Multimodal synthetic CT generation in tumor radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• MRI-guided radiation therapy (MRIgRT) offers advantages but requires CT for electron density.
• Current synthetic CT (sCT) generation methods face challenges in global and local accuracy.
• Efficient and accurate sCT generation from MRI remains a critical need for MRIgRT.

Purpose of the Study:

• To develop a novel deep learning model for improved sCT image generation from MRI.
• To enhance both global context and local detail accuracy in sCT synthesis.
• To achieve high-accuracy sCT with minimal increase in model complexity.

Main Methods:

• A generative adversarial network, RC-MambaGAN, integrating Mamba blocks and a Residual Constraint (RC) strategy was proposed.
• Mamba blocks were used to capture long-range contextual information while preserving local features.
• The RC strategy minimized residuals for better alignment between sCT and real CT images.

Main Results:

• RC-MambaGAN demonstrated high performance with MAE of 37.094 ± 8.761 HU, PSNR of 28.424 ± 1.145 dB, SSIM of 0.947 ± 0.004, and MI of 1.426 ± 0.012.
• External validation confirmed superior performance over state-of-the-art methods in image quality and quantitative accuracy.
• The model accurately depicted structures visible in MRI but not in CT.

Conclusions:

• RC-MambaGAN significantly enhances sCT image quality and anatomical accuracy from MRI.
• This advancement supports the clinical feasibility of MRI-only radiotherapy workflows.
• The model offers a promising solution for electron density estimation in MRIgRT.