Tumor Cellularity Assessment of Breast Histopathological Slides via Instance Segmentation and Pathomic Features...

Area of Science:

• Digital pathology
• Bioimage analysis
• Artificial intelligence in medicine

Background:

• Accurate cell nuclei segmentation and classification are crucial for bioimage analysis.
• Deep learning (DL) models excel in nuclei detection but lack interpretability, hindering clinical use.
• Pathomic features offer a more interpretable alternative for classifier decision-making.

Purpose of the Study:

• To develop an explainable computer-aided diagnosis (CAD) system for evaluating tumor cellularity in breast histopathology.
• To compare an end-to-end DL approach with a two-step pipeline using morphological and textural nuclei features.
• To enhance the interpretability of machine learning models for pathologist trust and clinical adoption.

Main Methods:

• Implemented an end-to-end DL approach using Mask R-CNN for instance segmentation.
• Developed a two-step pipeline extracting morphological and textural features for Support Vector Machine and Artificial Neural Network classifiers.
• Utilized SHAP (Shapley additive explanations) for feature importance analysis and an expert pathologist for validation.

Main Results:

• The two-stage pipeline, while slightly less accurate than end-to-end DL, provided clearer feature interpretability.
• SHAP analysis elucidated features used by machine learning models, aiding understanding of their decisions.
• The feature set was validated by an expert pathologist, confirming clinical usability.

Conclusions:

• Explainable AI approaches, particularly feature-based methods, can improve pathologist trust in AI-driven CAD systems.
• Interpretable features derived from morphological and textural characteristics are valuable for clinical decision support in digital pathology.
• The developed system, validated on an external dataset, supports tumor cellularity quantification and aids AI integration into clinical workflows.