Model-based deep learning with fully connected neural networks for accelerated magnetic resonance parameter mapping.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Quantitative Magnetic Resonance Imaging (qMRI)

Background:

• Quantitative magnetic resonance imaging (qMRI) holds significant potential for clinical research by imaging physical parameters of proton nuclear spin in tissue.
• Current qMRI methods face challenges with lengthy acquisition times, hindering clinical implementation and verification of accuracy and reliability.
• Deep learning (DL) has emerged as a promising technique to reduce imaging time and enhance image quality in medical imaging.

Purpose of the Study:

• To introduce and verify a novel deep learning framework, quantitative deep cascade of convolutional network (qDC-CNN), for accelerated quantitative parameter mapping in qMRI.
• To demonstrate that the proposed qDC-CNN model outperforms existing competing methods in terms of accuracy and efficiency.
• To address the need for reduced acquisition times in qMRI for practical clinical applications.

Main Methods:

• The study developed an integrated deep-learning framework, qDC-CNN, combining an unrolled image reconstruction network and a fully connected neural network for parameter estimation.
• Training and testing were performed using simulated multi-slice multi-echo (MSME) datasets from the BrainWeb database.
• Reconstruction error was evaluated using normalized root mean squared error (NRMSE) and compared against conventional DL-based methods under varying acceleration factors and numbers of contrast images.

Main Results:

• The proposed qDC-CNN achieved normalized root mean squared error (NRMSE) values within 10% for S0 and T2 parameters in most cases.
• Notably, the NRMSE values for T2 parameter estimation using qDC-CNN were significantly lower than those obtained with conventional methods.
• The framework demonstrated robust performance across different acceleration factors and reduced numbers of contrast images.

Conclusions:

• The qDC-CNN model exhibited significantly smaller reconstruction errors compared to conventional models, indicating superior performance.
• The proposed method is applicable to various qMRI sequences and offers flexibility through its modular design, allowing for performance improvements by replacing the image reconstruction module.
• This framework presents a viable solution for accelerating qMRI acquisition times, enhancing clinical relevance and enabling wider adoption.