Enhancing deep learning interpretability for hand-crafted feature-guided histologic image classification via...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Image analysis

Background:

• Deep learning (DL) models excel in histologic image analysis but lack interpretability.
• Hand-crafted features (HCFs) offer interpretability but have lower predictive power.
• The relationship between DL and HCFs is underexplored, hindering clinical adoption.

Purpose of the Study:

• To enhance DL model interpretability and performance in histologic image analysis.
• To integrate HCFs into DL models using a weak-to-strong generalization (WSG) framework.
• To explore the correlation between DL and HCFs for better clinical adoption.

Main Methods:

• Developed a WSG framework with an interpretable HCF-based 'weak' teacher model supervising a 'strong' DL student model.
• Designed an adaptive bootstrap WSG loss function for optimizing knowledge transfer from HCFs to DL features.
• Analyzed mutual information (MI) between HCFs and DL features to assess interpretability and correlations.

Main Results:

• WSG framework consistently improved classification performance across various models.
• Saliency-map analysis showed WSG supervision improved model focus on diagnostically relevant regions.
• Quantitative analysis revealed increased MI between HCFs and DL features post-WSG training.

Conclusions:

• The WSG framework effectively integrates HCFs into DL training, enhancing interpretability and predictive performance.
• Key HCFs driving DL predictions in histologic image classification were elucidated.
• Findings support broader clinical adoption of interpretable DL models in pathology.