S146
Brachytherapy - Physics
ESTRO 2026
independent modality for precise and effective skin cancer radiotherapy. Keywords: Brachytherapy, interventional radiotherapy, skin References: [1] E. Placidi et al., “Retrospective Dosimetric Comparison of HDR Interventional Radiotherapy (Brachytherapy) Versus Planning with VMAT and Electron Beam Therapy for Non-Melanoma Skin Cancer Treatment,” Applied Sciences (Switzerland), vol. 14, no. 22, Nov. 2024, doi: 10.3390/app142210669. Digital Poster Highlight 4596 Artificial Intelligence-Based Nomogram to Predict Achievable Dose Coverage in HDR Prostate Brachytherapy Joel Beaudry 1,2 , Tonghe Wang 1 , David Aramburu Nunez 1 , Shirin A Enger 2,3 , Antonio L Damato 1 1 Medical Physics, Memorial Sloan Kettering Cancer Center, New York, USA. 2 Medical Physics Unit, Department of Oncology, McGill University, Montreal, Canada. 3 Lady Davis Institute for Medical Research, Jewish General Hospital, Montreal, Canada Purpose/Objective: To develop and evaluate an artificial intelligence (AI)- based nomogram capable of predicting achievable post-implant dose coverage (V100%) for high-dose- rate (HDR) prostate brachytherapy. Conventional nomograms rely on empirical relationships between prostate volume and prescription dose coverage, often neglecting the spatial and geometric influence of dwell distributions. This study introduces a 3D convolutional neural network-based model that integrates patient-specific contour and dwell information to estimate achievable dose coverage, thereby supporting plan quality assurance (QA) and treatment standardization. Material/Methods: A total of 2,244 HDR prostate brachytherapy plans were retrospectively analyzed. For each case, 3D masks of the prostate, urethra, and rectum, along with corresponding dwell coordinate maps, were converted into voxelized model inputs. The target output was the binary dose distribution representing the prescribed V100% isodose. Data were partitioned into training, validation, and independent test subsets using a patient-level split (90%/10%). Model development used 5-fold cross-validation, and performance was evaluated on the held-out test set. An ensemble model was trained using a 3D nnUNet framework, combining Dice and binary cross-entropy losses with adaptive learning rate scheduling. Model performance was assessed using the Dice similarity coefficient (DSC), 99th percentile Hausdorff distance (HD99), average
consistency. All plans were generated independently at the validation center, keeping the same target delineations and prescription doses as in the Ir-192 study. Dosimetric metrics (V90%, V95%, V100%, D90%, Dmean, Dmax, and D2cc for the brain) were compared across modalities. Results: All techniques achieved V90%>99%, but Co-60 HDR-IRT provided the highest V100%, D90%, and Dmean values, consistent with the findings of the previous Ir-192 analysis. Differences in Dmean between HDR-IRT and VMAT/EBT remained statistically significant (p < 0.05), as shown in Figure 1. VMAT showed better dose homogeneity but lower target coverage, while EBT was feasible only for smaller, flatter lesions.
Figure 1. Dosimetric comparison of CTV coverage across three treatment modalities (Co- 60 IRT, VMAT and EBT) and prescription groups (A, B, and C): CTV Mean dose (a) and CTVMaximum dose (b). Brain D2cc doses were comparable among all modalities across the three prescription groups (Figure 2).
Figure 2. D2cc brain across three prescription groups (A–C) and techniques (Co-60 IRT, VMAT and EBT). Conclusion: This external validation confirmed the robustness and reproducibility of HDR-IRT in NMSC treatment. Even when replanned at a different center by different operators using a different radioactive source (Co-60 instead of Ir-192), HDR-IRT consistently outperformed VMAT and EBT. These results reinforce HDR-IRT as a reliable, source-independent, and operator-
Made with FlippingBook - Share PDF online