Purpose <p>Accurate segmentation of the pancreas and pancreatic tumors from abdominal CT scans is a critical task in medical imaging, complicated by the small size and irregular shape of tumors and significant class imbalance. This study aims to develop and evaluate a novel segmentation model that achieves high precision and efficiency for pancreatic tumor delineation.</p> Methods <p>We propose an AttentionUNet-MSAF architecture that incorporates attention mechanisms and Multi-Scale Attention Fusion to enhance segmentation accuracy. The model was trained and tested on the Medical Segmentation Decathlon Pancreas dataset, which includes 281 annotated 3D CT volumes. Performance was compared against thirteen state-of-the-art methods using metrics such as Dice Similarity Coefficient, Hausdorff Distance (95%), Intersection over Union, and classification metrics including false positive and false negative rates.</p> Results <p>The proposed model outperformed all baselines, achieving the highest mean DSC (0.8735) and tumor-specific DSC (0.7789), the lowest mean Hausdorff Distance (1.3705), and the highest mean IoU (0.8115). It also demonstrated the best balance in classification outcomes with the lowest false positive rate (4.32%) and shared the lowest false negative rate (1.14%), indicating robust sensitivity and specificity. In addition, the model maintains computational efficiency with moderate parameter count and inference time, making it suitable for real-time clinical applications.</p> Conclusion <p>The AttentionUNet-MSAF model effectively addresses the challenges of pancreas and tumor segmentation, achieving state-of-the-art accuracy and boundary precision while maintaining efficiency. Its strong performance in detecting small pathological regions suggests its potential for aiding clinical diagnosis and treatment planning in pancreatic diseases.</p>