An interpretable hierarchical deep learning framework for lung and colon cancer detection

Accurate and interpretable classification of lung and colon cancer from histopathological images remains challenging because of substantial inter-class similarities, morphological heterogeneity, and the limited transparency of conventional deep learning (DL) models. Most existing methods treat the problem as a flat multiclass classification task, requiring a single model to distinguish normal tissue and multiple cancer subtypes simultaneously. In this study, we propose an interpretable hierarchical DL framework that decomposes the five-class classification problem into clinically motivated coarse-to-fine decision stages. The framework initially separates normal tissue from cancerous tissue and then directs each sample to a specialized classifier for either normal tissue categorization or cancer subtype identification. Candidate DL architectures are independently evaluated at each stage, allowing the most suitable model to be selected for each specific classification task. The framework is evaluated on the LC25000 benchmark and further assessed on the independent LungHist700 dataset, which presents a distinct evaluation setting with different image characteristics and class distributions. The proposed approach achieves an accuracy of 98.50% on LC25000 and 82.01% on LungHist700. To improve model transparency, Grad-CAM, Grad-CAM++, LIME, and Occlusion Sensitivity are employed to highlight the image regions that contribute to the hierarchical predictions. The results show that the framework combines strong classification performance with a structured coarse-to-fine decision process and complementary interpretability techniques. Overall, this study demonstrates the potential of hierarchical decision decomposition and explainable DL to provide a more transparent and task-specific approach to computer-aided histopathological cancer classification, particularly for distinguishing normal tissue from lung and colon cancer subtypes.