Dual selective gleason pattern-aware multiple instance learning with uncertainty regularization for grade group...

Area of Science:

• Digital pathology
• Computational oncology
• Machine learning in medicine

Background:

• Accurate Gleason Grade Group (GG) prediction is crucial for prostate cancer management.
• Existing multiple instance learning (MIL) methods for Gleason grading often neglect the combined influence of different Gleason Patterns, impacting accuracy and interpretability.

Purpose of the Study:

• To develop an uncertainty-driven, dual-selective, Gleason Pattern-aware MIL model (DSPA-U-MIL) for improved patient-level GG prediction.
• To enhance the accuracy and interpretability of Gleason grading by incorporating domain knowledge of Gleason Patterns.

Main Methods:

• Proposed DSPA-U-MIL model integrating learnable pattern aggregation tokens and expert concept-guided patch-level aggregation.
• Employed a teacher-student knowledge distillation framework for cooperative prediction across Gleason Patterns.
• Introduced an uncertainty constraint to improve robustness against noisy labels.

Main Results:

• DSPA-U-MIL demonstrated superior performance compared to existing MIL methods across five datasets (10,809 WSIs, 1133 TMAs).
• Achieved up to a 6.7% improvement in quadratic weighted kappa (QWK) score over the CLAM baseline, particularly on the TCGA-PRAD dataset.
• Student model's gating weight distributions aligned with pathologist annotations, confirming interpretability.

Conclusions:

• DSPA-U-MIL offers a significant advancement in automated Gleason GG prediction, enhancing both predictive accuracy and clinical interpretability.
• The model's ability to leverage domain knowledge of Gleason Patterns and handle label noise makes it a valuable tool for prostate cancer risk stratification.