Explainable nucleus classification using Decision Tree Approximation of Learned Embeddings.

Area of Science:

• Computational pathology
• Digital pathology
• Histopathology image analysis

Background:

• Accurate nucleus detection, segmentation, and classification are crucial for analyzing the tumor microenvironment in whole-slide histopathology images.
• Deep learning models offer high performance but often lack explainability, hindering clinical trust and adoption.
• Existing explainability methods (e.g., Grad-CAM, LIME) are not ideal for nucleus classification, providing indirect or nonintuitive explanations.

Purpose of the Study:

• To present scalable techniques for nuclear detection, segmentation, and explainable classification.
• To improve upon existing deep learning architectures for histopathology image analysis.
• To develop an explainability method that provides clear, quantitative insights for pathologists.

Main Methods:

• Modified the Mask R-CNN architecture, decoupling detection and classification tasks for improved accuracy.
• Enabled learning from hybrid annotation datasets (NuCLS) with mixed bounding boxes and segmentation boundaries.
• Introduced Decision Tree Approximation of Learned Embeddings (DTALE) for global and individual nucleus classification explanations.

Main Results:

• Achieved improved accuracy in nucleus detection, segmentation, and classification through architectural modifications.
• DTALE provides simple, quantitative, and pathologist-interpretable explanations based on morphological features.
• The developed methods maintain model accuracy while offering enhanced explainability.

Conclusions:

• The presented techniques facilitate scalable and explainable nucleus analysis in histopathology.
• DTALE offers a novel, intuitive approach to model explainability for computational pathology.
• These advancements support the integration of computational pathology in computer-aided diagnosis and biomarker discovery.