Go to the Right: A Real-Time And Accurate Polyp Segmentation Model for Practical Use

Although deep learning-based approaches have achieved a fairly good performance in polyp segmentation, most of them are unable to be deployed in real-world medical scenarios due to the excessive computational effort and the large model size. To deal with this issue, we propose a real-time and accurate polyp segmentation model for practical use. Specifically, we construct it from two respects. First, we try to effectively extract global contextual features to improve segmentation performance. Thus, we leverage large kernel attention to effectively capture global contextual dependencies while maintaining the low computational effort. Second, we improve the learning capability of the decoder and introduce feature transitions to enhance the performance of cross-level feature fusion. Extensive experiments on five polyp segmentation benchmarks demonstrate that our proposed model achieves state-of-the-art performance with less than 5M parameters and runs at 82.66 FPS on a GeForce 1080Ti GPU. Code will be released.