FCMD-Net: A Full-Connection Multi-Decoder Network for Polyp Segmentation

Colorectal cancer is ranked 4th in the list of leading causes of death after lung, stomach, and liver cancers. Most colorectal cancers arise from growths in the mucosa of the colon called polyps. Therefore, early detection of polyps can help to faster monitor and treat for both patients and doctors. This requires an automatic, accurate, efficient, and reliable polyp segmentation system. Currently, U-shape architectures have shown promising performance with developments and advancements in encoder and decoder designs as well as skip connection solutions. However, existing skip connection approaches face some problems such as plenty of redundant information, or lack of generality causing to weakening of feature maps. To handle this challenge, we propose a new architecture namely Full-Connection Multi-Decoder Network (FCMD-Net) for the polyp segmentation task. First, we design a new model, using an encoder and three decoder branches to take advantage of multi-path information to increase inference speed and produce a better prediction mask. Besides, we create a skip connection method called full-connection to reduce effectively the semantic gap between two components encoder and decoder. In addition, we inherit and flexibly apply Atrous Spatial Pyramidal Pooling blocks to increase the performance of the model. The results on the Kvaisir-Seg dataset and CVC-ClinicDB dataset demonstrate that FCMD-Net has advantages over state of the art methods in the benchmark.