An Investigation into a Lung Nodule Segmentation Algorithm Based on Improved Swin-Unet

In the realm of medical picture segmentation, solving the problem of imbalanced sample distribution is an important research direction. The majority of existing work on the problems of sensitive initial contour position, difficult convergence, and poor repeatability of abnormal lung boundary depression focuses on model optimization of single loss functions, and there is little discussion and analysis of multi-loss function collaborative models. This paper investigates a method for segmenting lung nodules based on an improved Swin-Unet network model: a residual network module is used to simplify network training and avoid gradient disappearance; combining the Focal Loss function with the Dice Loss function can increase the weight of nodules and effectively handle the problem of uneven sample distribution. The segmentation network emphasizes the central position and edge details of pulmonary nodules in addition to the local properties of the nodules. Tests were carried out on the LNDb data set to verify the model’s performance, which can accurately separate pulmonary nodules and has better segmentation performance.