Deep learning for improving ZTE MRI images in free breathing

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonary Medicine

Background:

• Lung MRI faces challenges due to short T2/T2* relaxation times and motion.
• Zero Echo Time (ZTE) sequences offer potential for CT-like lung MR images.
• Existing ZTE methods like ZTE4D show motion artifacts, limiting clinical use.

Purpose of the Study:

• To develop and evaluate a deep learning pipeline using fully convolutional neural networks (FCNNs) to improve image quality of free-breathing 4D lung MRI.
• To reduce motion artifacts and enhance visualization of lung parenchyma in ZTE4D images.

Main Methods:

• A FCNN model was trained on ZTE breath-hold (ZTE-BH) images artificially degraded to simulate ZTE4D motion artifacts.
• The model learned to remove noise and transformations, aiming to reproduce high-quality images.
• The trained FCNN was tested on unseen ZTE4D free-breathing lung MRI data from healthy volunteers.

Main Results:

• The FCNN model significantly improved image quality, reducing ghosting artifacts and blurring.
• Quantitative improvements included a 1.98-fold increase in lung parenchyma SNR and a 4.2% increase in CNR for intrapulmonary vessels.
• Visual assessment showed sharper diaphragm contours and enhanced anatomical structure visibility.

Conclusions:

• Deep learning with FCNNs can effectively enhance the image quality of free-breathing ZTE4D lung MRI.
• This technique improves the visualization of lung parenchyma across different respiratory phases, making it more suitable for clinical applications.