Conformer: A Parallel Segmentation Network Combining Swin Transformer and Convolutional Neutral Network

Abdominal organ segmentation can help doctors to have a more intuitive observation of the abdominal organ structure and tissue lesion structure, thereby improving the accuracy of disease diagnosis. Accurate segmentation results can provide valuable information for clinical diagnosis and follow-up, such as organ size, location, boundary status, and spatial relationship of multiple organs. Manual labels are precious and difficult to obtain in medical segmentation, so the use of pseudo-labels is an irresistible trend. In this paper, we demonstrate that pseudo-labels are beneficial to enrich the learning samples and enhance the feature learning ability of the model for abdominal organs and tumors. In this paper, we propose a semi-supervised parallel segmentation model that simultaneously aggregates local and global information using parallel modules of CNNS and transformers at high scales. The two-stage strategy and lightweight network make our model extremely efficient. Our method achieved an average DSC score of 89.75% and 3.78% for the organs and tumors, respectively, on the testing set. The average NSD scores were 93.51% and 1.82% for the organs and tumors, respectively. The average running time and area under GPU memory-time curve are 14.85 s and 15963 MB.