Meta-learning and knowledge distillation improve CBCT image quality via diffusion models

A knowledge-distilled diffusion model with meta-guidance improved cone-beam CT image quality, achieving MAE 13.22 HU, SSIM 0.9516, and PSNR 30.35 dB, significantly outperforming baselines (p<0.01).

• Evaluated on 99 breast cancer patients (19 reserved for testing) and an external public dataset.
• Knowledge distillation leveraged unpaired CBCTs, while meta-guidance dynamically weighted samples to reduce noise.
• Ablation studies showed improved image quality and preserved daily anatomy with minimal feature hallucination.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.