DMC-UNet-Based Segmentation of Lung Nodules

The accurate and rapid segmentation of different categories of lung nodules is of great importance for the diagnosis of early stage lung cancer and to assist physicians in the diagnosis and treatment of the disease. In the segmentation process, there are various types of lung nodules with different shape characteristics and occupying small volumes, so the process of segmenting lung nodules out is challenging. The DMC-UNet network proposed in this paper is an improved network based on UNet. The DMC-UNet network combines a lightweight residual structure, multiscale feature upsampling fusion and X/Y Channel Attention Module and Coordinate Attention (CCA) attention mechanism. The overall framework of the network firstly replaces the convolutional units of U-Net with residual units, and replaces the traditional convolution in the residual units with Depthwise Separable Convolution (DSC) to reduce the number of parameters and computation of the model and improve the efficiency of model training and prediction, and secondly replaces the transposed convolution and PixelShuffle in the upsampling process of U-Net with parallel direction fusion to replace the transposed convolution used in the original U-Net, which can enable the model to better capture information at different scales, and the addition of a multiscale feature fusion module before PixelShuffle improves the traditional Efficient Sub-Pixel Convolutional Neural(ESPCN) model, which aims to expand the perceptual field, and finally, the addition of a CCA attention mechanism after the upsampling fusion can better recover spatial information. It is shown by experiments that the IoU and F1-score of DMC-UNet are 65.52%±0.71% and 76.02%±0.63%, respectively, on the lung nodules provided by the Department of Medical Imaging of the Fourth Affiliated Hospital of Guangxi Medical University (FAHGMU), and the absolute gains of IoU and F1-score compared with U-Net are 2.78% and 2.91% on the Lung Image Database Consortium(LIDC) public dataset, and 83.36% and 89.92% on the IoU and F1-score, respectively, with a gain of 1.37% and 0.73% on the IoU and F1-score, respectively, compared to the U-Net.