M3ResU-Net: a deep residual network for multi-center colorectal polyp segmentation based on multi-scale learning and attention mechanism

Abstract

Colorectal polyps are considered as an important precursor of colorectal cancer (CRC) in clinical diagnosis. A network automatically and accurately segmenting polyps can recognize, locate and finally help to remove polyps, greatly reducing the misdiagnosis rate. Although many neural networks for polyp segmentation have been proposed, there still exist some difficulties including the diversity of image backgrounds, the jelly effect, and the various shapes and sizes of different polyps. These factors lead to the segmentation accuracy remaining to be improved. In this paper, we propose M3ResU-Net including multi-scale learning and attention mechanisms, aiming to segment multi-center colorectal polyps. First, we implement the contrast limited adaptive histogram equalization (CLAHE) and data augmentation for multi-center data. Then, channel and spatial attention mechanisms are introduced to focus on polyp features and suppress interference features. Finally, in order to balance small target segmentation and the acquisition of global information, multi-scale learning with dilated convolutions is employed. We compared other five polyp segmentation methods on three publicly available datasets. In single-center experiments, M3ResU-Net reaches a Dice similarity coefficient (DSC) exceeding that of the best compared method by over 2%. In various multi-center experiments, M3ResU-Net all achieves a DSC over 0.8. The results demonstrate that M3ResU-Net is capable of assisting clinicians in polyp segmentation in the field of colonoscopy, which provides important and reliable support to improve diagnostic efficiency.