Article

Deep-learning algorithm for prostate cancer detection demonstrates early potential

Author(s):

The algorithm was applied to 50 patients who underwent radical prostatectomy between 2008 and 2018.

Researchers at the 2020 Society of Urologic Oncology Annual Meeting shared initial data for their novel deep-learning algorithm intended to facilitate the detection and grading of clinically significant prostate cancer.1

Their findings showed that the algorithm appeared to accurately identify cancer on whole mount prostate pathology, potentially alerting pathologists to suspicious areas of cancer burden prior to clinical assessment.

“Accurate Gleason Grade grouping determination at time of prostate cancer diagnosis is vital for patients’ decision making to pursue either surveillance, definite treatment, or adjuvant therapy. However, intercarrier variability in the grading of pathologic specimens remains a significant issue,” Nitin K. Yerram, MD, a urologic oncology fellow at the National Cancer Institute, said during a virtual presentation.

“Given that deep learning algorithms are being developed to help augment and standardize detection and grading for prostate cancer, our group has utilized AI algorithms to provide a solution to similar problems in bladder cancer pathology and prostate cancer imaging,” he added.

The researchers used previous studies to create a novel algorithm that was developed to help identify and grade prostate cancer on whole mount pathology slides.

The algorithm was previously trained for detection and grading of prostate cancer using publicly available datasets, totaling over 9000 biopsies, tissue microarrays, and surgical sections. Additionally, a small number of segmented whole mount prostatectomy slides were utilized for detection task only.

“The algorithm was designed to be agnostic of tissue source operating on patches abstracted from each image,” Yerram explained. “Each patch is representing 100 microns by 100 microns, or 200 by 200 pixels at 20x. Each patch is then run through our algorithm. These patches are spatially recreated to produce a probability map where each patch is assigned either a value of zero or 1, 1 being the highest likelihood of cancer.”

Burden of each foci within each slide were marked by ink under microscope and mapped digitally for quantitative comparison.

Congruent lesions >2 mm2 area were considered positive for deep learning-based detection.

“At the NCI, we sought to evaluate the congruence of the AI algorithm with pathologist annotated regions from routine clinical practice,” Yerram said, adding that the researchers then chose to convert the probability map into a detection mask, “which identifies high risk areas defined as congruent areas of high probability in a set pixel size.”

For statistical analysis, they then compared the detection rate to pathologist annotation. A true positive was called as any detection within a prospect annotation. Once the detected areas of probability that are considered cancer were found, they ran it through the grading algorithm. Each patient is given a prediction: either Gleason Grade 3, 4, or 5. Similarly, with the detection map, the researchers were able to visually look at probability maps for each one of these Gleason scores.

The algorithm was applied to 50 patients who underwent radical prostatectomy between 2008 and 2018 with available digitized whole-mount pathology and foci-level annotations of disease burden. In total, 24 patients had Gleason Grade 1 or 2 and 26 had Gleason Grade 3, 4, or 5 on surgical pathology.

Patient level detection accuracy of cancer or no cancer was 96%, with 2 patients classified as negative for cancer. However, these 2 patients had Grade 1 and 2 disease, indicating that no high-grade disease was missed, Yerram noted.

On annotated foci identification, the algorithm showed a 77.6% sensitivity, with a median rate of 1 false-positive per patient (range, 0-20). Positive predictive value was 38.4%.

“Our program performed quite well in the grading of cancer. Overall, a higher percent of the tumor grade and burden was detected in patients with Grade 3-5 disease versus patients with Grade 1-2 disease,” Yerram concluded.

“Our algorithm demonstrated excellent accuracy in identifying cancer on whole mount pathology,” he added. “Now, there is room for improvement in the first level accuracy, and our positive predictive value needs to be improved, but that will continue to be improved upon with further refinement. Systems such as this, we envision, can be used to alert pathologists to suspicious areas of cancer burden prior to clinical assessment. And lastly, spatial assessment of grade distribution can be achieved with further refinement of our algorithm.”

Reference

1. Yerram N, Harmon S, O'Connor L, et al. Application of Deep Learning Detection and Grading System for Identification of Clinically Significant Prostate Cancer on Whole Mount Pathology. Presented at: Society of Urologic Oncology 21st Annual Meeting; December 3, 2020. Abstract 3.

Related Videos
Chad Tang, MD: Considerations for SBRT in metastatic RCC
Interpreting ART toxicity and tolerability for bladder cancer, with Vedang Murthy, MD
Alexander Pastuszak, MD, PhD: Is hormone therapy safe after prostate cancer radiotherapy?
Refining prostate cancer therapy strategy to address RAPTOR findings
Considering patient-reported outcomes in kidney cancer care, with Nicholas Zaorsky, MD, PhD
Soumyajit Roy, MS, MBBS: The effect of prostate cancer patient history in RAPTOR
 Nicholas Zaorsky, MD, MS: Protecting kidney function after local renal cell carcinoma therapy
Nicholas van AS, MD, MBBCH: The case for SBRT as a standard of care for localized prostate cancer
Pierre Blanchard, MD, PhD: What can hydrogel space provide to optimal prostate cancer care?
Related Content
© 2024 MJH Life Sciences

All rights reserved.