Imaging error reduction in radial cine-MRI with deep learning-based intra-frame motion compensation

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Radial cine-Magnetic Resonance Imaging (MRI) enables high-speed imaging for motion monitoring during radiotherapy.
• Intra-fractional motion during reconstruction can cause target positioning errors, especially with fast physiological movements.
• Current methods struggle with accurate real-time motion compensation in dynamic scenarios.

Purpose of the Study:

• To enhance radial cine-MRI by implementing deep-learning-based intra-frame motion compensation.
• To develop a novel network, TransSin-UNet, for estimating the end-of-frame target position.
• To reduce target positioning errors in real-time MRI-guided radiotherapy.

Main Methods:

• A novel TransSin-UNet architecture combining transformer encoder and UNet subnetworks was proposed.
• The network models spatial-temporal dependencies in sinogram data for motion compensation.
• Training and evaluation were performed using simulated 4D digital lung cancer phantoms with motion-dependent radial sampling.

Main Results:

• TransSin-UNet achieved a 50% reduction in normalized root mean-squared error (from 0.188).
• Mean Dice similarity coefficient for the gross tumor volume improved from 85.1% to 96.2%.
• The method demonstrated precise derivation of final positions for deforming anatomical structures with minimal added computational cost (4.8 ms/frame).

Conclusions:

• The proposed TransSin-UNet effectively compensates for intra-frame motion in radial cine-MRI.
• This deep learning approach significantly enhances image quality and target positioning accuracy.
• It offers a promising solution for reducing errors in real-time motion management for MRI-guided radiotherapy.