Self-Learning for Weakly Supervised Gleason Grading of Local Patterns.

Area of Science:

• Pathology
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Prostate cancer diagnosis relies on manual Gleason grading of histology slides, a time-consuming and subjective process.
• Computer-aided diagnosis (CAD) systems offer potential but often require tedious pixel-level annotations.
• Existing CAD methods struggle with explainability, lacking precise tumor region identification.

Purpose of the Study:

• To develop a weakly-supervised deep learning model for prostate cancer grading.
• To eliminate the need for pixel-level annotations by leveraging global Gleason scores.
• To improve both patch-level grading and biopsy-level scoring with enhanced explainability.

Main Methods:

• A novel weakly-supervised deep learning model based on self-learning Convolutional Neural Networks (CNNs) was developed.
• The model was trained using only global Gleason scores from gigapixel whole slide images.
• Performance was evaluated on multiple external datasets for patch-level Gleason grading and global Grade Group prediction.

Main Results:

• The proposed weakly-supervised model significantly outperformed its fully-supervised counterpart in patch-level Gleason grading.
• It achieved an average improvement of nearly 18% in Cohen's quadratic kappa (κ) score for patch-level grading.
• The model demonstrated superior performance compared to state-of-the-art methods in global biopsy-level scoring.

Conclusions:

• Weakly-supervised learning, using global Gleason scores, leads to more robust and higher-performing models by reducing annotator bias.
• The model's ability to generalize and provide better feature representations enhances its reliability for pathologists.
• This approach offers a promising solution for efficient and accurate prostate cancer grading, addressing limitations of current CAD systems.