distribution histogram is normalized within a range. Here the high grade tumor would yield one sharp peak and small peaks around it, however low grade tumor would yield one peak with very few/no small peaks around it. Therefore, in this implementation low grade tumor will have higher kurtosis and high grade tumor will show low kurtosis.

With continuing reference to FIGURE 1 and with further reference to FIGURE 3, a suitable grading scheme based on the kurtosis measure for the heterogeneity/contrast map 60 is described. FIGURE 3, upper plot, shows a grading scale in which the tumor grade is monotonically non-decreasing with increasing value of the kurtosis metric. Representing the value of the kurtosis metric as K, in this grading scheme the tumor grade (in a range 1-4) is assigned as follows:

where T_±, T₂, T₃ are thresholds which are suitably determined by a calibration process in which a training set of oncology patients are both imaged to acquire ADC maps which are processed as described with reference to FIGURES 1 and 2 in order to generate kurtosis metric values for the patients, and also are biopsied to perform tumor histology grading to generate "ground truth" tumor grades. Thresholds Ύ , T₂, T₃ are then chosen to optimally correlate the tumor grade assigned based on kurtosis metric value with tumor histology grade for this training set. It will be appreciated that such training is readily performed for various types of tumors, e.g. brain tumors, breast cancer tumors, and so forth, and the kurtosis metric thresholds are then optimized to correlate with the specific tumor histology grading scheme used for each tumor type (where the tumor grade levels may have tumor type-specific nomenclatures).

With returning reference to FIGURE 1, a tumor grading operation 70 suitably uses the grading scheme of FIGURE 3 or the like to assign a tumor grade for the tumor represented by the ROI 32. In an operation 72, the tumor grade is displayed on the display device 22, optionally along with related information such as the heterogeneity/contrast map 60 to allow the radiologist to visually observe the heterogeneity across the tumor.

In this ID, we propose a method for quantifying heterogeneity in MRI. The heterogeneity value can be used to non-invasively infer tumour grade i.e. higher the tumour grade higher will be the heterogeneity.

In the following, actual training and testing of the disclosed MR-based tumor grading system for grading brain tumors is described. In these tests, the MR imaging employed a 1.5 Tesla MR scanner, with data acquired from 20 patients with Glioma brain tumors. Routine anatomic imaging was also performed using standard T1W, T2W, FLAIR and post-contrast T1W sequences. To acquire the ADC map of a patient, diffusion images were acquired by standard EPI-based diffusion weighted imaging (b-values of 0 and 1000). and the ADC map was generated using commercial software on the MR scanner console. The co-occurrence matrices at different offsets d_k were weighted by inverse of distance from the centre pixel i.e. l/d_k. A chessboard distance was used as a distance measure. The co-occurrence matrices were concatenated along the third dimension (k) as described in operation 50 of FIGURE 1 to generate the 3D co-occurrence matrices of size /, /, K, where / and / are number of grey levels used for computing co-occurrence matrices and K represents number of neighbouring voxels or number of offsets. From the 3D co- occurrence matrix obtained above, the contrast metric was computed as described with reference to operation 54. This procedure was performed on each pixel of the ROI to obtain a Heterogeneity (or contrast) map as described with reference to the map 60 of FIGURE 1.

With brief reference to FIGURE 4, two examples are shown of this result. In FIGURE 4, the left-hand image is the ADC map, while the right-hand image is the generated heterogeneity/contrast map (with the portions of the ADC map outside of the ROI shown in black).

To assess the Gioma brain tumor grades, the Heterogeneity kurtosis was computed from the histogram of heterogeneity map for the ROI corresponding to the brain tumor as already described with reference to FIGURE 1. As baseline comparisons: the mean value of the ADC map in the ROI ("ADC mean"), the skewness of the ADC map in the ROI ("ADC skewness") and the kurtosis of the ADC map in the ROI ("ADC kurtosis") were also computed. (Note that here the term "heterogeneity kurtosis" is used for the kurtosis metric computed from the heterogeneity map 60 in operation 64, to distinguish from the "ADC kurtosis" computed directly from the ADC map values.) Additionally, the tumor grade of each patient was assessed by surgical resection and tumor histology grading to obtain the "ground truth" tumor grade.

With reference to FIGURE 5, the results of these experiments are summarized in bar graphs. For each metric ("ADC mean", "ADC skewness", "ADC kurtosis", and "heterogeneity kurtosis"), four bars are given, corresponding to the ground truth tumor grades 1-4 obtained by histology. For each ground truth grade, the aggregate value of the metric for all patients with that grade is represented by the height of the bar. For the heterogeneity kurtosis, good correlation is seen: tumors with grade 1 have the lowest heterogeneity kurtosis, and as the tumor grade increases from 1 through 2, 3, and 4 the aggregate heterogeneity kurtosis value increases accordingly.

By contrast, no such correlation is seen for the other metrics. In the case of

ADC mean, the patient group with tumor grade 3 has the highest aggregate ADC mean, while the patient group with tumor grade 4 has a lower aggregate ADC mean. In the case of ADC skewness, the patient group with tumor grade 1 has the highest skewness. In the case of ADC kurtosis, tumor grades 1 and 3 show the highest aggregate ADC kurtosis.

The illustrative tumor scoring examples score the tumor based on peakedness, i.e. kurtosis metric, computed for the heterogeneity/contrast map that is generated from an ADC map. However, it is contemplated to apply this analysis to other types of MR images, besides an ADC map, exhibiting isotropic textural characteristics correlative with tissue differentiation.

The disclosure has been set forth with reference to the preferred embodiments. Obviously, modifications and alterations will occur to others upon reading and understanding the preceding detailed description. It is intended that the present application be construed as including all such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.

Claims

1. A tumor grading system comprising:

a computer (20) programmed to grade a tumor by performing a tumor grading method on a region of interest (ROI) (32) corresponding to the tumor in a magnetic resonance (MR) image (16), the tumor grading method including the operations of:

generating a heterogeneity map (60) corresponding to the ROI by computing a contrast measure for each pixel of the ROI;

constructing a histogram of the computed contrast measures over the heterogeneity map;

computing a kurtosis metric of the histogram; and

grading the tumor based on the kurtosis metric.

2. The tumor grading system of claim 1 further comprising:

a user interface including a display device (22) and at least one user input device (24), wherein the tumor grading method includes the further operation of receiving, via the user interface, identification of the ROI corresponding to the tumor in the MR image.

3. The tumor grading system of claim 2 wherein the tumor grading method performed by the computer further includes the operation of displaying a tumor grade generated by the grading along with the generated heterogeneity map on the display device of the user interface.

4. The tumor grading system of any one of claims 1-3 further comprising:

a magnetic resonance (MR) imaging system (10. 12);

wherein the tumor grading method includes the further operation of causing the MR imaging system to acquire the MR image as an apparent diffusion co-efficient (ADC) map (16).

5. The tumor grading system of any one of claims 1-4 wherein the operation of grading the tumor based on the kurtosis metric comprises: assigning a tumor grade to the tumor using a grading scale in which the tumor grade is monotonically non-decreasing with increasing value of the kurtosis metric.

6. The tumor grading system of any one of claims 1-5 wherein the operation of computing a contrast measure for each pixel of the ROI comprises, for each pixel of the ROI:

computing a plurality of co-occurrence matrices (42) with different offsets for an image kernel (40) associated with the pixel; and

computing a contrast measure for the pixel based on the plurality of computed co-occurrence matrices.

7. The tumor grading system of claim 6 wherein the image kernel associated with the pixel is an NxN sub-matrix of pixels of the ROI centered on the pixel where N is an odd integer.

8. The tumor grading system of any one of claims 6-7 wherein the operation of computing a contrast measure for the pixel based on the plurality of co-occurrence matrices comprises computing: contrast

where indices i, V and j, j' index gray scale levels of the ROI, index k, k' indexes the co-occurrence matrices, and d_k denotes an offset distance for the k^th co-occurrence matrix.

9. The tumor grading system of any one of claims 1-5 wherein the operation of computing a contrast measure for each pixel of the ROI comprises, for each pixel of the ROI:

computing texture features for an image kernel associated with the pixel; and computing a contrast measure for the pixel based on the texture features for the image kernel associated with the pixel.

10. The tumor grading system of any one of claims 1-9 wherein the operation of computing a kurtosis metric of the histogram comprises computing the kurtosis metric as one of:

the excess kurtosis of the histogram,

the ratio where is the fourth moment of the histogram and σ^Δ is the variance of the histogram,

the fourth moment μ⁴ of the histogram, and

an L-moment of the histogram.

11. The tumor grading system of any one of claims 1-10 wherein the tumor grading method including the further operation of, prior to generating the heterogeneity map:

re-scaling gray scale levels of the ROI.

12. A tumor grading method comprising:

acquiring an apparent diffusion co-efficient (ADC) map (16) using a magnetic resonance (MR) imaging system (10, 12);

receiving identification of a region of interest (ROI) (32) in the ADC map corresponding to a tumor via a user interface device (22, 24); and

using a computer (20):

computing a kurtosis metric of the histogram; and

displaying, on the user interface device, a tumor grade determined based on the kurtosis metric.

13. The tumor grading method of claim 12 wherein generating the heterogeneity map comprises: computing co-occurrence matrices (44) with different offsets for sub-matrices (40) defined in the ROI (32); and

computing pixels of the heterogeneity map (60) as contrast measures computed using the co-occurrence matrices.

14. The tumor grading method of claim 12 wherein generating the heterogeneity map comprises:

computing texture features for sub-matrices defined in the ROI; and

computing pixels of the heterogeneity map as contrast measures computed using the texture features computed for the sub-matrices.

15. The tumor grading method of any one of claims 12-14 wherein computing the kurtosis metric of the histogram comprises one of:

computing the excess kurtosis of the histogram,

computing the ratio where is the fourth moment of the histogram and σ^Δ is the variance of the histogram,

computing the fourth moment μ⁴ of the histogram, and

computing an L-moment of the histogram.

16. A non-transitory storage medium storing instructions readable and executable by a computer (20) to perform a tumor grading method including the operations of:

generating statistics on pixel gray level co-occurrences in a region of interest (ROI) (32) corresponding to a tumor in an apparent diffusion co-efficient (ADC) map (16); and

assigning a tumor grade to the tumor based on the generated statistics on pixel gray level co-occurrences in the ROI.

17. The non-transitory storage medium of claim 16 wherein the tumor grading method performed by the computer reading and executing the stored instructions includes the further operations of: causing a display device (22) to display the ADC map (16);

receiving, via a user input device (24), an identification of the ROI (32) corresponding to the tumor in the displayed ADC map; and

causing the display device to display the assigned tumor grade.

18. The non-transitory storage medium of any one of claims 16-17 wherein the operation of generating statistics on pixel gray level co-occurrences in the ROI includes: generating a set of co-occurrence matrices (44) with different offsets for each of a plurality of sub-matrices (40) spanning the ROI (32).

19. The non-transitory storage medium of any one of claims 16-18 wherein the operation of assigning a tumor grade to the tumor based on the generated statistics on pixel gray level co-occurrences in the ROI includes:

generating a contrast map for the ROI based on the generated statistics on pixel gray level co-occurrences in the ROI;

computing a kurtosis metric for a histogram of constrast values in the contrast map; and

assigning the tumor grade to the tumor based on the computed kurtosis metric.

20. The non-transitory storage medium of any one of claims 16-19 wherein the stored instructions are further readable and executable by a computer perform a tumor staging method (34) based at least on size and location of the ROI (32) corresponding to the tumor in the displayed ADC map (16).