2D Medical Image Segmentation Combining Multi-Scale Channel Attention and Boundary Enhancement

Abstract

Listen

Translate article

2D medical image segmentation plays an important role in disease diagnosis and computer-assisted treatment. Aiming at the problem that 2D medical image is difficult to accurately segment due to target size, shape and blurred boundaries, a method combining multi-scale channel attention and boundary enhancement is proposed. First, 2D medical image is received as input, and the encoder and boundary enhancement module are used to extract the high-level feature maps and boundary segmentation result respectively. Then the multi-scale channel attention module extracts context information of different scales from the high-level feature map, enhances the useful features and suppresses the useless feature response. Finally, the context information obtained in the previous step is passed into the decoder to obtain the region segmentation result, and integrated with the boundary segmentation result to obtain the final segmentation result. In order to resolve the problem of data imbalance in medical images, a custom loss function is proposed. Experimental results on the four datasets of tooth panoramas containing residual roots, tooth panoramas containing caries, retinal vessels, and skin lesions show that the segmentation precision of proposed method has reached 85.63%, 70.15%, 75.86% and 85.92%, respectively. Compared with other medical image segmentation methods, proposed method performs better.