nnU-Net-Based Pancreas Segmentation and Volume Measurement on CT Imaging in Patients with Pancreatic Cancer

Yang, Kim, Min, Jeong, Hwang, and colleagues developed a nnU-Net-based model to segment the pancreas and measure pancreatic volume from CT images in pancreatic cancer patients. The model was trained on 851 CT scans and validated against manual segmentations by expert radiologists. For the external test set, Aview software was used by radiologists to perform manual segmentation of the pancreas, establishing the reference standard to assess model performance. The nnU-Net achieved Dice similarity coefficients of 0.764 and 0.803 on internal and external test sets, respectively. Automatic volume measurements closely matched manual ones, showing minimal average differences. This model could enable non-invasive and efficient pancreatic volume assessments to support pancreatic cancer diagnosis and monitoring, potentially reducing manual workload. However, the study’s focus on pancreatic cancer alone and lack of data on non-cancer cases indicate a need for further research to enhance the model’s generalizability across various pancreatic conditions.