SUO 2020: Application of Deep Learning Detection and Grading System for Identification of Clinically Significant Prostate Cancer on Whole Mount Pathology
Treatment decision making for patients newly diagnosed with prostate cancer is in large part predicated on histologic grade. Thus, timely and accurate ascertainment of tumor grade is important to inform the decision to pursue initial active surveillance, upfront active treatment, and adjuvant therapy following surgery. However, there is a significance between reader variability. Thus, a standardized approach would be useful. The authors sought to assess a novel deep learning algorithm in the detection and grading of prostate cancer based on whole-mount pathology slides.
The authors developed a deep learning algorithm using publicly available data sources of prostate biopsies, tissue microarrays, and surgical sections. This algorithm used an ensemble of ResNet architectures for cancer detection and grading using image patches measuring 100x100µm at 20x.
The authors used a number of publicly available data sets for training and validation, demonstrating, for patch-level validation, the algorithm had 92% detection accuracy and 78% Gleason grade classification accuracy.
The authors then applied this system to patients with digitized whole-mount pathology and with foci-level annotations of disease burden, among 50 patients treated with radical prostatectomy at their institution. Among these patients, 24 had Gleason grade group 1 or 2 disease and 26 of whom had Gleason grade group 3, 4, or 5 based on surgical pathology assessment. The algorithm was compared to pathologist annotation performed in the course of routine clinical care.
The patient-level cancer detection accuracy rate was 96%, with the algorithm identifying two false-negative patients. At a foci-level (of which there were 85 among 50 patients), 66 foci were correctly identified (77.6% sensitivity), though 115 false positives were identified. There were 19 false-negative foci, the vast majority of which (16) were for patients with Gleason grade group 1 or 2 disease. Thus, at a foci-level, the positive predictive value was 36.5%. The authors demonstrated that false-positive lesions were more likely to be small, suggesting that these may in fact represent foci of disease that were just not annotated by the pathologists.
The authors then applied their algorithm for grading of the identified lesions.
Using the algorithm, grade was assigned on a per-patient basis. They noted that high-grade histology (Gleason grade group 3-5) was more commonly noted in the algorithm.
The authors conclude that this approach demonstrates excellent accuracy in prostate cancer diagnosis, though grading and foci-level assessment remains limited with improvements in positive predictive value required.
Presented by: Nitin K. Yerram, MD, Urologic Oncology Fellow, National Cancer Institute, Bethesda, MD
Written by: Christopher J.D. Wallis, MD, Ph.D., Instructor in Urology, Vanderbilt University Medical Center, Nashville, Tennessee @WallisCJD on Twitter at the 2020 Society of Urologic Oncology Annual Meeting – December 2-5, 2020 – Washington, DC