Adaptive radiation therapy (ART) relies on daily imaging, commonly cone-beam CT (CBCT), to capture anatomy changes and update plans. However, CBCT image quality is limited compared with planning CT (pCT), which can hinder accurate dose calculation when patients undergo substantial anatomical change. Improving CBCT-derived synthetic CT remains a practical barrier to more robust ART workflows.
The study proposes a training framework for diffusion-based synthetic CT generation that addresses a common clinical data imbalance: relatively scarce paired pCT–CBCT examples versus abundant unpaired CBCT scans. The framework centers on two complementary components: knowledge distillation to exploit large unpaired datasets, and gradient-based meta-guidance to stabilize and purify training by weighting unpaired samples based on alignment with trusted supervised gradients.
Knowledge distillation in this work refers to leveraging supervision obtained from paired pCT–CBCT training examples to guide learning on the much larger set of unpaired CBCTs. The distillation component enables the model to benefit from the broad distribution represented by unpaired CBCT scans while still being anchored to the reliable labels available in the paired subset. The abstract emphasizes that this strategy allows the model to use a vast unpaired dataset without discarding the value of scarce paired supervision.
To complement distillation, the authors introduce a gradient-based meta-guidance mechanism that dynamically updates an importance weight for each unpaired sample during training. The mechanism assigns higher weights to pseudo-labels whose gradients are consistent with gradients produced by trusted supervised examples. In doing so, meta-guidance effectively filters out noisy pseudo-labels from unpaired data and stabilizes training, reducing the risk that noisy or misaligned unpaired examples will degrade the learned mapping from CBCT to synthetic CT.
The method was evaluated on a cohort of 99 breast cancer patients; 19 patients were reserved for testing. The authors additionally evaluated the approach on a public dataset to assess generalization. The abstract specifies a focus on application in breast cancer and notes that the external public dataset was used to evaluate how well the approach generalizes beyond the institution cohort.
The proposed framework achieved the best quantitative performance reported in the abstract for CBCT-to-sCT generation: mean absolute error (MAE) 13.22 Hounsfield units (HU), structural similarity index measure (SSIM) 0.9516, and peak signal-to-noise ratio (PSNR) 30.35 dB. The approach significantly outperformed representative supervised, unsupervised, and standard distillation baselines, with statistical significance reported as p < 0.01 in comparisons described in the source.
Ablation experiments reported in the abstract indicate that the combined knowledge-distillation and meta-guidance design contributes meaningfully to improved image quality. The authors note that their method preserved the patient's daily anatomy and minimized feature hallucination compared with alternatives, implying better anatomical fidelity and fewer artificial features introduced by the model.
By uniquely combining knowledge distillation with gradient-based meta-guidance, the authors report an advance in generation of high-quality synthetic CT from CBCT under imbalanced paired/unpaired data conditions. The improved image quality and reduced hallucination suggest potential to enable more robust and accurate ART workflows, where reliable synthetic CTs can better support dose calculation and adaptive planning.
The abstract reports cohort size, evaluation metrics, statistical significance, and the high-level design of knowledge distillation plus meta-guidance. Specific implementation details such as the exact diffusion model architecture, training hyperparameters, pseudo-labeling procedure, computational cost, runtime, and the precise identity of the public dataset(s) used were not reported in the abstract and therefore are not included here. For those details, readers should consult the full text of the cited Med Phys article (DOI: 10.1002/mp.70530) if available.