Robust, credible, and interpretable AI-based histopathological prostate cancer grading.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Histopathology image analysis

Background:

• Prostate cancer (PCa) diagnosis relies on expert histopathological evaluation.
• Current AI grading tools show reduced performance on external datasets due to variations in sample preparation (staining, thickness, scanner).
• AI grading predicting ISUP grades perpetuates human annotation errors.

Purpose of the Study:

• To develop an AI-based prostate cancer grading framework (PCAI) robust to data variations and trained on patient outcomes.
• To enhance AI model robustness, interpretability, and prediction confidence for clinical application.

Main Methods:

• Developed PCAI, an AI framework trained on objective patient outcomes, not subjective ISUP grades.
• Generated a multicentric dataset (25,591 patients, 83,864 images) including high-variance data to optimize robustness.
• Evaluated PCAI on three external test cohorts (2,255 patients, 9,437 images).

Main Results:

• Sample processing variations (slide thickness, staining time) reduced AI grading performance by up to 6.2% (C-index).
• Algorithmic improvements (e.g., domain adversarial training) enhanced PCAI's robustness, interpretability, and credibility.
• PCAI systematically exceeded expert ISUP grading in C-index and AUROC by up to 22 percentage points.

Conclusions:

• Data variations pose significant risks to AI-based histopathological PCa grading.
• Algorithmic improvements and outcome-based training create robust AI models with superior PCa grading performance.