Anatomy-Aware MR-Imaging-Only Radiotherapy.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Computed Tomography (CT) synthesis supplements electron density data and corrects MR-CT registration errors for MR-only radiotherapy planning.
• Traditional MR-to-CT translation methods often require region-specific models due to anatomical variations, increasing development complexity.
• A unified approach is needed to handle diverse anatomical regions efficiently for MR-only radiotherapy.

Purpose of the Study:

• To develop a unified, prompt-driven model for MR-to-CT image translation that dynamically adapts to various anatomical regions.
• To improve structural consistency and anatomical detail in synthesized CT images from MRI data.
• To validate the model's performance against state-of-the-art methods and assess its dosimetric accuracy for radiotherapy applications.

Main Methods:

• Proposed a unified, prompt-driven model for MRI-to-CT image translation.
• Implemented a region-specific attention mechanism with region-aware vectors and dynamic gating factors.
• Evaluated the model on three anatomical datasets using qualitative, quantitative, and dosimetric analyses.

Main Results:

• The unified model successfully generated CT images with high structural consistency across multiple anatomical regions.
• Synthesized CT images exhibited superior clarity and anatomical detail compared to existing state-of-the-art translation models.
• Dosimetric analysis confirmed that the synthesized images yield dose distributions closely matching those from real CT scans.

Conclusions:

• The proposed unified, prompt-driven model effectively performs MRI-to-CT image translation for diverse anatomical regions.
• This approach shows significant potential for enabling robust MR-only radiotherapy planning across various anatomies.
• The model's source code has been publicly released to facilitate further research and application.