Unsupervised Bayesian generation of synthetic CT from CBCT using patient-specific score-based prior

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Machine Learning in Healthcare

Background:

• Cone-beam computed tomography (CBCT) is crucial for image-guided radiotherapy (IGRT) but suffers from artifacts and inaccurate Hounsfield Units (HU).
• These limitations hinder quantitative applications like segmentation and dose calculation, essential for adaptive radiotherapy (ART).
• Improving CBCT image quality is vital for implementing online ART protocols.

Purpose of the Study:

• To develop an unsupervised, patient-specific diffusion model for generating synthetic CT (sCT) images from CBCT.
• The goal is to enhance CBCT image quality for improved quantitative analysis in radiotherapy.

Main Methods:

• An unsupervised framework utilizing a patient-specific score-based diffusion model as an image prior.
• Customized total variation (TV) regularization enforces slice coherence.
• The model is trained on a patient's planning CT (pCT) to learn unique anatomical features.

Main Results:

• Significant reduction in CBCT artifacts across head and neck, pancreatic, and lung cancer patient data.
• In lung SBRT, Mean Absolute Error (MAE) improved from 47 HU to 13 HU, Non-Uniformity (NU) from 45 HU to 14 dB, and SSIM from 0.58 to 0.67.
• The proposed method outperformed other unsupervised learning-based CBCT correction algorithms.

Conclusions:

• The unsupervised diffusion model effectively generates synthetic CT (sCT) from CBCT with reduced artifacts and accurate HU values.
• This enables precise CBCT-guided segmentation and replanning for online adaptive radiotherapy (ART).
• The approach holds promise for advancing quantitative imaging in clinical radiotherapy workflows.