Multi-Convolutional Channel Residual Spatial Attention U-Net for Industrial and Medical Image Segmentation

Abstract

Image segmentation has demonstrated immense potential in computer vision. In particular, the U-Net architecture, built on fully convolutional networks, is highly suitable for image segmentation tasks. Its encoder-decoder structure effectively captures both local and global features. This approach has achieved remarkable outcomes across various sectors, most notably in medical diagnostics and industrial quality control. However, U-Net, by employing skip connections, fuses different low-level and high-level convolutional features between the encoder and decoder, limiting its ability to effectively integrate useful features and harness contextual information. To address these feature disparities between the encoder and decoder, this paper introduces a novel network structure named Multi-Convolutional Channel Residual Spatial Attention U-Net (MCRSAU-Net). Designed for industrial and medical image segmentation, this model is anchored on the U-Net architecture. It replaces the traditional skip connections with channel attention residual paths featuring multiple convolutions, retaining more low-level features. Moreover, spatial attention module is incorporated in the decoding path to ensure the model concentrates on crucial regions of the input space, enhancing its segmentation capability across varied tasks. The proposed method was subjected to 5-fold cross-validation and testing on three public datasets: Mvtec AD, CHASE DB1, and Kvasir SEG. MCRSAU-Net achieved average Dice coefficients of 0.7755, 0.7651, and 0.8958 for defect segmentation of bottles, woods, and tiles, respectively, with average accuracies reaching 0.9751, 0.9815, and 0.9841. For retinal blood vessel and colon polyp segmentation, it exhibited superior performance, achieving average Dice scores of 0.8540 and 0.7053, and average accuracies of 0.9465 and 0.9195, respectively. These results not only underscore MCRSAU-Net's strong performance in image segmentation tasks but also demonstrate its significant potential in addressing specific challenges encountered in industrial and medical image segmentation.