Abdominal Multi-organ Segmentation Algorithm Based on Small Samples and TransU-Net

Automated medical image segmentation technology assists doctors in making comprehensive judgments by combining multiple sources of information, enabling more accurate and detailed diagnoses for patients' conditions, the proposition of reasonable treatment plans, and reasonable disease prognosis. Such advancements significantly improve treatment outcomes, reduce surgical risks, and lower the incidence of postoperative complications. To address the poor performance of TransU-Net on small datasets and the issue of the loss of local information, we introduce a hierarchical feature extraction structure similar to the nnUNet [2] baseline network so that the algorithm can adapt to small sample datasets without compromising TransU-Net's capability for large sample operations or its transferable potential. Based on TransU-Net [1], this experiment proposes an algorithm designed for abdominal multi-organ segmentation on small sample training datasets. Experimental results demonstrate that the algorithm achieves a higher Intersection-over-Union (IoU) and a lower loss value while maintaining a nearly constant quantity of parameters. Compared with the NestedUNet model, which excels in small sample operations, the proposed model improves IoU by 2%. Compared to the lightweight TransU-Net model, it increases IoU by 10% and reduces the loss value by 5%.