Beyond transfer learning: Attention-enhanced deep learning framework for multiclass gastrointestinal disease classification

Gastrointestinal (GI) diseases pose a significant global health burden. Therefore, early and accurate diagnosis is paramount for effective clinical intervention. While deep learning (DL) has shown promise in automating GI disease classification, existing approaches often rely on conventional transfer learning (TL) methods that employ pretrained convolutional neural networks (CNN) either as standalone classifiers, ensemble components, or fixed feature extractors for subsequent feature selection and classification stages. These fragmented and generic pipelines suffer from several limitations, including limited adaptability to domain-specific features, increased computational complexity, and reduced clinical interpretability. To address these challenges, we propose a novel end-to-end DL framework, called Inception-Residual-Attention-DenseNet201 (IRA-DenseNet201), that redefines TL by embedding a DenseNet201 backbone within a unified task-specific architecture. The proposed framework encompasses multiscale Inception modules, residual bottleneck blocks, and a comparative study of five advanced attention mechanisms to dynamically refine feature representations. A global context block is appended to enhance semantic coherence, while gradient-weighted class activation mapping is employed to provide visual explanations of the model predictions. IRA-DenseNet201 achieved classification accuracy of 94.67 % on the Kvasir dataset. Squeeze and excitation mechanism delivered the best performance in GI disease classification. Comparative experiments revealed that the proposed model consistently outperformed conventional CNN architectures in GI detection. Unlike traditional pipelines, our approach bridges the gap between the simplicity of single pretrained models and the complexity of ensemble or multi-phase frameworks, offering a clinically viable solution for automated GI disease diagnosis. This work sets a new direction for adaptive and explainable TL in medical image analysis.