Multi-Scale Optimized Feature Network for Polyp Segmentation

Polyps are one of the most common problems in human beings where some part on the surface of the body may bulk. This could be cancerous or non-cancerous. However, to segment these polyps it is very difficult to region them properly because of the different anatomy of human body, polyp shape, size etc. The smaller polyps are harder to segment than the bigger ones. Thus, we have proposed a network which helps to segment the polyps despite of its shape and size. The network consists of two major modules. First module is the multi-scale feature fusion. The second module consists of attention module. The purpose of first module is to represent the image clearly as it works upon multiple dilations on different rate. The second module enhances the relevant channel in the feature map by learning weights. The combination of these modules helps in better representation and thus we can further segment it easily.