Synthesizing CT images from MR images with deep learning: model generalization for different datasets through...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• MR-only radiotherapy is gaining traction, necessitating reliable CT image synthesis from MR data.
• Current MR-to-CT models often lack generalizability across different institutions and imaging protocols.
• Addressing model generalization is crucial for the widespread adoption of MR-only workflows.

Purpose of the Study:

• To investigate and improve the generalizability of MR-to-CT conversion models.
• To evaluate different strategies for training MR-to-CT conversion models using CycleGAN.
• To assess model performance on diverse datasets from different hospitals and MR protocols.

Main Methods:

• Utilized CycleGAN for MR-to-CT image synthesis.
• Compared four training strategies: source, target, combined, and adapted models.
• Employed transfer learning for the adapted model, pre-training on a source domain and fine-tuning on a target domain.
• Quantitatively evaluated models using MAE, RMSE, PSNR, and SSIM metrics.

Main Results:

• The adapted model, using transfer learning, achieved superior quantitative performance compared to other methods.
• The adapted model demonstrated high generalization capabilities on unseen data from different acquisition centers and protocols.
• The source-only model exhibited the poorest performance, highlighting the need for robust training strategies.

Conclusions:

• Pre-trained CycleGAN models exhibit significant generalization potential for synthesizing CT images from limited MR data.
• The adapted model approach proves effective for MR-to-CT conversion in diverse clinical settings.
• This study validates the concept of creating reliable synthetic CT images for MR-only radiotherapy.