S1566
Physics - Autosegmentation
ESTRO 2026
Purpose/Objective: Pancreatic cancer (PC) has a 5-year survival rate of 3– 8% [1]. Online adaptive MRI-guided radiotherapy represents a promising treatment modality for PC through improved soft-tissue visualization. However, tumour delineation is labour-intensive and remains challenging due to poor image quality and low inter-observer agreement (generalized conformity index of 0.36 for GTV delineations) [2]. The aim of this study was to evaluate the performance of AI-based GTV segmentation of pancreatic tumours on MRI scans, as this is a substantial knowledge gap. Material/Methods: Segmentation models were developed using nnU-Net [3] on T2W MRI from 82 patients with locally advanced PC treated on a 1.5 T MRI-Linac (50Gy/5fx). T2W scans and pancreas delineations, converted to binary masks, served as input images with GTVs contoured by experienced clinicians as ground truth labels. Scans were acquired pre-treatment on a 1.5 T MRI scanner and at each treatment fractions on the MRI-Linac. Four nnU-Net models were trained using different combinations of MRI scans, MRI-Linac scans, and pancreas masks to assess which inputs best support accurate segmentation. All models were full-resolution 3D U-Nets with 5-fold cross-validation, default pre- processing, standard augmentation, and Dice–cross- entropy loss. Performance was evaluated on both MRI and MRI-Linac scans. Table 1 summarises train/test split and cohort characteristics. Table 1: Cohort characteristics for training and test sets. PatientsScans per patientMedian GTV (cm3)GTV range (cm3)Training72*1 MRI + 5 MRI-Linac28.221.05- 158.45Test101 MRI + 5 MRI-Linac39.856.70- 127.28*During 5-fold cross-validation, each fold contains 57 training and 15 validation patients. Results: Models trained on a single scan type showed reduced performance compared to models trained on both (see Figure 1 and 2). The MRI-only model, evaluated on MRI scans, achieved a median dice similarity coefficient (DSCmedian) (25th-75th percentiles) of 0.16 (0.00-0.49), compared to 0.47 (0.06-0.64) for the combined-scan model. Similarly, the MRI-Linac-only model, evaluated on MRI-Linac scans, achieved a DSCmedian of 0.37 (0.00-0.64) vs. 0.42 (0.04-0.60). These results indicate including both scan types improves model generalization. Incorporating pancreas masks further improved performance for both MRI segmentation with DSCmedian of 0.54 (0.44- 0.72) and MRI-Linac segmentation with DSCmedian of 0.45 (0.27-0.58), showing how providing additional classes of related organs can improve GTV segmentations.
Conclusion: Exposure to both MRI and MRI-Linac scans improved model generalization, and incorporating pancreas masks further enhanced performance. Although not yet sufficient for clinical use, these findings represent a step toward more consistent and efficient adaptive RT for PC. References: [1] Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2021;71:209–249.[2] Gurney- Champion, Oliver J et al. “Addition of MRI for CT-based pancreatic tumor delineation: a feasibility study.” Acta oncologica (Stockholm, Sweden) vol. 56,7 (2017): 923- 930. doi:10.1080/0284186X.2017.1304654[3] Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2), 203-211.
Made with FlippingBook - Share PDF online