Attention-enhanced residual U-Net with histogram equalization for automated classification of gastrointestinal bleeding disorders

초록

Gastrointestinal (GI) diseases range from benign conditions to life-threatening disorders, and their diagnosis traditionally relies on clinical evaluations, imaging techniques, and surgical procedures, which can be timeconsuming, costly, and prone to inconsistencies. To address these limitations, this study proposes a novel deep learning-based classification framework-Histogram Equalized Attention Residual Gate U-Net (HARUNet)-for accurate identification of GI disorders. The proposed research contributes in two key areas: first, by enhancing image quality through histogram equalization to improve feature extraction, and second, by advancing the Attention U-Net architecture with residual gating to improve classification performance. The upsampling decoder layers of HARU-Net integrate refined attention mechanisms derived from the Attention U-Net, while the ReLU layers in attention gates are replaced with normalized residual blocks to generate optimized feature maps. The model was trained and evaluated using the Gastrointestinal Bleeding Images Dataset from Kaggle, comprising 113 bleeding and 113 normal wireless capsule endoscopy images. After labeling and data augmentation, the dataset expanded to 4,746 images, with 3,146 used for training, 800 for validation and 800 for testing. Preprocessing techniques such as Gaussian smoothing, median and high-pass filtering, and histogram equalization were applied to enhance visual features. Experimental results demonstrate that HARU-Net significantly outperforms conventional deep learning models, achieving a classification accuracy of 98.21%, compared to over 85% with existing CNN architectures including ResNet and the standard Attention U-Net. By incorporating histogram equalization to enhance image quality and integrating residual gating into the Attention U-Net architecture, the model effectively refines feature extraction and boosts classification accuracy. These findings suggest that HARU-Net is a highly effective model for automated classification of gastrointestinal diseases, with potential to support faster and more reliable diagnostics. To promote transparency and reproducibility, the implementation and supplementary materials supporting this study are available at Zenodo (https://doi. org/10.5281/zenodo.15619214).

키워드