S1853
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
multiple metastases. Med Phys 49:4305-21, 2022. doi:10.1002/mp.15689[2] Mackeprang PH et al. Randomized safety study of tolerance to table rotation in dynamic trajectory radiotherapy in healthy volunteers. Phys Imaging Radiat Oncol 35:100796, 2025. doi:10.1016/j.phro.2025.100796[3] Gianna C et al. Thresholds for detection of motion direction during passive lateral whole-body acceleration in normal subjects and patients with bilateral labyrinthine loss. Brain Res Bull 40:443-7, 1996. doi:10.1016/0361- 9230(96)00140-2 Keywords: Patient Comfort, Couch Motion, Dynamic Trajectory Enabling Safe Patient-Specific Dose Prediction through Integrated Uncertainty Estimation in Online Adaptive Radiation Therapy Benjamin Roberfroid 1,2 , Andréanne Lapointe 3 , Houda Bahig 3 , Hugo Bouchard 1,2 , Arthur Lalonde 1,2 1 Department of Physics, Université de Montréal, Montreal, Canada. 2 Axe imagerie et ingénieurie, Centre de Recherche du Centre Hospitalier de l'Université de Montréal, Montreal, Canada. 3 Département de radio- oncologie, Centre Hospitalier de l'Université de Montréal, Montreal, Canada Proffered Paper 1667 Purpose/Objective: Deep learning-based dose prediction (DP) models rapidly and accurately predict radiotherapy dose distributions [1]. In adaptive radiotherapy (ART), where time constraints are critical, patient-specific (PS) approaches that fine-tune models on individual patients yield highly tailored predictions. However, clinical adoption remains limited by model opacity and inability to assess prediction reliability. This study proposes a novel PS strategy combining intentional deep overfitting (IDOL) [2] with simultaneous CT reconstruction (IDOL_CTr) [3]. This dual-task approach enables uncertainty estimation through reconstruction error, providing confidence measures essential for safe ART implementation. We evaluated this method against a general DP model and two state-of-the-art PS strategies. Material/Methods: All methods were implemented using the HD-UNet architecture [1] and a dataset of 81 head and neck cancer patients with repeated CTs and two prescribed dose levels (70 and 59.4Gy). A general DP model was trained on 55 patients (300 epochs), validated on 5, and tested on 21 patients. IDOL_CTr builds upon IDOL [2], which fine-tunes a general model on a patient's initial treatment plan. IDOL_CTr extends this with dual- decoder architecture for simultaneous dose prediction and CT reconstruction. Comparisons included:
that may be relevant to patient comfort. Material/Methods:
Six previously treated VMAT plans were re-optimized to create two dynamic trajectory plan types: dynamic VMAT and dynamic iABC1. Dynamic VMAT (one gantry arc and 4–6 couch arcs) used standard Varian optimization, while dynamic iABC (one gantry arc and two couch arcs) employed intra-arc binary collimation to enable variable-speed couch motion. Both dynamic plans incorporated dynamic gantry, couch, and collimator rotations. Gantry arcs enabled bidirectional couch motion, while couch arcs required continuous unidirectional couch motion. All plans were delivered on a Varian TrueBeam STx accelerator. Couch angular velocity was recorded using an Xsens MTw Awinda accelerometer, and angular acceleration and jerk were calculated from the data. Dynamic couch motion was quantified using root-mean-square, time spent at different velocity/acceleration/jerk ranges, and weighted peak rate. Results: Angular couch velocity, acceleration, and jerk ranged from 0.5–2.5 °/s, 0.5–6 °/s ² , and 0.5–50 °/s ³ , respectively. Dynamic iABC exhibited more frequent couch motion fluctuations than dynamic VMAT. Both dynamic techniques operated mainly at low couch speeds (80% of treatment time spent below 1.5 °/s) during gantry arcs, while greater speed variability was observed during couch arcs. Both techniques spent over 77% of treatment time below 3 °/s ² , remaining under the vestibular perception threshold3 (5–10 °/s ² ). However, their motion styles differed. During gantry arcs, both dynamic techniques exhibited similar couch motion characteristics, with dynamic iABC spending 2% more time in the highest acceleration (5–6 °/s ² ) and jerk (20–50 °/s ³ ) ranges, suggesting a comparable patient experience during gantry arcs. During couch arcs, dynamic iABC moved more slowly on average, but frequent speed fluctuations increased overall acceleration and jerk. However, these fluctuations were moderate and never exceeded 3 °/s ² . In contrast, dynamic VMAT maintained smoother, more gradual motion but spent 23% more time above 3 °/s ² . These motion style differences suggest potential variations in patient comfort between techniques. Conclusion: While dynamic iABC improves treatment quality and efficiency with fewer arcs1, our findings suggest it may introduce more frequent fluctuations and higher angular jerk during couch arcs, potentially affecting patient comfort. This work highlights the need to consider couch kinematics in future optimization to better balance dynamic trajectory-based planning with patient comfort. References: [1] Lee E et al. Intra-arc binary collimation with dynamic axes trajectory optimization for SRS of
Made with FlippingBook - Share PDF online