Deep learning for automatic Gleason pattern classification for grade group determination of prostate biopsies.

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Histopathologic grading of prostate cancer via Gleason patterns (GPs) exhibits significant inter-observer variability.
• This variability can lead to suboptimal patient treatment strategies.
• Digitization of prostate biopsies and whole-slide imaging enable computer-aided grading solutions.

Purpose of the Study:

• To develop and evaluate an automated system for detecting Gleason patterns (GPs) and determining prostate cancer Grade Groups (GG).
• To assess the accuracy and concordance of a convolutional neural network (CNN) based approach for prostate cancer grading.

Main Methods:

• A convolutional neural network (CNN), specifically Inception-v3, was re-trained for automated detection of GP 3 and GP ≥ 4 in digitized prostate biopsies.
• Pixel-level annotations were performed on 96 prostate biopsies from 38 patients.
• The CNN's output was converted into probability maps to determine the GG for the entire biopsy.

Main Results:

• The CNN achieved 92% accuracy in differentiating non-atypical from malignant (GP ≥ 3) areas, with 90% sensitivity and 93% specificity.
• Differentiation between GP ≥ 4 and GP ≤ 3 areas yielded 90% accuracy, 77% sensitivity, and 94% specificity.
• The automated GG determination showed substantial agreement with a genitourinary pathologist (65% concordance, κ = 0.70).

Conclusions:

• A CNN-based approach enables accurate differentiation between non-atypical and malignant areas in prostate biopsies based on GPs.
• The automated system demonstrates substantial agreement with expert pathologist grading, offering a promising tool for improving prostate cancer diagnosis and treatment selection.