ESTRO 2025 - Abstract Book

S3369

Physics - Machine learning models and clinical applications

ESTRO 2025

The framework generated high-quality SDs for every RWD, with the TVAE producing the five best performing SDs considering all metrics with an equal weight for clinical behaviour, privacy and feature distribution. Regarding data privacy, no more than 4% of rows overlapped between each synthetic and real-world dataset. Regarding clinical behavior, cox proportional hazards (CPH) models for the real-world and synthetic datasets achieved similar concordance indexes (average for real-world C-indexes vs for SD C-indexes: 0.701 vs 0.699), with every SD hazard ratio falling within the 95% confidence intervals of their real-world counterparts for 4 of the 5 RWCDs. Regarding feature distribution, the nearest neighbor adversarial accuracy was in average 0.47 for all datasets. In the attached Figure the framework including its most important steps are presented on the left part. On the right part, results of an exemplary dataset (Glioma) are presented by showing the differences between the original and the synthetic dataset regarding CPH modelling and Kaplan-Meier plot. Conclusion: The proposed framework enables the production, evaluation and selection of SDs that closely mirror RWCD characteristics, ensuring privacy and clinical utility in RO. This approach can facilitate data sharing in clinical research, addressing privacy-related barriers. Digital Poster A deep learning pipeline for real-time conformal palliative radiotherapy of spine metastases Mehan Haidari 1 , Dal Granville 2 , Elsayed Ali 1 1 Medical Physics, The Ottawa Hospital, Ottawa, Canada. 2 Medical Physics, Nova Scotia Health Authority, Halifax, Canada Purpose/Objective: To develop a deep learning pipeline towards a CT simulation(CTsim)-free conformal palliative radiotherapy workflow for spinal metastases for the entire spine. The pipeline consists of a novel combination of a two-stage network for streak artifact reduction and synthetic CT (sCT) generation from Cone Beam CT (CBCT) images, followed by a three-stage vertebral segmentation network to sCT images. The goal is to reduce barriers to palliative radiotherapy 1 . Material/Methods: First, a dataset of CTSim and CBCT images from 220 patients, spanning the entire spine, was used to train and validate a two-stage generative adversarial network-based pipeline for sCT generation after streak artefact reduction 2-6 . Image quality was evaluated using a distinct dataset of 33 patients undergoing same-day treatment, with dosimetric analysis conducted on a subset. Next, a three-stage U-Net-based network 7-9 was trained and validated for vertebral segmentation using an open-source dataset 10-12 and was used to generate pseudo-ground truth vertebrae labeling on a local CTsim test dataset to evaluate performance on sCT images. Results: For the first part of the pipeline, the two-stage network significantly improved the Hounsfield Unit (HU) accuracy of CBCT images, reducing the Mean Absolute Error (MAE) from 225±62 HU in CBCT to 86±24 HU in sCT images, improving the Mean Error (ME) from 178±91 HU to -8±20 HU, improving the Structural Similarity Index Measure (SSIM) from 0.73±0.10 to 0.86±0.06, and improving the Peak Signal-to-Noise Ratio (PSNR) from 22±2 dB to 26±2 dB. Mean dose discrepancy was lowered by an average of 4.5%, with a 22% improvement in average gamma pass rate. For the second part of the pipeline, the segmentation network achieved an Identification Rate of 93±11%, Mean Distance from Centroid of 1.8±0.7 mm, Hausdorff Distance of 10.5±4.4 mm, and DICE Similarity Coefficient of Keywords: synthetic data, machine-learning 759