PhysVENeT: a physiologically-informed deep learning-based framework for the synthesis of 3D hyperpolarized gas MRI...

Area of Science:

• Medical Imaging
• Pulmonary Medicine
• Artificial Intelligence

Background:

• Hyperpolarized gas MRI provides functional lung ventilation data but has limited clinical use.
• Proton (1H)-MRI ventilation mapping shows moderate correlation with hyperpolarized gas MRI.
• Deep learning (DL) shows potential for synthesizing functional lung images.

Purpose of the Study:

• To develop a novel 3D deep learning framework (PhysVENeT) for synthesizing 3D lung ventilation surrogates using multi-inflation structural 1H-MRI.
• To evaluate PhysVENeT's accuracy in reflecting ventilation defects compared to existing methods.
• To assess PhysVENeT's performance on a diverse dataset including post-COVID-19 patients.

Main Methods:

• A 3D multi-channel convolutional neural network (PhysVENeT) was designed.
• The network integrates physiologically-informed ventilation mapping with multi-inflation structural 1H-MRI data.
• Paired inspiratory/expiratory 1H-MRI and hyperpolarized gas MRI scans from 170 participants were used, with external validation on 20 post-COVID-19 patients.

Main Results:

• PhysVENeT significantly outperformed conventional 1H-MRI ventilation mapping and DL methods lacking structural integration.
• The framework accurately reflected ventilation defects.
• Minimal overfitting was observed on external validation data.

Conclusions:

• PhysVENeT offers an accurate method for synthesizing 3D lung ventilation surrogates from 1H-MRI.
• This DL approach provides a viable, accessible alternative to hyperpolarized gas MRI for assessing regional lung ventilation.
• PhysVENeT demonstrates robust performance and generalizability across different pulmonary pathologies.