S2082
Physics - Image acquisition and processing
ESTRO 2026
image quality of the predicted FBCTs using both pixel- wise metrics and auto-segmentation for organs-at-risk (OARs); and (3) Explore the feasibility of this approach for potential applications in adaptive radiotherapy planning. Material/Methods: Retrospective data from 466 NPC patients— comprising 8,880 FBCT scans acquired during radiotherapy between September 2021 and December 2024—were divided into training (298 patients, 64%), validation (75 patients, 16%), and test sets (93 patients, 20%). The textual data included TNM stage, age, sex, EBV DNA results, and time relative to the first radiotherapy session.Latent space features were extracted using a pre-trained 3D autoencoder (trained on the training set). Both training and inference of the diffusion model were conducted using these high- dimensional features. The 3D diffusion model training was structured in two stages: the first stage involved an upstream task where the backbone model synthesized CT images from textual features; the second stage introduced a downstream task, embedding a multi-time FBCT-guided control network into the backbone to generate future CT images.Performance evaluation was two-fold: first, image generation quality was assessed using Mean Absolute Error (MAE), Structural Similarity Index (SSIM), and Peak Signal-to-Noise Ratio (PSNR); second, an auto-segmentation model was applied to evaluate the quality of generated OARs using the Dice similarity coefficient and the 95th percentile Hausdorff Distance (HD95). Results: The vast majority of the current results are based on predicting CT images one day ahead on the test set. Table 1 presents a comparison of generation metrics. Built upon a high-performance autoencoder, the mean absolute error (MAE) of the predicted generated images has reached 9.5 Hounsfield Units (HU). Furthermore, the metrics for automatically segmented OARs on both the predicted and real images, as shown in Table 2, demonstrate the accuracy of the predicted generated images.
Conclusion: Developed a two-stage latent diffusion model to predict future FBCT images for NPC radiotherapy. The model achieved high accuracy (MAE: 9.5 HU) and generated images with clinically usable OAR auto- segmentation quality. This demonstrates significant potential for enabling proactive adaptive radiotherapy by forecasting anatomical changes. References: 1. Touvron H, et al. Llama 2: Open Foundation and Fine-Tuned Chat Models. *arXiv preprint* 2023; arXiv:2302.05543. Keywords: Temporal FBCT,Prediction, Nasopharyngeal carcinoma Mapping white matter tracts without diffusion MRI: an atlas-based tool for stereotactic brain radiotherapy Sophie Bockel 1 , Mohammed El Aichi 1 , Cristina Veres 1 , Camilla Satragno 1 , Cédric Yuste 1 , François Bidault 2 , Gabriel Garcia 2 , Frédéric Dhermain 1 , Eric Deutsch 1 , Charlotte Robert 1 1 Radiation Oncology, Gustave Roussy, Villejuif, France. 2 Radiology, Gustave Roussy, Villejuif, France Purpose/Objective: Preserving white matter tracts (WMTs) is essential to limit neurocognitive side effects in stereotactic brain radiotherapy (SBRT). However, diffusion MRI is rarely available in treatment workflows. Following our previous work, we validated TractoBrainRT, an open- source pipeline that integrates atlas-based WMTs, benchmarked against diffusion tractography and tested for registration robustness. Material/Methods: Each patient’s planning-CT and dose map were registered to the ICBM standardized space using a Proffered Paper 3890
Made with FlippingBook - Share PDF online