Efficient and rapid generation of high-quality CT images from CBCT using a flow matching model.
BACKGROUND Generating high-quality CT images from cone-beam computed tomography (CBCT) is critical for clinical applications such as online adaptive radiotherapy (ART). Although deep learning methods, including diffusion models, have shown remarkable advances, persistent challenges remain in terms of image quality and computational efficiency. PURPOSE This study aims to leverage the flow matching model for CBCT-to-CT translation, enabling the rapid and efficient synthesis of high-quality CT images for clinical application. METHODS Our proposed method is built upon the flow matching model, a powerful multi-step generative framework that is designed to construct a continuous trajectory that transforms data from a prior distribution to a complex target distribution. A cohort of 135 abdominal-pelvic CT scans was used for model training (95 cases), validation (15 cases), and testing (25 cases). To ensure accurate evaluation, a dataset of rigorously paired CBCT-CT images was generated by simulating the CBCT imaging process. A comprehensive assessment of the model's performance was conducted with respect to structural consistency, detail fidelity, image blurring, and texture realism. The proposed method was benchmarked against the state-of-the-art probabilistic diffusion model. RESULTS As the number of sampling steps increases, the metrics reflecting image sharpness exhibit a consistent upward trend, indicating progressive improvement in clarity. An optimal balance among textual realism, detail fidelity, and computational efficiency was achieved within the 20-50 step range. At the 20 sampling steps, the proposed model can generate a 100-slice CT volume within 20 s, with a mean absolute error (MAE) of 12.25, a structure similarity index measure (SSIM) of 0.977, and a peak signal-to-noise ratio (PSNR) of 42.03. Compared to the conventional diffusion model, the proposed method delivers superior performance with fewer sampling steps, demonstrating its potential for efficient generation of high-quality medical images. CONCLUSIONS The proposed flow matching model enables efficient and rapid generation of high-quality synthetic CT images from CBCT, substantially outperforming existing diffusion models and demonstrating strong potential for time-sensitive clinical applications such as online ART.