A new attempt at full-scale jump connectivity and Transformer

Abstract

Abstract In this research, we introduce SwinUnet3+, a pioneering algorithm that integrates Unet with Transformer, to facilitate the automatic segmentation of three primary tissues—subcutaneous fat layer, muscle, and intramuscular fat—in the thoracoabdominal region under challenging conditions, including subcutaneous soft tissue swelling, gas accumulation, artifacts, and fistulas. Our model showcases superior performance in body composition segmentation tasks, with improvements in DSC, IoU, sensitivity, and positive predictive value by 3.2%, 6.05%, 4.03%, and 2.34%, respectively. Notably, in segmenting subcutaneous fat, intramuscular fat, and muscle, SwinUnet3 + yielded the best outcomes. However, the model does exhibit certain limitations, such as a reliance on vast amounts of training data and potential challenges in handling certain image types. Additionally, high-resolution images may pose computational efficiency concerns. In conclusion, while SwinUnet3 + offers considerable advantages in complex medical image segmentation tasks, its limitations warrant acknowledgment. Future research will focus on addressing these challenges and enhancing the model's robustness and generalization capabilities.