A filter-level explainability framework for CNNs in histopathology image analysis

Convolutional neural networks (CNNs) have achieved remarkable accuracy in histopathology image classification, yet their decision logic remains largely opaque. Most explainability methods, such as Grad-CAM or SHAP, provide only coarse heatmaps, offering limited insight into the role of individual filters. This study introduces a filter-level explainability framework that quantitatively aligns single-filter activations with Grad-CAM outputs using similarity metrics (Pearson correlation, Dice coefficient, mean squared error), integrates progressive ablation experiments to assess filter contributions, and applies region-based scoring to capture intensity, frequency, and spatial distribution. Evaluations are conducted on the LC25000 dataset, a widely used benchmark comprising 25,000 histopathology image patches from lung and colon tissues (benign and malignant classes). The results demonstrate strong classification performance (Accuracy 0.964, F1-score 0.965, ROC-AUC 0.997) and high filter-to-Grad-CAM agreement (≈0.97 Pearson, ≈0.95 Dice), supported by statistical significance testing. In contrast to prior studies that mainly rely on qualitative visualization, the proposed framework delivers a systematic, statistically validated, and multidimensional approach to filter-level interpretability in CNN-based histopathology analysis. While demonstrated on the LC25000 dataset, the methodological contribution is designed in principle to be dataset-agnostic, as it operates on the final convolutional layer irrespective of dataset choice. By moving beyond qualitative heatmaps, the framework offers a rigorous and reproducible pathway to better understanding CNN decisions, with potential to enhance trust, transparency, and clinical adoption of AI in pathology. In addition, we outline key considerations for translating the framework into routine pathology practice.