KEY POINTS
- This narrative review examined deep-learning methods for generating synthetic CT (sCT) from treatment-fraction cone-beam CT for photon and proton adaptive radiotherapy. A structured PubMed search, last updated 30 July 2026, was supplemented by reference screening and targeted searches. The authors classified approaches into four overlapping methodological categories: supervised regression, adversarial generation, registration-aware methods, and diffusion-based generation. Unlike a systematic review or meta-analysis, the study did not perform formal risk-of-bias assessment, quantitative synthesis, or pooled estimates of model performance.
- Different AI architectures introduce distinct trade-offs between image quality, anatomical accuracy, and computational efficiency. Supervised models, including U-Net and related convolutional networks, provide efficient inference but depend on well-aligned paired CBCT–CT data and may produce overly smoothed images. Pix2Pix and CycleGAN offer improved texture reconstruction and support paired or unpaired translation, respectively, but can introduce anatomically incorrect or hallucinated structures. RegGAN-type approaches explicitly address residual image-registration errors, whereas conditional diffusion models may improve local detail at the expense of repeated sampling and greater computational requirements. No method consistently outperformed all alternatives across sites and imaging conditions.
- External generalizability remains insufficiently demonstrated despite encouraging internal results. One RegGAN study evaluated 140 head and neck patients, including 100 training and 40 test cases, using CBCT from four different linear accelerators within the same institution; this provides internal cross-device evidence, not independent multicenter validation. In contrast, the SynthRAD2023 benchmark included 270 brain and 270 pelvic cases from three Dutch centers, enabling more independent comparisons across submitted approaches. However, even multicenter benchmark performance does not establish reliable operation in a department's actual clinical workflow.
- Global image similarity metrics can conceal clinically important anatomical errors. Common endpoints such as mean absolute HU error, PSNR, SSIM, and normalized cross-correlation assess image-level agreement but may fail to detect localized inaccuracies in bone, air cavities, soft-tissue interfaces, target boundaries, and critical OARs. The review recommends combining regional HU analysis with Dice similarity, HD95, surface distances, expert assessment, and case-level failure detection. Interpretation also depends on the reference image: same-day repeat CT, registered planning CT, and deformably generated reference CT have different anatomical and methodological limitations and should not be treated as interchangeable ground truth.
- Dosimetric validation must be tailored to the treatment modality rather than inferred from image quality alone. For photon adaptive RT, recommended evaluation includes full dose-distribution comparison, DVHs, target D95/D98/D2, and clinically relevant OAR endpoints. Published studies have used gamma criteria such as 3%/3 mm and 2%/2 mm, but the authors caution that these are study-specific comparison settings, not universal sCT acceptance thresholds. For proton therapy, verification must additionally address relative stopping power, water-equivalent thickness, beam-path composition, and proton range errors, particularly near bone–air interfaces and regions with changing tissue density.
- Fast image generation does not necessarily translate into an efficient clinical adaptive workflow. In one reported comparison, diffusion-based synthesis required approximately 0.217 seconds per image, compared with 0.007–0.009 seconds for GAN-based approaches on the evaluated hardware. Another preclinical thoracic-phantom workflow generated sCT in approximately 15 seconds, but required 6–8 minutes for the complete adaptive process. The review emphasizes that preprocessing, registration, contour propagation, dose calculation, quality assurance, human review, and recovery from failures must all be included when evaluating operational feasibility; inference times reported per slice or volume cannot be directly compared across different studies.
- The authors propose a two-axis framework for evaluating clinical readiness. The first axis distinguishes five levels of validation maturity: proof of concept, internal validation, independent external validation, preclinical task validation, and prospective workflow deployment. The second evaluates four clinical reliability dimensions: data transferability, structural reliability, dosimetric validity, and deployment feasibility. Before clinical adoption, the review recommends locked-model testing across independent institutions, anatomical-site-specific dosimetry, calibrated uncertainty and failure detection, and prospective shadow-mode testing with predefined acceptance criteria, human override, and fallback procedures. The framework is an evaluation proposal, not a validated clinical acceptance standard.
CLINICAL TAKEAWAY
CBCT-derived synthetic CT has demonstrated encouraging technical performance, but accurate-looking images alone cannot establish the reliability required for adaptive dose calculation, particularly in proton therapy where local density errors may affect particle range. Clinical implementation should depend on independent anatomical and dosimetric validation, patient-level failure detection, and prospective end-to-end workflow assessment rather than global image metrics or inference speed alone.