ESTRO 2025 - Abstract Book

S3124

Physics - Inter-fraction motion management and offline adaptive radiotherapy

ESTRO 2025

2121

Digital Poster CBCT-based synthetic CT generation with diffusion models for adaptive radiotherapy Ping Lin Yeap 1,2 , Andrew Hoole 3 , Rajesh Jena 1,4

1 Department of Oncology, University of Cambridge, Cambridge, United Kingdom. 2 Division of Radiation Oncology, National Cancer Centre Singapore, Singapore, Singapore. 3 Department of Medical Physics, Addenbrooke's Hospital, Cambridge University Hospitals NHS Foundation Trust, Cambridge, United Kingdom. 4 Department of Oncology, Cambridge University Hospitals NHS Foundation Trust, Cambridge, United Kingdom Purpose/Objective: Cone-beam CT (CBCT)-based adaptive radiotherapy is limited by the challenge of inaccurate Hounsfield Unit (HU) values in CBCT images. Conventional methods to improve CBCT quality, such as Hounsfield Unit (HU) calibration and deformable image registration, have limitations, including reliance on operator skill and potential inaccuracies. To overcome these issues, the study explores the use of CBCT-guided denoising diffusion probabilistic models (DDPMs) and their performance with small patient training datasets. The study aims to generate high-fidelity synthetic CT (sCT) images, enabling effective CBCT-based adaptive radiotherapy workflows. Material/Methods: This study utilised a dataset of 40 head-and-neck cancer patients, with paired planning CT (pCT) and CBCT images, pre-processed through resampling, rigid registration, and intensity scaling. Using 25 of the patients for training, a 2D U-Net-based DDPM [1] was employed to generate sCT images. During training of DDPMs, Gaussian noise was iteratively added to degrade the data, and a neural network then learned to reverse this process akin to “de noising”, thereby recovering the data. To generate sCT images with the quality of the pCT yet preserve the anatomical geometry of the CBCT, the model incorporated CBCT guidance to condition the sCT generation process [2,3]. Results: The results demonstrated the utility of the DDPM in sCT generation. We observed that a conventional DDPM with a linear noise schedule performed poorly with our limited dataset, yielding masked mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and normalized cross-correlation (NCC) values of 436 ± 261 HU, 14.2 ± 5.1 dB, and 0.63 ± 0.23, respectively, compared to 130 ± 16 HU, 23.7 ± 1.5 dB, and 0.94 ± 0.02 for the original CBCT images. By incorporating appropriate guiding mechanisms, we achieved substantially improved metrics of 48 ± 5 HU, 26.5 ± 1.1 dB, and 0.97 ± 0.01. These findings highlight the effectiveness of choosing the right sampling strategy in generating high-fidelity sCT images, especially with a small training dataset. Figure 1: 4 example sCT slices generated with our model. The corresponding CBCT and pCT slices, as well as HU difference maps and histogram of CT numbers, were displayed.