Reference-Driven Compressed Sensing MR Image Reconstruction Using Deep Convolutional Neural Networks without...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning significantly reduces Magnetic Resonance Imaging (MRI) scan times and improves image reconstruction in Compressed Sensing MRI (CS-MRI).
• Clinical applications of CS-MRI are hindered by the need for extensive, high-quality, patient-specific training datasets for deep learning models.

Purpose of the Study:

• To develop a novel deep learning-based CS-MRI reconstruction method that eliminates the requirement for pre-training or large training datasets.
• To reduce reliance on patient-specific datasets for clinicians and improve the efficiency of MRI reconstruction.

Main Methods:

• The proposed method utilizes the Deep Image Prior (DIP) framework, incorporating a high-resolution reference MRI as input to a convolutional neural network.
• A reference-driven strategy is employed to guide the network's learning process and induce structural priors.
• A k-space data correction step is integrated to ensure consistency with measured data, enhancing reconstruction accuracy.

Main Results:

• The reference-driven DIP framework demonstrated improved efficiency and effectiveness in network learning for MRI reconstruction.
• The addition of k-space data correction further refined image reconstruction accuracy.
• Experiments on in vivo MR datasets confirmed the method's ability to achieve superior reconstruction from undersampled k-space data.

Conclusions:

• The developed deep learning method offers a viable solution for CS-MRI reconstruction without extensive training data requirements.
• This approach significantly lowers the barrier for clinical adoption by reducing dependence on large, curated datasets.
• The method shows promise for faster, more accurate MRI acquisition and reconstruction in clinical settings.