mcU-TransNet: A more Comprehensive UNet-Transformer Network for Polyp Segmentation

Purpose: Polyp segmentation is crucial for localizing areas that are at risk of early cancer. Traditional Convolutional Neural Networks (CNNs) have limitations in capturing the global context and long-range dependencies. In this paper, a more comprehensive UNet-Transformer network (mcU-TransNet) is proposed for polyp segmentation. Methods: mcU-TransNet combines the strengths of CNN-based approaches with Transformer-based approaches. To tackle the challenge of detecting different types of polyps at the same location, an attention-enhanced UNet backbone is introduced. This backbone integrates a stem block into the UNet encoder. Spatial and channel attention mechanisms in the UNet decoder are employed to learn specific polyp features, thereby enhancing contour extraction accuracy. To overcome the limited fine-grained perception of the Transformer model, a Transformer head with deformable separable kernel attention (DSKA) is utilized to capture global context information. Convolutional parallel dilated convolutions are employed to gather both local and distant context information, thereby addressing spatial perception deficiencies in the unfolding sequence. Results: Experimental results on the CVC-ClinicDB and Kvasir-SEG datasets demonstrate that mcU-TransNet outperforms existing state-of-the-art methods, showcasing its robustness and ease of training. Conclusion: mcU-TransNet could comprehensively learn dataset features and enhance colonoscopy interpretability for polyp detection.