Dual-Domain Feature Interaction Network for Automatic Colorectal Polyp Segmentation

Recently, many deep neural network-based methods have been proposed for polyp segmentation. Nevertheless, most methods primarily analyze spatial information and usually fail to accurately localize polyps with inconsistent sizes, irregular shapes, and blurry boundaries. In this article, we propose a dual-domain feature interaction network (DFINet) for automatic polyp segmentation to overcome these difficulties. DFINet has an encoder-decoder structure that includes two key modules: a spatial and frequency feature interaction (SFFI) module and a boundary enhancement (BE) module. To learn shape-aware information, the SFFI module is deployed at each layer of the encoder, where spatial and frequency features are simultaneously extracted and fused using the attention mechanism. Such a module helps the network adjust to the polyps with irregular shapes and blurry boundaries. The BE module is used to enhance the boundary areas by integrating cross-layer features of SFFI modules with the prediction map of the adjacent high layer. Since there is no higher layer for the top layer, we integrate the multiscale features of the encoder to generate a prediction map for the BE module at the top layer. Such configuration helps the network handle the challenge of inconsistent sizes. By connecting the BE modules from top to bottom and applying deep supervision, DFINet can generate coarse-to-fine prediction maps. Results of both in-domain and out-of-domain tests show that DFINet achieves good segmentation results, with stronger learning ability and better generalization ability than 11 state-of-the-art methods.