Real time volumetric MRI for 3D motion tracking via geometry-informed deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Volumetric MRI is crucial for radiotherapy but limited by long acquisition times.
• Accurate 3D motion tracking is essential for precise radiation delivery.
• Existing methods struggle with the temporal resolution needed for real-time adjustments.

Purpose of the Study:

• To develop a geometry-informed deep learning framework for sub-second volumetric MRI acquisition.
• To enable high-precision 3D motion tracking for MRI-guided radiotherapy.
• To overcome the limitations of long image acquisition times in current MRI techniques.

Main Methods:

• A 2D-3D deep learning network with a geometry module was designed to embed k-space encoding priors.
• A 2D generation network augmented sparse data, and a geometry module unfolded it to 3D.
• A 3D refinement network produced high-resolution volumetric images from patient-specific models.

Main Results:

• The framework achieved median GTV centroid distances of 0.4-0.5 mm compared to ground truth.
• 95-percentile Hausdorff distances for GTV contours were 3.2-4.7 mm, comparable to image resolution.
• The model demonstrated robustness to longitudinal anatomical changes in patient data.

Conclusions:

• Integrating geometric priors into deep learning enhances volumetric imaging capabilities.
• The developed framework offers high spatial and temporal resolution for 3D motion tracking.
• This technology holds significant potential for improving the precision of MRI-guided radiotherapy.