Lung Nodule Segmentation Algorithm With SMR-UNet

Abstract

Accurate segmentation of lung nodules is of great significance for the early diagnosis of lung cancer. However, due to the diverse shapes and small sizes of lung nodules, lung nodule segmentation is a difficult task. In this paper, we propose an improved U-Net network called SMR-UNet, which integrates self-attention, multi-scale features and residual structures for lung nodule segmentation. The framework replaces the U-Net's convolutional units with residual units to ensure fast convergence, enhances the network's global modeling capability with Transformer, restores more detailed information with PixelShuffle, and enlarges the receptive field with a multi-scale feature fusion module before upsampling. The experiments show that on the LIDC dataset, the SMR-UNet achieves a Dice index of 0.9187 and an IoU of 0.8688, which are improved by 1.33% and 2.36% espectively compared to U-Net. When tested on lung nodules provided by the Department of Medical Imaging of the Fourth Affiliated Hospital of Guangxi Medical University, the SMR-UNet achieves a Dice index of 0.7785 and an IoU of 0.6541, which are improved by 4.22% and 4% espectively compared to U-Net.