Entropy-Regularized Attention for Explainable Histological Classification with Convolutional and Hybrid Models.

Area of Science:

• Histopathology
• Medical Imaging
• Computer Vision

Background:

• Deep learning models like CNNs and ViTs excel at histological image classification.
• However, these models often lack interpretability, hindering clinical trust and adoption.
• Understanding model decisions is crucial for reliable diagnostic tools.

Purpose of the Study:

• To develop a unified framework to enhance the interpretability of deep learning models in histological image classification.
• To improve the quality of visualization heatmaps generated by methods like Grad-CAM.
• To quantitatively evaluate the impact of the proposed framework on explainability and classification performance.

Main Methods:

• Introduced a unified framework incorporating an attention branch and CAM Fostering (an entropy-based regularizer).
• Trained six backbone architectures (ResNet-50, DenseNet-201, EfficientNet-b0, ResNeXt-50, ConvNeXt, CoatNet-small) on five H&E-stained datasets.
• Evaluated explanation quality using metrics like coherence, complexity, confidence drop, and their harmonic mean (ADCC).

Main Results:

• The proposed method increased ADCC in five of six backbones, with ResNet-50 showing the largest gain (+15.65%).
• CoatNet-small achieved the highest overall ADCC score (+2.69%), reaching 77.90% on a non-Hodgkin lymphoma dataset.
• Classification accuracy remained stable or improved in four of the tested models.

Conclusions:

• Combining attention mechanisms and entropy-based regularization produces clearer and more informative heatmaps.
• The framework enhances model interpretability without degrading classification performance.
• The contributions include a modular architecture for diverse models and a quantitative evaluation suite for explainability.