SAM-guided structural consistency constraints for unsupervised MR-to-CT synthesis.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• MRI-only radiotherapy requires accurate CT synthesis for treatment planning.
• Unsupervised MR-to-CT synthesis methods often lack structural consistency, causing anatomical misalignments.
• Existing methods struggle with precise anatomical localization, impacting radiotherapy efficacy.

Purpose of the Study:

• To develop an unsupervised MR-to-CT synthesis method with enhanced structural consistency.
• To leverage the Segment Anything Model (SAM) for improved anatomical alignment between MR and synthesized CT images.
• To validate the proposed method's performance in synthesizing CT images for nasopharyngeal carcinoma radiotherapy.

Main Methods:

• Integration of the Segment Anything Model (SAM) into unsupervised MR-to-CT synthesis.
• Implementation of SAM-guided structural consistency constraints for anatomical alignment.
• Utilizing prompt augmentation during training to improve generalization across anatomical structures.

Main Results:

• The proposed SAM-guided method achieved superior performance on a nasopharyngeal carcinoma dataset.
• Quantitative metrics included Mean Absolute Error (MAE) of 108.12 HU, Peak Signal-to-Noise Ratio (PSNR) of 24.24 dB, and Structural Similarity Index (SSIM) of 0.753.
• Exceptional anatomical consistency in synthesized CT images was confirmed through quantitative analysis and dose distribution assessments.

Conclusions:

• The SAM-guided approach significantly improves anatomical consistency in synthesized CT images.
• This method holds potential for enhancing anatomical localization accuracy in MRI-only radiotherapy.
• The findings support the advancement of MRI-only radiotherapy workflows through improved image synthesis techniques.