Gland and Colonoscopy Segmentation Method Combining Self-Attention and Convolutional Neural Network

The automatic segmentation of glands and polyps is the foundation for the diagnosis of artificial intelligence-assisted colorectal adenocarcinoma. However, the size and shape of segmentation targets in medical images vary considerably, and the automatic segmentation approach based on a convolutional neural network has thus run into a hindrance. Therefore, a dual branch network (LG UNet) combining convolutional neural network and self attention is proposed to improve the accuracy of segmentation. First, the Local UNet branch was developed based on U-Net, and the convolutional neural network's benefits were employed to elucidate the segmentation target's local information. Subsequently, the segmentation details were optimized using the Transformer's learning ability of global dependencies in the Global Transformer branch. Finally, during the encoding process, feature maps of the Local and Global branches were merged by a cross-fusion module to complement their benefits. The two test subsets of Glas and findings of LG UNet were 93. 62% and 88. 44% for Test A and 88. 17% and 80. 49% for Test B, employing the Dice coefficient and intersection and union (IOU) coefficient as the primary examination indexes. Furthermore, the Dice and IOU coefficients in the polyp segmentation dataset Kvasir-SEG were 85. 63% and 77. 82%, respectively. The experimental findings demonstrate that LG UNet exhibits better performance efficiency in gland and polyp segmentation by combining the benefits of the Transformer and convolutional neural network.