S1609
Physics - Autosegmentation
ESTRO 2026
Poster Discussion 4648 One-shot patient-specific fine-tuning of a non- specialized autosegmentation model provides high quality CTV delineations Nicole Ferreira Silverio, Joren Brunekreef, Alice Couwenberg, Olga Hamming-Vrieze, Jan-Jakob Sonke, Tomas Janssen Department of Radiation Oncology, The Netherlands Cancer Institute, Amsterdam, Netherlands Purpose/Objective: Autosegmentation of CTVs is challenging. Their definition relies on clinical judgment or patient-specific factors not explicitly visible in imaging data. In (online) adaptive radiotherapy, this patient-specific information is available during treatment as earlier delineations, which can be utilized to fine-tune a pre- trained model and improve autosegmentations. Previous work [1,2] focused on fine-tuning task- specific models, prompting the question whether non- specialized models or models trained on limited data can serve as a basis for patient-specific fine-tuning. This study aimed to determine data requirements for CTV autosegmentation models with patient-specific fine-tuning in (online) adaptive radiotherapy of MRI- based rectal cancer (RC) and CT-based head-and-neck cancer (HNC). Material/Methods: 3D nnU-Net [3] baseline models were trained to segment the RC CTV on a population of 5-10-30 patients, treated with MR-guided online adaptive radiotherapy. Per patient, one delineated MRI-scan (3D-T2) was included (80/20 train-validation split). The MRI-based total body TotalSegmentator model [4] served as the fourth baseline, without additional population-based training. Inference was performed on fractions 2-4 for 14 additional RC testpatients.This was repeated for CT-based delineation of the primary CTV of 5-30-1033 HNC patients. The CT-based total body TotalSegmentator model served as the fourth baseline. Twentyfive HNC patients with a replanning scan two weeks after treatment start served as test data, with inference performed on this second scan.Patient-specific fine-tuning was done on the test data, by training the baseline models until convergence on fraction 1 (RC) or planning CT (HNC). Predictions were compared to clinical delineations on the Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (95HD) and added path length (APL). All patients provided written informed consent. Results: Figure 1-2 demonstrate that the performance gains from increasing the baseline training dataset are less pronounced (RC: median DSC 0.82 for 5pts, 0.87 for 30pts, HNC: 0.12 for 5pts, 0.64 for 1033pts) than those
achieved through patient-specific fine-tuning (0.92 (RC) and 0.91 (HNC) using 5pts baseline model). All task- specific baseline models performed similarly after fine-tuning, regardless of the number of patients. TotalSegmentator performed comparable to the task- specific models after patient-specific fine-tuning (DSC 0.92 (RC) and 0.89 (HNC)).
Conclusion: Fine-tuning TotalSegmentator on a single scan of a patient resulted in a similar performance as fine- tuning a specialized model for CTV segmentation. These results hold for two vastly different applications: RC CTV on MRI and HNC CTV on CT. This suggests that high quality autosegmentation of targets in (online) adaptive radiotherapy can be performed without training specialized models. References:
Made with FlippingBook - Share PDF online