S1586
Physics - Autosegmentation
ESTRO 2026
1 Department of Oncology, University Hospital & Faculty of Medicine, Ostrava, Czech Republic. 2 Department of Deputy director for science, University Hospital Ostrava, Ostrava, Czech Republic. 3 Department of Biomedical Engineering, Faculty of Electrical Engineering and Communication, Brno University of Technology, Brno, Czech Republic. 4 Department of Neurosurgery, University Hospital & Faculty of Medicine, Ostrava, Czech Republic Purpose/Objective: To describe our first clinical experience with a physician ‐ tailored deep ‐ learning model for automatic segmentation of brain metastases and to quantify its impact on SRS planning. Material/Methods: We trained a baseline nnU ‐ Net on institutional T2 ‐ weighted navigation MRI. The model was then tailored to one attending radiation oncologist according to prior contours. Second fine ‐ tuning targeted small metastases using a subset enriched for sub ‐ centimeter lesions. The tool was integrated via DICOM ‐ RT and evaluated prospectively in 15 consecutive patients (55 metastases). We recorded time to first auto-contour, manual contouring time, Dice similarity coefficients (DSC) for gross tumor volume (GTV) and planning target volume (PTV) vs predicted volume, lesion counts, and volume agreement. Margin for GTV-PTV was 0.3 mm isometrically. Results: Median time to first auto-contour was 1 min (IQR 1 –1; max 2). De ‐ novo manual contouring had a median duration of 23 min (IQR 13–31; range 11–150). Median DSC was 0.94 for GTV (IQR 0.89–0.95) and 0.95 for PTV (IQR 0.88–0.95). PTV DSC ≥0.90 occurred in 10/15 cases (67 %) and ≥0.95 in 8/15 (53 %).Volume agreement was high: the mean absolute volume difference was 0.49 cc for GTV (median 0.40 cc) and 0.73 cc for PTV (median 0.60 cc). The mean absolute percentage difference was 6.53 % for GTV and 11.61 % for PTV. The predicted lesion count matched or exceeded the reference in 14/15 patients (64 predicted vs 55 annotated), yielding additional candidates for review; one case was under ‐ predicted. The lowest PTV DSC (0.61) occurred in the smallest ‐ volume case (0.30 cc). The higher DSC agreement sometimes favored GTV and sometimes PTV. At given voxel scale, residual misalignment largely reflects the choice of boundary at the lesion edge (inner vs outer voxel tracing); the model’s output is stable, but the manual trace may shift with boundary contrast.
Conclusion: There is currently substantial heterogeneity in clinician-based auto-contouring assessment. Inconsistent scale use and ambiguous descriptors increase subjectivity, and concerns must be raised about the quality of auto-contouring assessments currently in use. Establishing a consensus in methodology should be a priority, in line with recommendations for standardised quality assurance procedures in international radiotherapy trials4. References: 1. Rong, Y., et al., NRG Oncology Assessment of Artificial Intelligence Deep Learning-Based Auto- segmentation for Radiation Therapy: Current Developments, Clinical Considerations, and Future Directions. Int J Radiat Oncol Biol Phys, 20232. Mackay, K., et al., A Review of the Metrics Used to Assess Auto-Contouring Systems in Radiotherapy.Clin Oncol (RCR), 20233. Aggarwal, A., et al., ARCHERY: a prospective observational study of artificial intelligence-based radiotherapy treatment planning for cervical, head and neck and prostate cancer - study protocol. BMJ Open, 2023. 4. Melidis, C., et al., Global harmonization of quality assurance naming conventions in radiation therapy clinical trials. Int J Radiat Oncol Biol Phys, 2014 Keywords: Qualitative assessment, auto-contouring evaluation Digital Poster 3647 Physician ‐ Tailored Auto ‐ Segmentation for Brain Metastases: First Clinical Experience Lukas Knybel 1 , Michal Nohel 2,3 , Jana Jackaninova 1 , Romana Kaplanova 2 , Stefan Reguli 4 , Jiri Chmelik 3 , Jakub Cvek 1
Made with FlippingBook - Share PDF online