Author: Manju Liu, Ning Wen, Fuhua Yan, Yanzhao Yang, Zhenyu Yang, Haoran Zhang, Lei Zhang, Yajiao Zhang 👨🔬
Affiliation: Department of Radiology, Ruijin Hospital Shanghai Jiaotong University School of Medicine, Duke Kunshan University, Medical Physics Graduate Program, Duke Kunshan University 🌍
Purpose: Pancreatic ductal adenocarcinoma (PDAC) is a highly aggressive malignancy where precise segmentation of tumors and adjacent vessels is crucial for effective treatment planning. This study developed a deep learning model to automate segmentation of key structures, including the common hepatic artery (CHA), celiac artery (CA), superior mesenteric artery (SMA), superior mesenteric vein (SMV), portal vein (PV), and tumors, to improve surgical and radiotherapy planning.
Methods: This study included 1,108,000 arterial phase and 1,329,600 portal phase slices from 2,216 patients across four medical centers. Radiologists annotated 1,000 cases to train a hybrid model combining convolutional neural networks (CNNs) for spatial feature extraction and graph neural networks (GNNs) for vascular topology. To ensure vessel continuity, the model incorporated a connectivity map to refine segmentation. The resulting Teacher model generated pseudo-labels for the unannotated data, which, together with the original annotations, were used to train a Student model in a semi-supervised framework. Dice loss and Boundary loss were employed to address class imbalance and optimize segmentation performance. Model evaluation combined objective metrics on a test set with radiologists' five-point subjective assessments of clinical usability.
Results: The model achieved high Dice similarity coefficients: 0.88 for PV, 0.87 for CA, 0.83 for SMV, 0.82 for tumors in the portal phase, 0.80 for SMA, 0.78 for tumors in the arterial phase, and 0.70 for CHA. Subjective evaluations aligned closely, with PV (4.57) and CA (4.50) receiving the highest scores and lowest variability, while SMV (4.38) and tumors in the portal phase (4.32) also showed strong agreement. Lower scores for SMA (4.06), tumors in the arterial phase (3.87), and CHA (3.02) reflected their comparatively lower performance, highlighting areas for improvement.
Conclusion: The hybrid model achieves accurate segmentation of pancreatic tumors and major vessels, providing valuable support for surgical planning and radiotherapy precision.