Towards Efficient and Explainable Colorectal Histopathology Classification in IoT-Enabled eHealth via Compact Attention Fusion

Colorectal cancer (CRC) remains a leading cause of cancer-related mortality. This creates a strong need for accurate and efficient computer-aided diagnosis in Internet of Things (IoT)-enabled eHealth workflows, such as digital pathology and edge-assisted screening. Deep learning has improved histopathology image classification. However, many high-performing models are computationally demanding, and compact alternatives may miss fine-grained tissue patterns. We propose a compact attention-fusion framework for nine-class CRC histopathology patch classification that balances accuracy and efficiency for deployment on resource-constrained edge gateways. The model extends a DenseNet121 backbone by integrating window self-attention (WSA) and the convolutional block attention module (CBAM). Their outputs are fused adaptively. Separable convolutions then refine features to capture local morphology and global context while reducing redundancy. On the NCT-CRC-HE-100K dataset, the proposed method achieves 99.68% accuracy, 99.70% precision, 99.60% recall, and 99.65% F1-score, outperforming six CNN baselines. With only 0.75M parameters and a 2.87 MB footprint, it enables efficient on-site inference and reduced transmission requirements. Explainable AI visualizations are also provided to highlight tissue regions that drive model decisions.