Few-shot CBCT-based synthetic CT generation with denoising diffusion probabilistic model.

Abstract

BACKGROUND: In adaptive cancer radiotherapy, treatment plans were periodically re-optimised according to anatomical variations, with the goal of maintaining or improving clinical outcomes. This required high-quality in-room imaging for daily monitoring, tumour/organ segmentation and dose evaluation. However, in-room cone-beam computed tomography (CBCT) images suffered from inaccurate Hounsfield Unit (HU) values and poor image quality due to higher artifact presence and lower signal-to-noise ratio. While CBCT-to-CT generation models have shown promising results, many conventional networks require large datasets for training. PURPOSE: To address this, we proposed a denoising diffusion probabilistic model (DDPM) with a modified sampling process to enable few-shot synthetic CT (sCT) generation from CBCT images. METHODS: We used a denoising diffusion probabilistic model trained on rigidly registered CT-CBCT image pairs in our approach, without requiring deformable registrations. Our model was trained on 25 patients from a head-and-neck cancer cohort at our institution, and validated and tested on separate sets of seven and eight patients. Our sampling method leveraged channel- and noised-conditioning, demonstrating high-quality sCT generation with limited training data. By adding noise to guiding CBCT images during sampling, we exploited the convergence of CBCT and planning CT (pCT) representations in latent space. This allowed sampling from a noisy representation of the data, achieving sCTs with high fidelity. RESULTS: Evaluation on a head-and-neck (H&N) cancer dataset showed that sCTs outperformed CBCTs, reducing masked mean absolute error (MAE) from 131 ± 17 HU to 49 ± 7 HU, and improving peak signal-to-noise ratio (PSNR) from 20.0 ± 0.9 dB to 22.9 ± 1.1 dB and normalised cross-correlation (NCC) from 0.93 ± 0.01 to 0.96 ± 0.01. We further evaluated the method's generalisability on phantoms and publicly available patient datasets. Without retraining, the model achieved moderate improvements over CBCT (e.g., masked MAE 87 ± 17 HU, PSNR 19.4 ± 1.7 dB, NCC 0.93 ± 0.04) on the external H&N dataset. However, after retraining on 27 in-distribution patient cases, the model achieved comparable performance to our internal dataset, with MAE 48 ± 12 HU, PSNR 22.2 ± 2.5 dB, and NCC 0.95 ± 0.03. We also extended the model to the pelvis site, achieving similarly strong results (masked MAE 44 ± 9 HU, PSNR 21.9 ± 2.1 dB, NCC 0.96 ± 0.03), demonstrating the feasibility of our method across multiple anatomical sites. Visual inspections showed improved image quality, significantly reduced artifacts, and better anatomical preservation in sCTs. Clinician evaluations revealed a higher preference to use sCTs over pCTs for fractional evaluation. CONCLUSIONS: Our work offered a practical, data-efficient and site-robust solution for clinics to generate high-fidelity sCTs, facilitating CBCT-based dose evaluation and plan adaptation in adaptive radiotherapy workflows.