Hybrid TransUNet and Attention U-Net for AI-Based Colonoscopy Polyp Detection

Colonoscopy remains the gold standard for the detection of polyps and the early detection of colorectal cancer. Automated polyp identification based on deep learning can reduce human error and improve diagnosis accuracy. This study proposes a hybrid model that combines Attention U-Net as the decoder and TransUNet as the encoder to enhance polyp segmentation. The encoder uses transformer-based global feature extraction, while the decoder uses attention techniques to refine the segmentation mask. A hybrid loss function that combines Dice and Focal loss is used to improve boundary precision and address class imbalance. Better segmentation performance is demonstrated by experimental tests on the CVC dataset, where the proposed model outperforms traditional approaches.