PanDiffuNet: Multi-Phase Diffusion Learning for Pancreatic Tumor Detection in Contrast-Enhanced CT

Pancreatic cancer remains one of the most fatal malignancies, and computed tomography (CT) is the main imaging route for its assessment. Early tumors are small and difficult to distinguish from surrounding tissue, and a significant proportion of them are not detected during the routine reading. Although conventional convolutional networks like three-dimensional (3D) U-Net and subsequently transformer models have been able to boost detection, they lack the ability to cope with limited tumor samples and varying contrast phases. This work introduces a phase-conditioned diffusion framework (PanDiffuNet) for pancreatic tumor detection, localization, and segmentation. It combines diffusion-based augmentation with phase-aware denoising, cross-phase consistency and anatomical context guidance. The PanTS dataset consists of 9,901 3D abdominal CT scans and voxel-wise tumor masks in the arterial, portal-venous and delayed phases for training and evaluation. PanDiffuNet achieves an accuracy of 96.4%, a sensitivity of 95.3%, an area under the receiver operating characteristic (ROC) curve (AUROC) of 0.987, a Dice score of 0.842, and 6.9 mm 95th-percentile Hausdorff distance (HD95). The framework demonstrates that a combination of balanced synthetic augmentation and phase conditioning improves the detection accuracy and the accuracy of the boundaries for pancreatic CT analysis.