S2000
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
simple non-modulated treatments may seem straightforward; however, extending such estimates to VMAT irradiations remains challenging and typically requires Monte Carlo simulations or dense measurements—both too slow and cumbersome for routine planning. We propose a deep-learning method that extrapolates the in-field dose distribution into out-of-field regions, balancing fidelity and speed to enable seamless integration into clinical workflows. Material/Methods: From the French Childhood Cancer Survivor Study (FCCSS) dataset, we selected 1,579 photon-based treatment cases delivered using linear accelerators split into training (70%), validation (15%) and test (15%) sets. The FCCSS cohort contains 7,205 pediatric radiotherapy records, each with a whole-body 3D dose map reconstructed retrospectively at 2 × 2 × 2mm ³ resolution using an analytical model[1]. We evaluated several encoder–decoder architectures for reconstructing the normalized out-of-field (OOF) dose from the in-field dose distribution and a binary whole- body mask. We first refined a 3D U-Net baseline from a previous proof-of-concept study [2], and then introduced Masked Instance Normalization to reduce the large inter-patient anatomical variability characteristic of pediatric data. Building on these steps, we ultimately selected the Swin-UNETR as our proposed model, as its shifted-window attention mechanism captures long-range spatial dependencies more effectively than convolution-only encoders. To isolate the effect of the encoder/backbone, all transformer variants used the same decoder as the U- Net models. Because lower-dose regions span several orders of magnitude, we used a log-MSE loss, which makes the gradient sensitive to relative errors. To stabilize training near zero values, we applied a softplus transformation before taking the logarithm. Model performance was assessed using the Mean- Absolute-Error and the Root mean-square deviation computed exclusively on the OOF region. Results: Using the Swin-UNETR, we improved on our baseline models, achieving a MAE of 0.08 (0.05) , 0.13 (0.11) , and 0.14 (0.13) on train/val/test, with RMSE of 0.11 (0.07) , 0.19 (0.14), and 0.20 (0.16) Visual comparison also shows that the Swin-UNETR closely reproduces the ground-truth out-of-field dose distribution, as illustrated in Fig.1.
Conclusion: Transformer-based architectures show strong potential for out-of-field dose reconstruction; future work will validate our model using anthropomorphic and clinical measurements and explore implementation of distance-aware patch selection. References: [1] Veres C, et al. Retrospective reconstructions of active bone marrow dose-volume histograms. Int J Radiat Oncol Biol Phys. 2014;90(5):1216–24. [2] Benzazon N, et al. Deep-Learning for Rapid Estimation of the Out-of-Field Dose in External Beam Photon Radiation Therapy – A Proof of Concept. Int J Radiat Oncol Biol Phys. 2024;120(1):253–264. doi:10.1016/j.ijrobp.2024.03.007 [3] Hatamizadeh A, Nath V, Tang Y, Yang D, Roth HR, Xu D. Swin UNETR: Swin Transformers for Semantic Segmentation of Brain Tumors in MRI Images. In: Crimi A, Bakas S (eds) BrainLes 2021. Lecture Notes in Computer Science, vol 12962. Springer, Cham. https://doi.org/10.1007/978-3- 031-08999-2_22 Keywords: Radiation-induced cytotoxicity, Deep Learning Dosimetric Impact of Bolus Placement Timing in Post-Mastectomy Breast Radiotherapy: A 30- Patient Study. Md Abdul Mannan 1 , Ahammad Al Mamun Sweet 1 , Md. Ruhul Amin Bhuiyan 2 , Md Samiul Alim 3 , MDJobairul Islam 1 , Priyanka Poddar 3 1 Clinical Oncology, Labaid Cancer Hospital and Super Specialty, Dhaka, Bangladesh. 2 Clinical Oncology, North East Cancer Centre, Shylet, Bangladesh. 3 Radiotherapy, Institute of Nuclear Medical Physics, Dhaka, Bangladesh Purpose/Objective: In post-mastectomy breast radiotherapy, accurate dose delivery to the chest wall while controlling skin toxicity is critical. Bolus is commonly used to increase surface dose, but discrepancies may occur if the bolus is applied only during treatment and not included during CT simulation. This study evaluates the Digital Poster 4491
Made with FlippingBook - Share PDF online