A SYMMETRIC THREE-DIMENSIONAL ConvNeXt NETWORK FOR LUNG NODULE SEGMENTATION IN CT IMAGES

Accurately segmenting lung nodules from the background to provide doctors with crucial information, such as shape and size, remains a challenge in medical image diagnosis. This paper employs a three-dimensional ConvNeXt as the backbone network to construct a symmetric segmentation network model. Residual connections are added to the downsampling and upsampling modules in this model to minimize information loss during the sampling process. Considering the diverse morphologies of lung nodules, the model also incorporates a spatial attention mechanism to enhance its perception of critical information, thus improving generalization and overall segmentation results. To validate the performance of the segmentation method, comparative experiments were conducted on the LUNA-16 and LNDb-19 datasets. The model demonstrated superior performance in the main metrics such as the Dice similarity coefficient, achieving 89.15% and 84.30%, respectively.