Multiscale Segmentation-Guided Diffusion Model for CBCT-to-CT Synthesis.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Synthetic CT (sCT) generation from Cone-Beam CT (CBCT) is crucial for adaptive radiotherapy.
• Current methods often struggle with anatomical accuracy and image fidelity.
• Improving sCT quality can enhance treatment planning and delivery.

Purpose of the Study:

• To propose a novel multiscale segmentation-guided diffusion framework for improved sCT generation.
• To integrate anatomical priors across multiple resolutions for enhanced image quality.
• To evaluate the framework's performance against baseline models on a clinical dataset.

Main Methods:

• A multiscale diffusion framework incorporating a segmentation mask pyramid.
• A scale-specific loss function to guide learning at different resolutions.
• Evaluation on the SynthRAD2023 brain dataset using standard image quality metrics.

Main Results:

• Achieved a Mean Absolute Error (MAE) of 61.82 HU.
• Attained a Peak Signal-to-Noise Ratio (PSNR) of 32.05 dB.
• Obtained a Structural Similarity Index (SSIM) of 0.90, outperforming baseline models.

Conclusions:

• Multiscale anatomical guidance significantly improves sCT fidelity and consistency.
• The proposed framework facilitates high-quality CBCT-to-CT translation for radiotherapy.
• This approach holds promise for advancing adaptive radiotherapy workflows.