A dynamic reconstruction and motion estimation framework for cardiorespiratory motion-resolved real-time volumetric...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy

Background:

• Cardiorespiratory motion significantly impacts the accuracy of Magnetic Resonance (MR) imaging during radiotherapy.
• Accurate real-time tracking of this motion is essential for adaptive radiotherapy strategies.

Purpose of the Study:

• To develop DREME-MR, a system for joint reconstruction of patient anatomy and a patient-specific cardiorespiratory motion model from 3D pre-treatment MR scans.
• To enable real-time volumetric MR imaging and motion tracking with minimal intra-treatment data using a learned motion encoder.

Main Methods:

• DREME-MR integrates dynamic MRI reconstruction and real-time imaging in a dual-task learning framework using spatiotemporal implicit-neural-representation (INR).
• A multilayer perceptron (MLP)-based motion encoder is trained to infer motion coefficients from k-space data for real-time tracking.
• Cardiac and respiratory motion basis components (MBCs) are decoupled using a frequency-guided strategy.

Main Results:

• DREME-MR achieved real-time 3D cardiorespiratory motion tracking with latency under 165 ms.
• Evaluations on a digital phantom showed low center-of-mass tracking errors for a lung tumor (0.73±0.38mm) and the left ventricle (1.69±1.12mm).
• Human scan data demonstrated strong motion correlations for the liver (0.96) and left ventricle (0.65).

Conclusions:

• DREME-MR successfully enables low-latency, real-time 3D MRI and cardiorespiratory motion tracking.
• The system has the potential to significantly advance intra-treatment MR-guided adaptive radiotherapy, including real-time multileaf collimator tracking.