In recent years, U-Net has shown excellent performance in medical image segmentation, but it cannot accurately segment nodules of smaller size when segmenting pulmonary nodules. To make it more accurate to segment pulmonary nodules in CT images, U-Net is improved to REMU-Net. First, ResNeSt, which is the state-of-the-art ResNet variant, is used as the backbone of the U-Net, and a spatial attention module is introduced into the Split-Attention block of ResNeSt to enable the network to extract more diverse and efficient features. Secondly, a feature enhancement module based on the atrous spatial pyramid pooling (ASPP) is introduced in the U-Net, which is utilized to obtain more abundant context information. Finally, replacing the skip connection of the U-Net with a multi-scale skip connection overcomes the limitation that the decoder subnet can only accept same-scale feature information. Experiments show that REMU-Net has a Dice score of 84.76% on the LIDC-IDRI dataset. The network has better segmentation performance than most other existing U-Net improvement networks.
Keywords: Deep learning; Pulmonary nodules; Segmentation; U-Net.
© 2022. Australasian College of Physical Scientists and Engineers in Medicine.