Integration of 3D Attention Mechanism and Atrous Convolution for Lung Nodule Segmentation

Lung nodules, as an initial symptomatic indication of lung cancer, have been a significant concern for human health and well-being. In order to achieve more accurate segmentation of lung nodules in CT images, a 3D U-net segmentation method is proposed that incorporates the attention mechanism and dense cavity convolution to address the incomplete acquisition of spatial features of lung nodules by the U-net network. Firstly, a 3D Convolutional Block Attention Module is added after the jump connection to avoid the loss of spatial dimensional information due to the sequence-dependent channel attention by optimizing the parallel channel and spatial attention modules, and to improve the attention of the model to the key features, which helps the network focus on important lung nodule characteristics. Second, by adding the Residual-Atrous Spatial Pyramid Pooling block at the bottleneck of the network, the network can improve the receptive field while maintaining high resolution, and capture more information between layers of CT slices. In this paper, Luna16 and a case dataset from a hospital in Southwest China are used, and the experimental results show that the four metrics of Dice, Recall, Precision, and Jaccard have been improved by 3.14%, 2.06%, 1.67%, and 1.69%, respectively, and the segmentation results of the model in this paper are also closer to the gold standard compared to other mainstream methods.