S2201
Physics - Intra-fraction motion management and real-time adaptive radiotherapy
ESTRO 2026
3 Trinity St. James’s Cancer Institute, Trinity College Dublin, Dublin, Ireland
mean HU offset was –5 HU. In dosimetry, average differences in PTV metrics (V100%, D98%, V95%, D50%) were below 1.2%, with a maximum deviation of 5.5%. 3D gamma analysis showed pass rates of 97.3% (2%/2 mm).
Purpose/Objective: Population-based PTV recipes (e.g. van-Herk) overlook large patient-to-patient variation in intrafraction motion. We developed and evaluated a Bayesian Neural Network (BNN) to predict an individualised intrafraction PTV margin (PTVintra) with a calibrated uncertainty (95% high-density interval, HDI). Unlike a standard neural network that returns one “best” set of weights/parameters, a BNN averages over many plausible networks (model uncertainty) and also models measurement uncertainty (data’s inherent variability). Material/Methods: Open-mask head-and-neck patients in the OPEN randomised trial (N=100; 30-35 fractions) had 3D surface-guided radiotherapy (SGRT) motion recorded during beam-on for each fraction. Using only the first N fractions (N ∈ {1,3,5,10,15,20,25,30}) we derived (i) pooled summaries of overall motion and (ii) fraction- to-fraction variability features. The BNN (Figure.1) was created with one shared 32-unit hidden layer; two heads for predicting the systematic( Σ ) and random components( σ ), expressing uncertainty in the weights and biases rather than fixing them. We centered the priors for the output heads on our institution’s population H&N intrafraction data as a safe population baseline, which the model could move if the data supported. The BNN model produces likely distributions for Σ and σ , resulting in patient specific margin estimates, and 95% HDIs including both model and data uncertainty. Evaluation used patient-grouped 10-fold cross-validation. Evaluation metrics were mean absolute error (MAE), calibration (predicted vs. observed with HDIs), and paired tests (Wilcoxon signed-rank for paired differences and Levene for variance).
Fig2. Dosimetric differences between sCT and deformed CT for PTV metrics and 3D gamma analysis (test cohort, n=13). Conclusion: Compared with recent literature [1–4], our approach great HU accuracy (18 HU). This model enables the generation of reliable sCTs from head and neck CBCTs, with potential integration into ART workflows at conventional accelerators and possible extension to other anatomical sites. References: [1] O’Hara E, et al. Assessment of CBCT-based synthetic CT generation accuracy for adaptive radiotherapy. J Appl Clin Med Phys. 2022;23(6):e13639.[2] Chen L, et al. Synthetic CT generation from CBCT images via unsupervised deep learning. Phys Med Biol. 2021;66(11):115015.[3] Pang G, et al. A physics-informed deep learning framework for synthetic CT generation from CBCT in adaptive radiotherapy. Med Phys. 2023;50(2):857–869.[4] Altalib A, McGregor S, Li C, Perelli A. Synthetic CT image generation from CBCT: a systematic review. IEEE Trans Radiat Plasma Med Sci. 2025;9(6):691–707. doi:10.1109/TRPMS.2025.3533749 Keywords: Synthetic CT, Adaptive Radiotherapy, Deep Learning Bayesian neural network for personalised, uncertainty-aware intrafraction PTV margin adaptation in open-facemask head-and-neck radiotherapy using SGRT Ciaran Malone 1,2 , Jill Nicholson 2,3 , Claire Fitzpatrick 2 , Sinead Brennan 2,3 , Frances Duane 2,3 , Orla McArdle 2 , Ruth Woods 2 , Pat McCavana 2 , Gerard G Hanna 2,3 , Ben Heijmen 1 , Brendan McClean 2 1 Department of Radiotherapy, Erasmus MC Cancer Institute, University Medical Center Rotterdam, Rotterdam, Netherlands. 2 Radiation Oncology, St. Luke’s Radiation Oncology Network, Dublin, Ireland. Proffered Paper 368
Results: Patient-specific ground-truth PTVintra spanned 0.3-2.7 mm. At N=1, predictions conservatively reduced to a cohort baseline population margin (~1mm) with uniformly wide HDIs (showing appropriate behaviour under sparse evidence). As information accrued (N=3- >5->10), predictions individualised, scatter tightened around y=x, and HDIs narrowed in the common 0.5- 1.5mm range while remaining wider for rare large
Made with FlippingBook - Share PDF online