US-12626360-B2 - Automated prostate cancer detection and diagnosis using a boosted ensemble of bagging ensemble models

Abstract

A computer system that analyzes medical-imaging data to assess a risk for prostate cancer is described. The computer system may compute features (including intensity, texture and/or spatial features) based at least in part on the medical-imaging data. Then, using a pretrained predictive model, the computer system may determine cancer predictions on a voxel-by-voxel basis, based at least in part on the computed features. Note that the pretrained predictive model may include a boosted parallel random forests (BPRF) model with a boosted ensemble of bagging ensemble models, where a given bagging ensemble model includes an ensemble of random forests models. Next, the computer system may provide feedback based on the cancer predictions for the voxels. For a given voxel, the feedback may include a cancer prediction and a location. In some embodiments, for the given voxel, the feedback may include an aggressiveness of the predicted cancer and/or a recommended therapy.

Inventors

• Jasser Dhaouadi
• Ethan J. Ulrich
• Randall W. Jones

Assignees

• Bot Image, Inc.

Dates

Publication Date

20260512

Application Date

20230224

Claims (20)

1. 1 . A computer system, comprising: an interface circuit; a computation device coupled to the interface circuit; and memory, coupled to the computation device, configured to store program instructions, wherein, when executed by the computation device, the program instructions cause the computer system to perform one or more operations comprising: receiving medical-imaging data associated with a pelvic region of an individual; computing features associated with voxels corresponding to a prostate of the individual based at least in part on the medical-imaging data, wherein, for a given voxel, the features comprise: intensity features, texture features, and a spatial feature corresponding to a distance from a peripheral zone of the prostate to a transition zone of the prostate; determining, on a voxel-by-voxel basis, cancer predictions for the voxels based at least in part on the computed features and a pretrained predictive model, wherein the pretrained predictive model comprises a boosted parallel random forests (BPRF) model, wherein the pretrained BPRF model comprises a boosted ensemble of bagging ensemble models, and wherein a given bagging ensemble model comprises an ensemble of random forests models; and providing feedback based at least in part on the cancer predictions for the voxels, wherein, for the given voxel, the feedback comprises: a cancer prediction and a location.
2. 2 . The computer system of claim 1 , wherein the medical-imaging data comprise magnetic-resonance-imaging studies of the individual; and wherein the magnetic-resonance-imaging studies comprise: transverse relaxation time-weighted (T2 W) images, apparent diffusion coefficient (ADC) images, and diffusion-weighted imaging (DWI) images.
3. 3 . The computer system of claim 2 , wherein the operations comprise registering a first volume corresponding to the ADC images and a second volume corresponding to the DWI images with a third volume corresponding to the T2 W images; and wherein the registration comprises aligning the first volume and the second volume with the third volume, and correcting the second volume for distortion.
4. 4 . The computer system of claim 3 , wherein the registration is based at least in part on mutual information between a give pair of volumes in the first volume, the second volume and the third volume.
5. 5 . The computer system of claim 3 , wherein the registration is based at least in part on a Bayesian technique.
6. 6 . The computer system of claim 1 , wherein the operations comprise segmenting the medical-imaging data to identify the voxels corresponding to the prostate.
7. 7 . The computer system of claim 6 , wherein the segmenting further identifies sub-regions of the prostate.
8. 8 . The computer system of claim 1 , wherein the intensity features comprise radiomics features and the texture features comprise Haralick texture features.
9. 9 . The computer system of claim 1 , wherein the boosted ensemble is based at least in part on an adaptive boosting technique and the bagging ensemble is based at least in part on a Bayesian estimator technique.
10. 10 . The computer system of claim 1 , wherein the boosted ensemble is computed sequentially and the bagging ensemble is computed in parallel.
11. 11 . The computer system of claim 1 , wherein the feedback comprises an image indicating first regions of the prostate where the cancer predictions exceed a first threshold value.
12. 12 . The computer system of claim 11 , wherein the feedback comprises the image indicating second regions of the prostate where the cancer predictions are less than the first threshold value and greater than a second threshold value.
13. 13 . The computer system of claim 1 , wherein the feedback comprises an image with an at least partially transparent three-dimensional (3D) rendering of the prostate and one or more color-coded regions in the 3D rendering corresponding to cancer predictions exceeding a threshold value.
14. 14 . The computer system of claim 1 , wherein, for the given voxel, the feedback indicates an aggressiveness of predicted cancer based at least in part on the cancer predictions.
15. 15 . The computer system of claim 1 , wherein the feedback comprises or corresponds to a recommended therapy based at least in part on the cancer predictions.
16. 16 . A non-transitory computer-readable storage medium for use in conjunction with a computer system, the computer-readable storage medium configured to store program instructions that, when executed by the computer system, causes the computer system to perform one or more operations comprising: receiving medical-imaging data associated with a pelvic region of an individual; computing features associated with voxels corresponding to a prostate of the individual based at least in part on the medical-imaging data, wherein, for a given voxel, the features comprise: intensity features, texture features, and a spatial feature corresponding to a distance from a peripheral zone of the prostate to a transition zone of the prostate; determining, on a voxel-by-voxel basis, cancer predictions for the voxels based at least in part on the computed features and a pretrained predictive model, wherein the pretrained predictive model comprises a boosted parallel random forests (BPRF) model, wherein the pretrained BPRF model comprises a boosted ensemble of bagging ensemble models, and wherein a given bagging ensemble model comprises an ensemble of random forests models; and providing feedback based at least in part on the cancer predictions for the voxels, wherein, for the given voxel, the feedback comprises: a cancer prediction and a location.
17. 17 . The non-transitory computer-readable storage medium of claim 16 , wherein the boosted ensemble is based at least in part on an adaptive boosting technique and the bagging ensemble is based at least in part on a Bayesian estimator technique.
18. 18 . A method for providing feedback, comprising: by a computer system: receiving medical-imaging data associated with a pelvic region of an individual; computing features associated with voxels corresponding to a prostate of the individual based at least in part on the medical-imaging data, wherein, for a given voxel, the features comprise: intensity features, texture features, and a spatial feature corresponding to a distance from a peripheral zone of the prostate to a transition zone of the prostate; determining, on a voxel-by-voxel basis, cancer predictions for the voxels based at least in part on the computed features and a pretrained predictive model, wherein the pretrained predictive model comprises a boosted parallel random forests (BPRF) model, wherein the pretrained BPRF model comprises a boosted ensemble of bagging ensemble models, and wherein a given bagging ensemble model comprises an ensemble of random forests models; and providing the feedback based at least in part on the cancer predictions for the voxels, wherein, for the given voxel, the feedback comprises: a cancer prediction and a location.
19. 19 . The method of claim 18 , wherein the boosted ensemble is based at least in part on an adaptive boosting technique and the bagging ensemble is based at least in part on a Bayesian estimator technique.
20. 20 . The method of claim 18 , wherein the boosted ensemble is computed sequentially and the bagging ensemble is computed in parallel.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of U.S. Provisional Application No. 63/313,604, filed Feb. 24, 2022, the entirety of which is incorporated herein by reference. FIELD The described embodiments relate to processing of medical images. Notably, the described embodiments relate to a processing of medical images to detect and/or diagnose a type of cancer (such as prostate cancer) using a pretrained predictive model that includes a boosted ensemble of bagging ensemble models. BACKGROUND Prostate cancer is the second most-common type of cancer among men. If detected and diagnosed early in the disease progression, prostate cancer can typically be treated successfully. Screening tests are often performed in an attempt to find early-stage prostate cancer in individuals before they have symptoms. For example, prostate-specific antigen (PSA) blood tests are used to look for changes in the PSA level, because the probability of having prostate cancer increases with the PSA level. However, there is no set cutoff point that indicates whether or not a man has prostate cancer. Consequently, it remains unclear whether the benefits of prostate cancer screening outweigh the risks for most men. When prostate cancer is suspected, e.g., based on the results of a screening test or patient systems, additional tests are usually used to detect and diagnose prostate cancer. For example, a urologist may perform a prostate biopsy on a patient. However, a prostate biopsy is an invasive procedure in which a thin, hollow needle is repeatedly inserted into the prostate to remove small cylindrical cores of prostate tissue for assessment by a pathologist. Alternatively or additionally, medical imaging tests are often used to detected and diagnose prostate cancer. For example, different types of magnetic-resonance-imaging (MRI) scans or studies (such as diffusion weighted imaging or DWI, dynamic contrast enhanced or DCE MRI, and/or MR spectroscopy or MRS) are routinely used to acquire images of the prostate. In principle, MRI studies can be analyzed to: determine parameters of prostate tissue, identify abnormal areas, diagnose prostate cancer, and/or assess the extent (or stage) of the cancer. In practice, accurate interpretation of medical images (such as MRI images) and prediction of prostate cancer remain challenging. This is the case whether the interpretations and predictions are performed solely by physicians (such as radiologists), jointly by physicians and computers (computer-aided diagnosis), or in an automated manner (solely by computer). For example, when medical images are interpreted by radiologists, there is often high inter-reader variability and a high false positive rate. Consequently, the limitations of existing analysis techniques adversely impact analysis performance and increase the costs of treating suspected or actual prostate cancer, which in turn increase the morbidity and mortality associated with this disease. SUMMARY A computer system that provides feedback is described. This computer system includes: an interface circuit; a computation device (such as a processor, a graphics processing unit or GPU, etc.) that executes program instructions; and memory that stores the program instructions. During operation, the computer system receives medical-imaging data associated with a pelvic region of an individual. Then, the computer system computes features associated with voxels corresponding to a prostate of the individual based at least in part on the medical-imaging data, where, for a given voxel, the features include: intensity features, texture features, and a spatial feature corresponding to a distance from a peripheral zone of the prostate to a transition zone of the prostate. Moreover, the computer system determines, on a voxel-by-voxel basis, cancer predictions for the voxels based at least in part on the computed features and a pretrained predictive model. This pretrained predictive model includes a boosted parallel random forests (BPRF) model, and the pretrained BPRF model includes a boosted ensemble of bagging ensemble models (such as classifiers), where a given bagging ensemble model includes an ensemble of random forests models. Next, the computer system provides the feedback based at least in part on the cancer predictions for the voxels, where, for the given voxel, the feedback includes: a cancer prediction and a location. Note that the medical-imaging data may include MRI studies of the individual. For example, the MRI studies may include: transverse relaxation time-weighted (T2 W) images, apparent diffusion coefficient (ADC) images, and/or diffusion-weighted imaging (DWI) images. Moreover, the computer system may segment the medical-imaging data to identify the voxels corresponding to the prostate. In some embodiments, the segmenting may identify sub-regions of the prostate. Furthermore, the computer system may register a first volume corresponding to the ADC images