Lung Nodule Segmentation in LDCT: Modified 3D nnUNet with Unified Focal Loss

Cancer ranks first among the top ten causes of death in Taiwan, and lung cancer has the highest mortality rate among all cancers. Pulmonary nodules are early signs of lung cancer. The growth rate, shape, location, and density of pulmonary nodules are all crucial information for evaluating the degree of malignancy. To calculate these features, accurate segmentation of pulmonary nodules is a necessary base. This paper contributes to the improvement of two existing problems: (1) applying Unified Focal Loss to greatly improve the segmentation accuracy of ground glass opacifications (GGO), and (2) improving the existing nnUNet model, using Res2Net Block combined with Dilated Convolution to strengthen the semantic communication between the encoding layer and the decoding layer and add multi-scale information to improve the segmentation performance of the Model. The model training uses the public data set of LIDC-IDRI (Lung Image Database Consortium Collection and Image Database Resource Initiative) and the pathological and health examination data provided by National Cheng Kung University Hospital. Our improved nnUNet can achieve an average Dice score of 83.4% on the public dataset LIDC-IDRI for 5-Fold Validation. Experiments show that our results have very competitive results in terms of stability and segmentation accuracy.