Distortion-corrected image reconstruction with deep learning on an MRI-Linac.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Magnetic Resonance Imaging (MRI) offers superior soft-tissue contrast and avoids ionizing radiation, making it valuable for image-guided radiotherapy.
• Geometric distortions from gradient nonlinearities (GNLs) in MRI can reduce anatomical accuracy, potentially impacting radiotherapy precision.
• Slow MRI acquisition and reconstruction times hinder real-time image guidance during radiotherapy.

Purpose of the Study:

• To develop a deep learning-based method for rapid, distortion-corrected image reconstruction from raw k-space data for MRI-guided radiotherapy.
• To address the limitations of geometric inaccuracies and slow processing in current MRI techniques for radiotherapy.

Main Methods:

• A Distortion-Corrected Reconstruction Network (DCReconNet) was developed using interpretable unrolling networks and convolutional neural networks (CNNs).
• DCReconNet learns effective regularizations and nonuniform fast Fourier transforms for gradient nonlinearity encoding.
• The network was trained on public brain MRI data and validated on phantom, brain, pelvis, and lung images from a 1.0T MRI-Linac, comparing against Compressed Sensing (CS), Parallel Imaging (PI), and UNet methods.

Main Results:

• DCReconNet demonstrated superior preservation of image structures compared to CS and PI reconstruction methods.
• Achieved the highest Structural Similarity Index Measure (SSIM) (0.95 median) and lowest Root Mean Square Error (RMSE) (<0.04) on accelerated brain images.
• DCReconNet reconstruction was over 10 times faster than traditional iterative, regularized reconstruction techniques.

Conclusions:

• DCReconNet offers a solution for fast and geometrically accurate image reconstruction in MRI-guided radiotherapy.
• The method has significant potential to improve the quality and efficiency of tumor treatments using MRI guidance.