Knowledge-distilled diffusion models for improving cone-beam CT image quality with meta-learning under imbalanced...

Adaptive radiation therapy (ART) relies on daily cone-beam CT (CBCT), yet its limited image quality hinders accurate dose calculation, particularly under substantial anatomical changes.

To overcome the clinical challenge of scarce paired planning CT (pCT) data versus abundant unpaired CBCTs, we propose a framework driven by two core components: knowledge distillation and gradient-based meta-guidance.

The knowledge distillation strategy enables the model to leverage the vast unpaired dataset. Complementing this, the meta-guidance mechanism stabilizes training by dynamically updating the weight of each unpaired sample; it assigns higher importance to pseudo-labels that align with trusted supervised gradients, effectively filtering out noise. Our method was evaluated on a cohort of 99 breast cancer patients (with 19 reserved for testing) and further evaluated on a public dataset to assess the generalization capability.

The proposed approach demonstrated superior performance, achieving the best quantitative metrics (MAE 13.22 HU, SSIM 0.9516, PSNR 30.35 dB). It significantly outperformed representative supervised, unsupervised, and standard distillation baselines ( $p<0.01$ ). Ablation studies confirm that our method enhances image quality while preserving the patient's daily anatomy by minimizing feature hallucination.

By uniquely combining knowledge distillation with meta-guidance, our method advances the frontier of high-quality synthetic CT, enabling more robust and adaptive ART workflows.