Intra-frame motion deterioration effects and deep-learning-based compensation in MR-guided radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Hybrid magnetic resonance linear accelerators (MR-Linacs) use 2D+t cine MR imaging for intra-fractional motion monitoring.
• Limited temporal resolution of cine MR imaging causes motion artifacts and latency, especially during fast breathing.

Purpose of the Study:

• Investigate intra-frame motion deterioration effects in MR-guided radiotherapy (MRgRT).
• Explore deep-learning-based compensation using k-space data for motion correction.

Main Methods:

• Simulated intra-frame motion using 4D digital phantoms of lung cancer patients.
• Developed a motion pattern perturbation scheme with three degrees of freedom.
• Trained U-Net models (NNImgLoss-L1, NNFloss-L1, NN L2-Loss) with different loss functions for image reconstruction.

Main Results:

• Intra-frame motion caused ~50% imaging latency in Cartesian trajectories.
• All compensation models reduced image errors (MSE/MAE) and GTV position offset.
• Gross tumor volume (GTV) contouring accuracy improved: Dice Similarity Coefficient (DSC) increased from 88% to 96%, and Hausdorff distance (HD95) decreased from 4.8mm to 2.1mm.

Conclusions:

• Deep learning effectively compensates for intra-frame motion in MRgRT.
• Utilizing later acquired k-space data improves earlier k-space component convergence.
• This technique enhances image quality and precision in motion-guided radiotherapy.