Related Experiment Video
Updated: Aug 5, 2026
05:49
Reliability of Artificial Intelligence-Based Cone Beam Computed Tomography Integration with Digital Dental Images
Published on: February 23, 2024
Knowledge-distilled diffusion models for improving cone-beam CT image quality with meta-learning under imbalanced
Joonil Hwang1,2, Sangjoon Park1,2, Seungryong Cho3,4
1Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Seoul, South Korea.
Medical Physics
|July 31, 2026
Summary
This study introduces a novel framework using knowledge distillation and meta-guidance to improve synthetic CT image quality for adaptive radiation therapy (ART). The method enhances accuracy despite limited paired data, advancing ART workflows.
Area of Science:
- Medical Imaging
- Radiotherapy Physics
- Machine Learning in Healthcare
Background:
- Adaptive radiation therapy (ART) requires high-quality daily imaging, often limited by cone-beam CT (CBCT) image quality.
- Substantial anatomical changes during treatment further degrade CBCT accuracy, impacting dose calculations.
- A scarcity of paired planning CT (pCT) and CBCT data poses a challenge for developing accurate synthetic CT generation models.
Purpose of the Study:
- To develop a robust framework for generating high-quality synthetic CT (sCT) images from limited paired data.
- To overcome the challenge of abundant unpaired CBCT data and scarce paired pCT data for ART.
- To improve the accuracy of dose calculations in ART by enhancing daily image quality.
Main Methods:
- A novel framework combining knowledge distillation and gradient-based meta-guidance.
- Knowledge distillation leverages large datasets of unpaired CBCT images.
- Meta-guidance stabilizes training by dynamically weighting unpaired samples based on pseudo-label alignment with supervised gradients.
Main Results:
- Achieved superior performance with quantitative metrics: MAE 13.22 HU, SSIM 0.9516, PSNR 30.35 dB.
- Significantly outperformed supervised, unsupervised, and standard distillation baselines (p < 0.01).
- Ablation studies confirmed enhanced image quality and preservation of daily patient anatomy, minimizing feature hallucination.
Conclusions:
- The combined knowledge distillation and meta-guidance approach significantly advances high-quality synthetic CT generation.
- This method enables more robust and adaptive adaptive radiation therapy (ART) workflows.
- The framework effectively addresses the limitations of current CBCT imaging in ART.