ELRFNet: Effective Large Receptive Field Based ConvNext for Medical Segmentation

The use of colonoscopy to detect intestinal polyps is the key to early detection and treatment of colorectal cancer, but this method requires a large labor cost. So utilizing deep learning to automatically segment the location of colorectal polyps undoubtedly alleviates the heavy burden on healthcare. In this paper, the ELRFNet model is proposed to automatically segment the polyp location, the Encode of the model adopts ConvNext, the Decode uses UperNet, the overall morphology is U-shaped structure. And the ELRF module is used on the first layer of the feature map, which boosts the model's understanding of the global features through the decomposition of large convolutional kernels. Then, the extracted feature matrix undergoes Singular Value Decomposition for decomposition, and subsequently, the number of channels is adjusted based on the results to alleviate the computational burden. Ultimately, the Dice coefficient for rectal polyp localization reached 90.91, with an average Dice coefficient of 95.165, surpassing that of other mainstream methods.