This paper proposes a pancreatic inflammation segmentation model based on the UNet architecture, integrating the SE module and multi-scale feature fusion mechanism. Experimental validation demonstrates that the improved model exhibits excellent performance in pancreatic region segmentation tasks. The main contributions of this work include: (1) Integration of the SE module into the skip connections of UNet, enhancing channel attention to improve the model’s focus on critical pancreatic features; (2) Adoption of a multi-scale feature fusion mechanism, effectively combining high-resolution and low-resolution features during the progressive upsampling and skip connection process, further improving segmentation accuracy; (3) Design of a combined loss function incorporating Dice, Focal, and IoU losses, which significantly enhances segmentation performance in complex scenarios. Experimental results show that the proposed method achieves a Dice score of 87.32%, IoU of 85.13%, and Precision of 86.04%. Compared to the traditional UNet model, this model more accurately segments the pancreatic region and shows significant advantages in handling blurred boundaries and complex lesions.