SinoSynth: A Physics-Based Domain Randomization Approach for Generalizable CBCT Image Enhancement

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Cone Beam Computed Tomography (CBCT) is vital in medicine, but image quality is often compromised by noise and artifacts.
• Existing artifact correction methods struggle with diverse degradations due to limited training data.

Purpose of the Study:

• To introduce SinoSynth, a novel physics-based model for generating diverse synthetic CBCT images with realistic artifacts.
• To improve the performance of deep learning models for CBCT artifact reduction.

Main Methods:

• Developed a physics-based degradation model (SinoSynth) to simulate CBCT-specific artifacts from high-quality CT images.
• Generated a diverse dataset of synthetic CBCT images without requiring pre-aligned data.
• Trained generative networks on synthesized data for artifact correction.

Main Results:

• Generative networks trained on SinoSynth data achieved superior performance on multi-institutional CBCT datasets compared to those trained on real data.
• The model successfully generated high-quality, structure-preserving synthetic images.
• Demonstrated the ability to enforce anatomical constraints in generative models.

Conclusions:

• SinoSynth offers an effective solution for data augmentation in CBCT artifact correction.
• Physics-based simulation enhances the robustness and generalizability of deep learning models for medical imaging.
• This approach addresses the challenge of limited and varied artifact data in CBCT.