WaveAttn-CRC: Wavelet-Enhanced Attention for Multi-Magnification Colorectal Histopathology Classification with Explainable AI

The following are the contributions of this work:

1. We approach the colorectal histopathology classification problem as a multi-magnification task, where the model needs to identify five tissue states at four different magnifications (40x, 100x, 200x, and 400x) and make predictions for both the patches and the source image.

2. We introduce WaveAttn-CRC, a compact dual-stream architecture integrating FastViT spatial features and Haar wavelet details in orientation-specific ways. Multi-head attention is used to refine the fused representation and FiLM is used to condition it on magnification. This design enables the model to incorporate three features: tissue structure, high-frequency detail and optical scale in a single 12.47 million parameter network.

3. It is a controlled comparison of one multi-magnification model and four specialist models with different magnifications. The joint model achieves an image-level accuracy of 0.940 and a macro-F1 of 0.907 overall. It is also ahead of the 200x specialist by 0.042 macro-F1 and the 400x specialist by 0.011 macro-F1, and demonstrates classification problems at 40x.

4. We introduce an explainability assessment, which integrates Grad-CAM++, Score-CAM, LIME and SHAP, along with a quantitative deletion-insertion analysis. This is a way to relate visual explanations to a quantifiable faithfulness test, rather than just qualitative heatmaps.