MRgRT real-time target localization using foundation models for contour point tracking and promptable mask refinement

Area of Science:

• Medical physics and imaging
• Artificial intelligence in healthcare
• Radiotherapy technology

Background:

• Real-time target tracking is crucial for accurate MRI-guided radiotherapy (MRgRT).
• Evaluating novel AI approaches is essential to enhance MRgRT precision and efficiency.
• Foundation models offer potential for advanced image analysis in medical applications.

Purpose of the Study:

• To evaluate two foundation AI-based real-time target tracking methods for MRgRT.
• To compare these novel approaches against a transformer-based registration model (TransMorph) and inter-observer variability.
• To assess the performance using metrics like Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD).

Main Methods:

• Developed and evaluated a point-tracking model and a Segment Anything Model 2 (SAM2)-based video-object-segmentation model.
• Tested on 2D cine MRI datasets from 33 patients across two institutions.
• Compared results against TransMorph (with and without patient-specific fine-tuning) and expert annotations.

Main Results:

• Both contour tracking and SAM2 approaches achieved target segmentation comparable or superior to TransMorph without patient-specific fine-tuning.
• SAM2 showed slightly better performance than contour tracking but with higher computational cost.
• All evaluated AI methods, including TransMorph with patient-specific fine-tuning, surpassed inter-observer variability.

Conclusions:

• Foundation models demonstrate significant potential for high-quality real-time target tracking in MRgRT.
• These AI models offer performance matching state-of-the-art methods without the need for patient-specific fine-tuning.
• The findings suggest a pathway towards more accurate and efficient MRgRT treatments.