Pancreatic multi structure segmentation on CT using transformer and deep learning approaches

To comprehensively evaluate and compare the performance of six advanced deep learning architectures for automatic segmentation of pancreatic tumors and adjacent anatomical structures in contrast-enhanced CT images. The multi-center study included 1968 CT scans that were acquired from 6 different medical institutions and thus covered a large range of imaging hardware and acquisition protocols. Six anatomical structures - pancreatic tumour, parenchyma, duct, common bile duct, venous system, and arterial system were manually segmented by experienced radiologists. Intensity normalization, spatial resampling and data augmentation were used as pre-processing steps. The evaluated architectures were U-Net, nnU-Net, ATTENTION-BASED RESIDUAL U-Net, DEEPLABV3+, SEGNET and SWIN-UNET. Each network was trained using a Dice loss together with cross-entropy and the Adam optimizer with a cosine learning rate scheduler. Internal and external validation was performed on an entirely independent external cohort (n = 536) using the following measures: Dice coefficient, Jaccard Index, Precision, Recall and 95th percentile Hausdorff Distance (HD95). Wilcoxon signed-rank tests were used to compare the two groups; Bonferroni correction for multiple comparisons was used. Swin-Unet outperformed all other models in terms of volumetric and boundary metrics: Dice scores were obtained ranging from 0.94 to 0.96 and HD95 as low as 2.9 mm. nnU-Net achieved the second place and proved to be highly generalizable, although slightly lower accuracy was obtained for small/material tortuous structures. Transformer and attention-enhanced models significantly outperformed classical CNNs, especially in segmenting ducts and vascular anatomy. We demonstrated that the robust generalisation of Swin-Unet was validated externally with minimal degradation in performance when generalized to different institutions. Statistical results showed that Swin-Unet is significantly superior (p < 0.001) in all considered structures. Among the six contemporary deep learning architectures compared in this large multi-center CT study with external validation, Swin-Unet achieved the highest segmentation accuracy and generalization, particularly for vascular structures and pancreatic parenchyma, while performance on thin and low-contrast ductal structures remains an area for further improvement.