Denoising low-field MR images with a deep learning algorithm based on simulated data from easily accessible...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Low-field MRI (LFMRI) suffers from poor image contrast and noise, limiting its clinical utility.
• Existing denoising methods struggle to effectively enhance contrast in LFMRI data.
• Lack of sufficient in-vivo LFMRI datasets hinders the development of data-driven solutions.

Purpose of the Study:

• To develop and validate a deep learning-based denoising method for improving contrast ratio in LFMRI.
• To address the challenge of limited in-vivo training data by utilizing synthetic datasets.
• To compare the proposed method against established non-deep learning techniques.

Main Methods:

• An advanced 3D deep convolutional residual network was designed for image denoising.
• Synthetic brain imaging datasets were generated to mimic LFMRI contrast and noise.
• The model was trained on synthetic data and evaluated on both synthetic and in-vivo LFMRI datasets.
• Performance was compared against the BM4D algorithm using Relative Contrast Ratio (RCR) and spatial frequency preservation.

Main Results:

• The deep learning model significantly increased the Relative Contrast Ratio (RCR) in synthetic LFMRI data.
• Similar RCR improvements were observed in in-vivo LFMRI data across various imaging conditions.
• The proposed method demonstrated superior performance compared to BM4D in enhancing RCR.
• The model effectively preserved high spatial frequency components in in-vivo LFMRI images.

Conclusions:

• The developed 3D deep convolutional residual network is an effective denoising method for LFMRI.
• The use of synthetic data is a viable strategy to overcome limitations of in-vivo LFMRI datasets for deep learning.
• This approach offers a promising solution for enhancing LFMRI image quality and diagnostic potential.