DiffuPan: Diffusion-Based Framework for Multi-Phase Contrast-Enhanced CT Pancreatic Tumor Detection

Pancreatic cancer has a high mortality rate, and outcomes improve when tumors are identified early and delineated with precision. Contrast-enhanced CT is central to diagnosis and planning, yet classical segmentation approaches often miss irregular boundaries and show weak generalization across contrast phases. Recent convolutional and transformer architectures, including U-Net, Attention U-Net, and TransUNet, have raised baseline performance, but they typically rely on a single phase and struggle to capture complementary information across arterial, venous, and delayed acquisitions. This work presents DiffuPan, a diffusion-based encoder–decoder that performs cross-phase attention with residual feature fusion to couple information from all three phases. Training uses hybrid supervision that combines Dice, Focal, and SSIM losses to encourage accurate boundaries and coherence of fine structures. Experiments were run on the TCIA Pancreas-CT cohort comprising 300 patients and roughly 80,000 annotated slices. Ablation studies were designed to isolate the contributions of multi-phase fusion and diffusion guidance. DiffuPan obtained a Dice score of 92.3%, a precision of 93.1%, a recall of 92.0%, and an AUC of 0.97. These results exceed nnU-Net (88.2% Dice) and TransUNet (87.4% Dice) on the same data. The false-positive rate was 3.2% and the false-negative rate was 4.5%. The results suggest that the results suggest that the use of the diffusion-guided multi-phase integration is likely to result in more accurate tumor segmentations and more robust applicability across different scans, thus making it a proper choice for clinical segmentation of pancreatic lesions.