Lung Nodule Segmentation Method Based on 3D Rem-Unet

Abstract

For the problems of inaccurate positioning and imprecise edge segmentation of lung nodules in 3D medical computed tomography (CT) images, a 3D Rem-UNet method for lung nodule segmentation is proposed. Using 3D UNet as the reference model, a multi-branch hybrid attention (MHA) module is added to the encoder structure to improve the extraction of detailed features in the target region; The residual connection is used to realize feature reuse; And the Group Normalization (GN) algorithm is introduced to achieve the feature normalization. Finally, ablation experiments are conducted on the LUNA16 dataset to verify the effectiveness of the model. The experimental results show that 3D Rem-UNet has the fastest loss convergence and the most stable loss change during the training process. Meanwhile, compared with 3D UNet model, the dice coefficient, accuracy, recall and F1 score are increased by 4.73%, 14.6%, 1.76% and 4.71%, respectively. The model can effectively improve the segmentation effect of lung nodules.