S2185
Physics - Inter-fraction motion management and daily adaptive radiotherapy
ESTRO 2026
planning CT (pCT) and an average of 6 weekly treatment verification CTs (vCTs). Twenty-four patients required plan adaptation during treatment.For each patient, the pCT was augmented by 80 random treatment isocenter shifts up to ±15mm to simulate setup variations. A PS model was generated via transfer learning using dose distributions from the original and shifted pCTs at the time of planning in ~60 minutes. Subsequently, weekly vCTs were similarly augmented with 16 random isocenter shifts (±7mm), and the corresponding dose distributions were used as model inputs to predict treatment isocenter shifts at the beginning of each treatment session. The standard deviation of predicted isocenter positions on each vCT was computed, and its relative difference from the pCT standard deviation ( Δ SD) was used as an adaptation trigger. Values exceeding the empirically defined 60% threshold suggested a potential need for plan adaptation.The predictive power of Δ SD of predicted isocenter positions was validated by comparing primary and elective clinical target volume (CTV) V95 values on all vCTs against the clinical goal (V95>98%). Results: The PS model was sensitive to dose variations throughout treatment course. All adaptive patients were correctly identified using the Δ SD metric (sensitivity=1.0), while six non-adaptive patients were incorrectly flagged as adaptive (specificity=0.84); three of these had temporary anatomical changes that resolved without replanning, while the other three had clinically acceptable target coverage deviations. Δ SD increased markedly in adaptive patients (mean=127.2%, 95%-CI:98.2-156.2%) but remained stable in non-adaptive ones (mean=7.0%, 95%-CI:2.6- 11.5%) (Figure 1). Figure 2 illustrates Δ SD and V95 evolution for representative adaptive and non- adaptive patients. Model inference per vCT completed within 1.5 minutes, supporting potential clinical use.
Conclusion: The proposed automated patient-specific framework enables fast triggering of treatment adaptation by monitoring changes in the standard deviation of predicted treatment isocenter positions, without requiring target re-segmentation. Future work will extend this framework to adaptive photon therapy. References: [1] H. Bahrdo, G.G. Marmitt, F. Oosterhof, J.A. Langendijk, S. Both, 3978 A Deep Learning Model for Fast and Accurate Oropharynx Patients’ Pretreatment Positioning Prediction in Proton Therapy, Radiother. Oncol.206 (Suppl. 1), S3440–S3442 (2025). https://doi.org/10.1016/S0167-8140(25)03050-6 Keywords: Adaptation Trigger, Patient Specific Deep Learning
Made with FlippingBook - Share PDF online