Attention-Driven, Normalization-Free Deep Learning for Robust and Explainable Gastrointestinal Disease Detection.

Accurate classification of gastrointestinal (GI) diseases from endoscopic images is challenging due to visual similarity among disease categories, class imbalance, and variability in imaging conditions. To address these issues, this study proposes an attention-driven and normalization-free deep learning framework based on the ECA-NFNet-L0 architecture for robust GI disease classification. Efficient Channel Attention (ECA) is integrated into NFNet blocks to enhance channel-wise feature representation while avoiding batch normalization, enabling stable and efficient training. The model is trained using weighted random sampling to mitigate class imbalance, TrivialAugmentWide for data augmentation, and label-smoothed cross-entropy loss with differential learning rates for effective fine-tuning. Experiments conducted on the GastroVision dataset demonstrate strong performance, achieving Matthews Correlation Coefficient (MCC) scores of 84.80\%, 79.71\%, and 86.29\% for the combined, upper-GI, and lower-GI subsets, respectively, outperforming several state-of-the-art CNN baselines. Furthermore, explainability analyses using Grad-CAM and LIME confirm that the proposed model focuses on clinically relevant regions of endoscopic images, supporting interpretability and clinical reliability. The proposed framework also shows strong generalization capability across different GI subsets, highlighting its robustness in real-world diagnostic scenarios. These findings suggest that attention-enhanced normalization-free networks can serve as a reliable foundation for future video-based and multi-dataset GI disease classification systems.