A Multi-feature Extraction Decoder for Polyp Detection

Early detection of human rectal colorectal cancer is an important problem. The research in this field shows that early detection lowers the disease’s fatality rate. The gold standard for identifying rectal polyps is colonoscopy. Accurately segmenting polyps from it frequently yields important clues for prompt course of action. Even though the accuracy of the deep learning models now in use is high, their computational complexity is high and this frequently leads to overfitting or poor model generalization. This difficulty is frequently linked to the capability of the features that are extracted from the input images. This work proposes an enhanced architecture for encoder-decoder with a Multiscale Deep Context Feature Refinement module for polyp region segmentation using colonoscopy images at the decoder. Overall, the proposed network is reasonably less complex with less number of parameters. This helps in tackling the problem of low model generalization and improved segmentation performance.