A systematic comparison of deep learning methods for Gleason grading and scoring.

Area of Science:

• Computational pathology
• Digital histopathology image analysis
• Artificial intelligence in oncology

Background:

• Prostate cancer diagnosis relies on Gleason score, assessing cellular abnormality in tissue samples.
• Computational pathology offers advanced algorithms for analyzing digitized histopathology images.
• Lack of consensus exists on optimal deep learning methods for Gleason grading and scoring tasks.

Purpose of the Study:

• To systematically compare state-of-the-art deep neural network training approaches for Gleason grading and scoring.
• To evaluate performance across nine diverse datasets from pathology institutes and repositories.
• To provide guidance on selecting the best methods based on task and data characteristics.

Main Methods:

• Systematic comparison of nine datasets using various deep neural network training approaches.
• Evaluation of fully-supervised learning, weakly-supervised learning, semi-supervised learning, Additive-MIL, Attention-Based MIL, Dual-Stream MIL, TransMIL, and CLAM.
• Application of these methods to Gleason grading and scoring tasks on digitized histopathology images.

Main Results:

• Fully supervised learning demonstrated superior performance for Gleason grading tasks.
• CLAM (Contextualized Learning Attention Model) proved most effective for Gleason scoring tasks.
• Performance varied across methods depending on dataset characteristics and label availability.

Conclusions:

• The study identifies optimal deep learning strategies for prostate cancer Gleason assessment.
• Fully supervised learning is recommended for Gleason grading, while CLAM is preferred for Gleason scoring.
• Findings guide researchers in choosing appropriate computational pathology methods for specific diagnostic challenges.