Semantic segmentation of pulmonary nodules based on attention mechanism and improved 3D U-Net

Abstract

In lung CT images, the detection and diagnosis of pulmonary nodules is one of the important criteria for many pulmonary diseases. In recent years, image semantic segmentation technology has developed rapidly and has been gradually applied to the medical field. However, the existing segmentation methods of lung nodules need a lot of labeling work by professionals before training, and the segmentation results have some problems such as low accuracy and blurred image edges. In order to improve the above problems, in this study, the existing 3D U-Net network is improved, and the double Attention structure is applied to the 3D semantic segmentation network structure. The structure can focus Attention on the edge of the segmentation target, so as to improve the detail accuracy of the segmentation edge of the target lung nodule. Aiming at the problem of inconsistent output information of Attention structure, a new joint loss function is used to optimize. The trained network structure is tested on LUNA16 dataset, and the Dice value is 90.31%. The comprehensive performance of other test indexes is also better than that of other network structures. This study can provide a reference for semantic segmentation of lung CT images using deep learning methods.