AI
New AI model converts low-quality cone-beam CT scans into clinical-grade CT images for radiotherapy
Researchers have developed a diffusion-based deep learning system that synthesizes high-fidelity CT images from cone-beam CT scans, improving image accuracy for radiation treatment planning.
Cone-beam CT (CBCT) scanners are a staple of modern radiotherapy facilities because they provide real-time imaging at low cost and radiation dose. However, CBCT images suffer from artifacts caused by photon scattering and beam obstruction, resulting in inaccurate measurements of tissue density (Hounsfield Units) that undermine the reliability of dose calculations in adaptive radiation treatment planning.
Standard CT imaging, by contrast, delivers superior image quality and precise calibration—but it's typically acquired offline and cannot capture anatomical changes during treatment. A team of researchers has now proposed a machine learning solution to bridge this gap.
How the new system works#
The researchers introduced EqDiff-CT, a diffusion-based generative model designed to synthesize high-quality CT images directly from CBCT scans. The system uses a denoising diffusion probabilistic model (DDPM) to iteratively add and remove noise while learning the underlying patterns needed to reconstruct anatomically faithful CT images.
A key innovation is the use of a group equivariant U-Net backbone with steerable convolutions (e2cnn layers). This architectural choice enforces rotational equivariance under C4 cyclic symmetry, meaning the model learns to recognize anatomical patterns regardless of orientation—a property that helps preserve fine structural details while suppressing noise and imaging artifacts.
Validation and performance#
The team trained and tested EqDiff-CT on the SynthRAD2025 dataset, which includes paired CBCT and CT scans from multiple head and neck anatomical sites. When compared with established methods such as CycleGAN and standard DDPM, EqDiff-CT demonstrated measurable improvements across multiple metrics:
- Enhanced structural fidelity and anatomical accuracy
- Better Hounsfield Unit accuracy for dose calculations
- Sharper soft tissue boundaries
- More realistic bone reconstructions
Why it matters#
The results suggest that diffusion models offer a robust and generalizable framework for improving CBCT image quality in clinical workflows. By reducing the imaging quality gap between CBCT and CT, the approach could enhance the reliability of adaptive radiotherapy—allowing clinicians to use real-time, in-treatment imaging for better treatment planning without sacrificing accuracy.
:
