Pathologist-like explainable AI for interpretable Gleason grading in prostate cancer.

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Computational pathology

Background:

• Prostate cancer aggressiveness is determined by the Gleason scoring system from histopathology.
• Current AI models for Gleason scoring lack explainability, hindering clinical adoption.
• High interobserver variability exists in Gleason scoring among pathologists.

Purpose of the Study:

• To develop an inherently explainable AI model for Gleason pattern segmentation in prostate cancer.
• To improve the robustness and interpretability of AI-driven histopathological analysis.
• To address the challenge of interobserver variability in medical image segmentation.

Main Methods:

• Trained an AI model on 1,015 prostate tissue microarray core images.
• Utilized detailed pattern descriptions annotated by 54 international pathologists.
• Employed pathologist-defined terminology and soft labels to manage data uncertainty.

Main Results:

• Achieved robust Gleason pattern segmentation comparable to direct segmentation methods (Dice score: 0.713 ± 0.003).
• The AI model provided interpretable outputs, enhancing clinical trust.
• Demonstrated superior performance in segmentation accuracy compared to conventional methods.

Conclusions:

• An inherently explainable AI approach can effectively segment Gleason patterns despite high interobserver variability.
• This method offers a promising alternative to conventional AI, enhancing clinical acceptance.
• The released dataset will foster research in subjective medical image segmentation and pathologist reasoning.