Fusing Transformer and FCN for Polyp Segmentation

Colonoscopy is a medical examination method used to identify colon polyps and plays a crucial role in preventing the occurrence of serious colon abnormalities. This is because colonoscopy enables the direct visualization of the affected area, enhancing its significance in healthcare. However, manually segmenting polyps in colonoscopy images is very difficult, so it is important to achieve accurate polyp segmentation using deep learning techniques. The polyp images in different colonoscopies have complex diversity, and existing deep learning models are prone to overfitting and cannot handle unexposed polyp images, Therefore, enhancing the models’ generalization capability is crucial.In addressing this issue, we introduce a novel model for polyp segmentation. Our model adopts a 4-layer pyramid structure, and the encoder combines Transformer and FCN, where FCN can better extract features containing local information, while Transformer can extract global features better through its self-attention mechanism. By combining FCN and Transformer, the encoder of the model can extract both local features and global features at multiple scales to better understand the image content and improve the generalization ability of the model. In addition, our model uses a PLD decoder. Our model excels in the field of image segmentation, demonstrating outstanding generalization performance that surpasses that of currently available models.