Predicting prostate cancer risk of progression with multiparametric magnetic resonance imaging using machine learning and peritumoral radiomics

1. A non-transitory computer-readable storage device storing computer-executable instructions that when executed cause a processor to perform operations, the operations comprising:

accessing a multi-parametric magnetic resonance imaging (MRI) image associated with a patient, where the image includes a region of interest (ROI) demonstrating prostate cancer (PCa) pathology, the image having a plurality of pixels, a pixel having an intensity;

segmenting a tumoral region represented in the image, where segmenting the tumoral region includes defining a tumoral boundary;

defining a peritumoral region based on the tumoral boundary;

extracting a set of radiomic features from the image, where the set of radiomic features includes at least one radiomic feature extracted from the peritumoral region, and at least one, different radiomic feature extracted from the tumoral region;

providing the set of radiomic features to a machine learning classifier trained to distinguish a first class from a second, different class based on the set of radiomic features;

receiving, from the machine learning classifier, a probability that the ROI is a member of the first class, where the machine learning classifier computes the probability based on the set of radiomic features;

classifying the ROI as a member of the first class or the second, different class based, at least in part, on the probability; and

displaying the classification,

where defining the peritumoral region comprises performing a dilation of the tumoral boundary, where the peritumoral region includes a plurality of annular rings.

2. The non-transitory computer-readable storage device of claim 1 , where the multi-parametric MRI image is a bi-parametric MRI (bpMRI) image comprising a T2W MRI image and an apparent diffusion coefficient (ADC) map.

3. The non-transitory computer-readable storage device of claim 1 , where performing a dilation of the tumoral boundary comprises dilating the tumoral boundary 12 mm, and where the plurality of annular rings comprises four annular rings.

4. The non-transitory computer-readable storage device of claim 1 , where the machine learning classifier is a quadratic discriminant analysis (QDA) classifier.

5. The non-transitory computer-readable storage device of claim 1 , the operations further comprising:

generating a personalized treatment plan based, at least in part, on the classification; and

displaying the personalized treatment plan.

6. The non-transitory computer-readable storage device of claim 1 , the operations further comprising training the machine learning classifier on a multi-institutional cohort.

7. The non-transitory computer-readable storage device of claim 2 , where the first class is low risk of progression, and where the second class is high risk of progression.

8. The non-transitory computer-readable storage device of claim 2 , where the first class is low risk of progression, and where the second class is intermediate risk of progression or high risk of progression.

9. The non-transitory computer-readable storage device of claim 2 , the operations further comprising preprocessing the image.

10. The non-transitory computer-readable storage device of claim 6 , the operations further comprising testing the machine learning classifier on a multi-institutional testing cohort.

11. The non-transitory computer-readable storage device of claim 7 , where the set of radiomic features includes an intratumoral Laws ( 15 ) feature extracted from the T2W image, an intratumoral Canny feature extracted from the T2W image, a peritumoral co-occurrence of local anisotropic gradient orientations (CoLIAGe) entropy feature extracted from the ADC map, an intratumoral Laws ( 11 ) feature extracted from the ADC map, an intratumoral Haralick (entropy) feature extracted from the ADC map, an intratumoral CoLIAGe feature extracted from the ADC map, a peritumoral Haralick (info measure 1 ) feature extracted from the T2W image, a peritumoral Laws ( 17 ) feature extracted from the ADC map, an intratumoral Haralick (info measure 2 ) feature extracted from the T2W image, and an intratumoral Haralick (info measure 2 ) feature extracted from the ADC map.

12. The non-transitory computer-readable storage device of claim 8 , where the set of radiomic features includes an intratumoral Gabor (6 Hz, 2.0 rad) feature extracted from the T2W image, a peritumoral Gabor (6 Hz, 2.8 rad) feature extracted from the T2W image, a peritumoral Haralick (momentum sum) feature extracted from the ADC map, an intratumoral Gabor (6 Hz, 1.8 rad) feature extracted from the ADC map a mean filter (3 pixel by 3 pixel window) response feature extracted from the peritumoral region of the T2W image, an intratumoral Gabor (2.5 Hz, 0.4 rad) feature extracted from the T2W image, an intratumoral Gabor (3 Hz, 0.4 rad) feature extracted from the T2W image, an intratumoral Gabor (3.5 Hz, 0.4 rad) feature extracted from the T2W image, an intratumoral Gabor (5 Hz, 1.6 rad) feature extracted from the ADC map, and an intratumoral Gabor (6 Hz, 1.6 rad) feature extracted from the ADC map.

13. The non-transitory computer-readable storage device of claim 9 , where preprocessing the image comprises:

resampling the image to a pixel size of 0.5×0.5 mm 2 ;

cropping the ROI with 2 mm padding along the x axis and y axis;

interpolating the image to a 3 mm slice thickness;

verifying that the T2W image and ADC map represent the same three dimensional (3D) space; and

correcting scanner variability using a drift correction approach.

14. An apparatus comprising:

a processor;

a memory configured to store a bi-parametric magnetic resonance imaging (bpMRI) image associated with a patient, where the image includes a region of interest (ROI) demonstrating prostate cancer (PCa) pathology, the bpMRI image having a plurality of pixels, a pixel having an intensity, the bpMRI image comprising a T2W MRI image and an apparent diffusion coefficient (ADC) map;

an input/output (I/O) interface;

a set of circuits; and

an interface that connects the processor, the memory, the I/O interface, and the set of circuits, the set of circuits comprising:

an image acquisition circuit configured to:

access the bpMRI image;

a region definition circuit configured to:

segment a tumoral region represented in the bpMRI image, where segmenting the tumoral region includes defining a tumoral boundary; and

define a peritumoral region based on the tumoral boundary, where the peritumoral region includes a plurality of annular rings;

a radiomic feature circuit configured to:

extract a first set of radiomic features from the bpMRI image, where the first set of radiomic features includes at least one radiomic feature extracted from the peritumoral region, and at least one, different radiomic feature extracted from the tumoral region;

a PCa classification circuit configured to:

compute a first probability that the patient associated with the ROI has a low-risk of PCa progression, or a high-risk of PCa progression based on the first set of radiomic features;

generate a first classification of the patient as having low-risk of PCa progression, or a high-risk of PCa progression based, at least in part, on the probability; and

a display circuit configured to display the first classification.

15. The apparatus of claim 14 , where the set of radiomic features includes an intratumoral Laws ( 15 ) feature extracted from the T2W image, an intratumoral Canny feature extracted from the T2W image, a peritumoral co-occurrence of local anisotropic gradient orientations (CoLIAGe) entropy feature extracted from the ADC map, an intratumoral Laws ( 11 ) feature extracted from the ADC image, an intratumoral Haralick (entropy) feature extracted from the ADC map, an intratumoral CoLIAGe feature extracted from the ADC map, a peritumoral Haralick (info measure 1 ) feature extracted from the T2W image, a peritumoral Laws ( 17 ) feature extracted from the ADC map, an intratumoral Haralick (info measure 2 ) feature extracted from the T2W image, and an intratumoral Haralick (info measure 2 ) feature extracted from the ADC map.

16. The apparatus of claim 14 ,

where the radiomic feature circuit is further configured to:

extract a second set of radiomic features from the bpMRI image, where the second set of radiomic features includes at least one radiomic feature extracted from the peritumoral region, and at least one, different radiomic feature extracted from the tumoral region;

where the PCa classification circuit is further configured to:

compute a second probability that the patient associated with the ROI has a low-risk of PCa progression, or a high-risk of PCa progression or intermediate-risk of PCa progression based on the second set of radiomic features; and

generate a second classification of the patient as having low-risk of PCa progression, or a high-risk or intermediate-risk of PCa progression based, at least in part, on the second probability; and

where the display circuit is further configured to display the second classification.

17. The apparatus of claim 14 , where the PCa classification circuit is configured as a

quadratic discriminant analysis (QDA) classifier.

18. The apparatus of claim 14 , where the set of circuits further comprises:

a pre-processing circuit configured to preprocess the bpMRI image by:

resampling the image to a pixel size of 0.5×0.5 mm 2 ;

cropping the ROI with 2 mm padding along the x axis and y axis;

interpolating the image to a 3 mm slice thickness;

verifying that the T2W image and ADC map represent the same three dimensional (3D) space; and

correcting scanner variability using a drift correction approach.

19. The apparatus of claim 14 , where the set of circuits further comprises:

a PCa personalized treatment plan circuit configured to:

generate a personalized treatment plan based, at least in part, on the classification;

and where the display circuit is further configured to display the personalized treatment plan.

20. The apparatus of claim 14 , where the set of circuits further comprises:

a training and testing circuit configured to:

train the PCa classification circuit on a multi-institutional cohort; and optionally

test the PCa classification circuit on a multi-institutional testing cohort.

21. The apparatus of claim 16 , where the second set of radiomic features includes an intratumoral Gabor (6 Hz, 2.0 rad) feature extracted from the T2W image, a peritumoral Gabor (6 Hz, 2.8 rad) feature extracted from the T2W image, a peritumoral Haralick (momentum sum) feature extracted from the ADC map, an intratumoral Gabor (6 Hz, 1.8 rad) feature extracted from the ADC map, a mean filter (3 pixel by 3 pixel window) response feature extracted from the peritumoral region of the T2W image, an intratumoral Gabor (2.5 Hz, 0.4 rad) feature extracted from the T2W image, an intratumoral Gabor (3 Hz, 0.4 rad) feature extracted from the T2W image, an intratumoral Gabor (3.5 Hz, 0.4 rad) feature extracted from the T2W image, an intratumoral Gabor (5 Hz, 1.6 rad) feature extracted from the ADC map, and an intratumoral Gabor (6 Hz, 1.6 rad) feature extracted from the ADC map.

22. A non-transitory computer-readable storage device storing computer-executable instructions that when executed cause a processor to perform operations, the operations comprising:

accessing a training dataset of a plurality of bi-parametric magnetic resonance imaging (bpMRI) images of tissue demonstrating prostate cancer (PCa), where each bpMRI image of the plurality of bpMRI images comprises an associated plurality of pixels, each pixel of the associated plurality of pixels having an associated intensity, where each bpMRI image is associated with a patient, and where a PCa progression risk for each patient is known;

determining, for each radiomic feature of N radiomic features extracted from each bpMRI image, N being an integer, an associated value for that radiomic feature for each bpMRI image of the training dataset, where the N radiomic features includes at least one tumoral radiomic feature and at least one peritumoral radiomic feature; and training a quadratic discriminant analysis (QDA) classifier based on the training dataset, the associated values for each radiomic feature of the N radiographic features for each bpMRI image of the training dataset, and the known risk of PCa progression associated with each bpMRI image of the training dataset,

wherein the at least one tumoral feature is associated with a tumoral region, wherein the at least one peritumoral feature is associated with a peritumoral region defined by performing a dilation of a tumoral boundary of the tumoral region, and wherein the peritumoral region includes a plurality of annular rings.

23. The non-transitory computer-readable medium of claim 22 , where the N radiomic features comprise the N most distinguishing radiomic features as determined using one of a Wilcoxon rank-sum test, or a minimum redundancy maximum relevance (mRMR) feature selection technique.