Attention-Driven Feature Extraction for XAI in Histopathology: Leveraging a Hybrid Xception-CBAM Architecture for Multi-Cancer Diagnosis

The precise and automated classification of histopathology images is essential for early detection of cancer, particularly for widespread cancers like Colorectal Cancer (CRC) and Lung Cancer (LC). However, traditional deep learning models frequently encounter challenges due to significant intra-class variability, similarities between different classes, and inconsistent image quality. To overcome these limitations, a detailed multi-layer diagnostic framework is proposed. This method begins with a robust preprocessing pipeline that includes gamma correction, bilateral filtering, and adaptive CLAHE, leading to substantial improvements in quantitative metrics of image quality. A deep learning architecture based on hybrid attention mechanisms has been introduced, which integrates an Xception backbone, a Convolutional Block Attention Module (CBAM), a Transformer block, and an MLP classifier to effectively merge local features with global context. When evaluated on three publicly accessible datasets, the proposed model attained exemplary results, reaching classification accuracies of 99.98% on LC-2500, 99.58% on CRC-VAL-HE-7K, and 99.29% on NCT-CRC-HE-100K. To improve transparency, thorough explain ability analyses are performed utilizing layer-wise feature visualization and Grad-CAM. Lastly, the practical application of this framework is showcased through its implementation on a web-based platform, offering a valuable and user-friendly tool to assist in pathological diagnosis.