Performance Analysis of Pancreas Segmentation from CT Images Using Deep Learning

Pancreatic diseases pose a serious threat to human health. Computed Tomography (CT) is the primary imaging technique for diagnosis, and pancreas segmentation plays a crucial role in computer-aided diagnosis. To address the challenges of variable pancreatic morphology and blurred boundaries, this study systematically compares the performance of four widely used deep learning models—TransUNet, SwinUNet, nnUNet, and nnFormer—in pancreatic segmentation tasks. Experiments were conducted on the public NIH dataset and the upper abdominal CT dataset from Ruian People’s Hospital, with comprehensive evaluation using metrics including the Dice coefficient, Jaccard index, and HD95. Results demonstrate that TransUNet exhibits outstanding robustness and generalization capability across different datasets, achieving a Dice coefficient of 88.65% and an HD95 of 5.32 on the Ruian dataset, representing the best overall performance. Ablation experiments further validated the impact of input resolution and the number of skip connection on segmentation performance, indicating that higher input resolution and an appropriate number of skip connections significantly enhance model performance. This study provides a reference for model selection and a basis for methodological improvements in pancreatic CT image segmentation, contributing to the early diagnosis of pancreatic diseases.