Interpretable weakly-supervised learning through kernel density matrices: A digital pathology use case.

Area of Science:

• Computational pathology
• Machine learning
• Deep learning

Background:

• Deep learning classification faces challenges in uncertainty quantification and interpretability with fully-supervised vs. weakly-supervised methods.
• A unified framework for both supervision modes with quantifiable interpretation metrics is needed.

Purpose of the Study:

• To introduce WiSDoM (Weakly-Supervised Density Matrices), a novel framework unifying fully-supervised and weakly-supervised classification.
• To enable quantifiable interpretation metrics and uncertainty quantification within a single model.

Main Methods:

• Utilized kernel matrices to model probability distributions of data and labels.
• Integrated differentiable kernel density matrices for optimization, local-global attention for feature weighting, and prototype generation via kernel space sampling.
• Employed ordinal regression through density matrix operations for classification.

Main Results:

• Achieved high performance in supervised patch classification (AUC = 0.896) and weakly-supervised whole-slide classification (AUC = 0.930) on histopathology images.
• Generated posterior probability distributions, variance-based uncertainty maps, and interpretable phenotype prototypes.
• Demonstrated consistent performance (AUC > 0.89) and high expert agreement (0.88) in Gleason grading tasks across supervision modes.

Conclusions:

• Kernel density matrices provide a robust foundation for classification models requiring both interpretability and uncertainty quantification.
• WiSDoM successfully unifies different supervision modes, offering enhanced insights in computational pathology.
• The framework's outputs, including uncertainty maps and prototypes, improve model transparency and clinical utility.