Deep Leaning Based Multi-Modal Fusion for Fast MR Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) utilizes T1-weighted (T1WI) and T2-weighted (T2WI) images for clinical insights.
• Long MRI acquisition times increase susceptibility to motion artifacts, compromising image quality.
• Current accelerated MRI reconstruction methods often rely on single-modality data.

Purpose of the Study:

• To develop a deep learning approach for accelerating T2WI reconstruction by leveraging multi-modal MR acquisitions.
• To fuse complementary T1WI and under-sampled T2WI data for enhanced image reconstruction.
• To validate the efficacy of the proposed method in terms of speed, artifact reduction, and signal-to-noise ratio (SNR).

Main Methods:

• A novel deep learning architecture, Dense-Unet, was developed for multi-modal MR image reconstruction.
• The Dense-Unet model was trained to reconstruct fully-sampled T2WI from under-sampled T2WI and T1WI data.
• The method was evaluated on its ability to reconstruct 3D T2WI volumes rapidly and its transferability across different MRI scanners.

Main Results:

• Dense-Unet successfully reconstructed 3D T2WI volumes in under 10 seconds with an 8-fold k-space under-sampling rate.
• The reconstructed images exhibited negligible aliasing artifacts and minimal signal-noise-ratio (SNR) loss.
• The model demonstrated excellent transferability, performing well on datasets from diverse MR scanners.

Conclusions:

• The proposed Dense-Unet method significantly accelerates T2WI reconstruction using multi-modal MR data.
• This deep learning approach offers a promising solution for reducing MRI scan times and mitigating motion artifacts.
• The method holds substantial potential for various clinical applications requiring faster and high-quality MRI.