RPMS-DSAUnet: A Segmentation Model for the Pancreas in Abdominal CT Images

Automatic pancreas segmentation in CT scans is crucial for various medical applications including early disease detection, treatment planning and therapeutic evaluation. However, the pancreas’s small size, irregular morphology, and low contrast with surrounding tissues make accurate pancreas segmentation still a challenging task. To address these challenges, we propose a novel RPMS-DSAUnet for accurate automatic pancreas segmentation in abdominal CT images. First, a Residual Pyramid Squeeze Attention module enabling hierarchical multi-resolution feature extraction with dynamic feature weighting and selective feature reinforcement capabilities is integrated into the backbone network, enhancing pancreatic feature extraction and improving localization accuracy. Second, a Multi-Scale Feature Extraction module is embedded into the network to expand the receptive field while preserving feature map resolution, mitigate feature degradation caused by network depth, and maintain awareness of pancreatic anatomical structures. Third, a Dimensional Squeeze Attention module is designed to reduce interference from adjacent organs and highlight useful pancreatic features through spatial-channel interaction, thereby enhancing sensitivity to small targets. Finally, a hybrid loss function combining Dice loss and Focal loss is employed to alleviate class imbalance issues. Extensive evaluation on two public datasets (NIH and MSD) shows that the proposed RPMS-DSAUnet achieves Dice Similarity Coefficients of 85.51% and 80.91%, with corresponding Intersection over Union (IoU) scores of 74.93% and 67.94% on each dataset, respectively. Experimental results demonstrate superior performance of the proposed model over baseline methods and state-of-the-art approaches, validating its effectiveness for CT-based pancreas segmentation.