Enhancing GI Abnormality Classification through Attention: A Parameter-Efficient and Explainable Dual-Pool MSCA Architecture

Automated classification of gastrointestinal (GI) abnormalities offers significant potential to reduce clinical workload, enhance diagnostic accuracy, and improve patient outcomes. However, many existing methods rely on computationally intensive deep learning (DL) architectures, which limit their practicality in real-world and resource-constrained settings. This study presents a lightweight diagnostic framework by integrating attention mechanisms into the MobileNet-v2 backbone. Central to the approach is a novel Dual-Pool Multi-Scale Channel Attention (MSCA) module that extends conventional Channel Attention (CA) through multi-scale feature enhancement. We evaluate the framework using transfer learning strategies, including both pre-trained and fine-tuned models, on the Kvasir-v2 dataset, which contains 8,000 labeled endoscopic images across eight GI classes. A five-fold cross-validation ensures robust and reliable performance assessment. The proposed Dual-Pool Dual Block MSCA configuration achieves an accuracy of 93.84%, surpassing established attention mechanisms—Squeeze-and-Excitation (SE), Spatial Attention (SA), and Convolutional Block Attention Module (CBAM)—as well as non-attention-based transfer models. To improve interpretability, we visualize activation responses before and after applying attention using average channel maps and Gradient-weighted Class Activation Mapping (Grad-CAM). Additional analysis using confusion matrices and ROC-AUC metrics further confirms the model’s reliability and robustness. Overall, the proposed method contributes to the advancement of resource-efficient DL in medical imaging by balancing computational feasibility with diagnostic precision.