DF-ResTransUNet: Dual-Stream Feature Fusion Residual TransUNet for Polyp Segmentation

Over 90% of bowel cancers stem from polyps, making accurate polyp segmentation vital for doctors to assess conditions and plan surgical interventions. The current leading polyp segmentation methods typically employ U-shaped encoder–decoder architectures that integrate Transformer modules. By introducing Transformer modules into traditional UNet-like convolutional networks, these models enhance their capacity to capture long-distance dependencies. However, they often grapple with challenges like limited multi-scale capabilities, vanishing gradients due to overly deep Transformer layers, and semantic discontinuities caused by skip connections. These issues constrain the potential for further enhancing the model’s polyp segmentation performance. To tackle these problems, we introduce an enhanced TransUNet model, named Dual Stream Feature Fusion Residual TransUNet (DF-ResTransUNet). This model incorporates inter-layer residual connections within the Vision Transformer block to alleviate gradient vanishing. Additionally, we replace traditional skip connections with a novel dual-stream feature fusion module, equipping the model with a robust ability to learn multiscale features. Furthermore, our new feature fusion screening module within the decoder section contributes to further improvements in segmentation performance. Extensive experiments on Kvasir-SEG, CVC-612, and ETIS datasets demonstrate that our method achieves state-of-the-art polyp segmentation performance. Ablation studies and evaluations on model efficiency further validate the effectiveness and practicality of our proposed modules and the simplified model variant.