An EfficientNet-based hierarchical dual-encoder framework for multi-scale gastrointestinal disease detection

Abstract

Abstract Gastrointestinal (GI) diseases represent a major global health burden, making accurate and early diagnosis critical for improved clinical outcomes. We propose a dual-backbone convolutional framework that integrates EfficientNet-B0 and EfficientNet-B4 to jointly capture fine-grained local details and high-level global context in endoscopic images. The two feature streams are fused through residual learning with channel expansion–reduction and 1 $$\times$$ 1 convolutions, followed by a Convolutional Block Attention Module (CBAM) that adaptively emphasizes diagnostically relevant regions while suppressing background noise. To improve generalization and training stability, MixUp augmentation and Stochastic Weight Averaging (SWA) are employed, and a dropout-regularized classifier is used to mitigate overfitting. Experimental results demonstrate that the proposed method achieves an overall accuracy of 84.11% and a Macro-F1-score of 72.11%, consistently outperforming single-backbone EfficientNet variants and other baseline models, validating the effectiveness of the proposed architecture in both performance and efficiency.