Abstract
Pancreatic cancer, as a highly lethal malignant tumor, presents significant challenges for early diagnosis and treatment. Accurate segmentation of the pancreas and tumors is crucial for surgical planning and treatment strategy development. However, due to the variable morphology, blurred boundaries, and low contrast with surrounding tissues in CT images, traditional manual segmentation methods are inefficient and heavily reliant on expert experience. To address this challenge, this study proposes a lightweight automated 3D segmentation algorithm-Multi-Dimensional Decoupled Multi-Scale U-Net (MDMU-Net). First, depthwise separable convolution is employed to reduce model complexity. Second, a multi-dimensional decoupled multi-scale module is designed as the primary encoder module, which independently extracts features along depth, height, and width dimensions through parallel multi-scale convolutional kernels, achieving fine-grained modeling of complex anatomical structures. Finally, cross-dimensional channel and spatial attention mechanisms are introduced to enhance recognition capability for small tumors and blurred boundaries. Experimental results on the MSDPT and NIHP datasets demonstrate that MDMU-Net exhibits competitive advantages in both pancreatic segmentation DSC (0.7108/0.7709) and tumor segmentation DSC (showing an 11.8% improvement over AttentionUNet), while achieving a 15.3% enhancement in HD95 boundary accuracy compared to 3DUX-Net. While maintaining clinically viable precision, the model significantly improves computational efficiency, with parameter count (26.97M) and FLOPs (84.837G) reduced by 65.5% and 71%, respectively, compared to UNETR, providing reliable algorithmic support for precise diagnosis and treatment of pancreatic cancer.