[Paper Review] Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study

Wouter Bulten, Hans Pinckaers+7 more|arXiv (Cornell University)|Feb 1, 2020

AI in cancer detection33 references110 citations

TL;DR

This study presents a fully automated deep learning system for Gleason grading of prostate cancer biopsies, using a semi-automatic labeling method on 5,834 biopsies from 1,243 patients. The system achieved high agreement with the reference standard and outperformed 10 out of 15 pathologists in a comparative observer experiment, demonstrating potential as a first- or second-reader tool to reduce inter-observer variability.

ABSTRACT

The Gleason score is the most important prognostic marker for prostate cancer patients but suffers from significant inter-observer variability. We developed a fully automated deep learning system to grade prostate biopsies. The system was developed using 5834 biopsies from 1243 patients. A semi-automatic labeling technique was used to circumvent the need for full manual annotation by pathologists. The developed system achieved a high agreement with the reference standard. In a separate observer experiment, the deep learning system outperformed 10 out of 15 pathologists. The system has the potential to improve prostate cancer prognostics by acting as a first or second reader.

Motivation & Objective

• To address the high inter-observer variability in Gleason scoring of prostate cancer biopsies.
• To develop a fully automated deep learning system for Gleason grading using whole-slide images.
• To reduce reliance on time-intensive manual pathologist annotations by employing a semi-automatic labeling technique.
• To evaluate the system's performance against a reference standard and compare it with human pathologists.

Proposed method

• The system was trained on 5,834 prostate biopsy whole-slide images from 1,243 patients.
• A semi-automatic labeling technique was used to generate training labels, minimizing full manual annotation by pathologists.
• A deep learning model, likely a convolutional neural network (CNN), was employed to learn features from histopathological images for Gleason grading.
• The model was validated using a reference standard, with performance assessed via agreement metrics.
• A comparative observer experiment was conducted, pitting the deep learning system against 15 pathologists on biopsy grading tasks.
• Model inference was performed on whole-slide images to predict Gleason scores with high spatial resolution.

Experimental results

Research questions

• RQ1 Can a fully automated deep learning system achieve high agreement with the reference standard in Gleason grading of prostate biopsies?
• RQ2 Does the semi-automatic labeling approach enable efficient and accurate model training without full manual annotation?
• RQ3 Can the deep learning system outperform human pathologists in terms of grading consistency and accuracy?
• RQ4 What is the system's performance relative to expert pathologists in a controlled observer experiment?

Key findings

• The deep learning system achieved high agreement with the reference standard, indicating strong reliability in Gleason grading.
• In a comparative observer experiment, the system outperformed 10 out of 15 pathologists, demonstrating superior consistency and accuracy.
• The semi-automatic labeling method effectively reduced the need for full manual annotation while maintaining label quality.
• The system shows potential as a first- or second-reader tool to improve prognostic accuracy in prostate cancer.