Pulmonary Ventilation Maps Generated with Free-breathing Proton MRI and a Deep Convolutional Neural Network

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Hyperpolarized noble gas MRI is valuable for lung ventilation assessment but faces limited clinical use.
• Free-breathing proton MRI offers a contrast-free method to quantify lung function using existing systems.
• This approach may reveal ventilation details not visible or easily segmented in standard MRI.

Purpose of the Study:

• To develop and validate deep convolutional neural networks (DCNNs) for generating synthetic MRI ventilation scans.
• To use DCNN-derived ventilation maps from free-breathing MRI as a surrogate for noble gas MRI.
• To assess the DCNN approach across diverse lung diseases.

Main Methods:

• A U-Net-based DCNN model was trained using paired noble gas MRI and free-breathing MRI scans from 114 participants.
• The DCNN mapped free-breathing proton MRI to hyperpolarized helium-3 (³He) MRI ventilation data.
• Validation involved comparing DCNN-generated maps with ³He MRI using correlation coefficients and Dice similarity coefficients (DSC).

Main Results:

• DCNN ventilation maps showed a high correlation with ³He MRI (mean r = 0.87).
• The mean DSC between DL ventilation MRI and ³He MRI was 0.91, indicating strong agreement.
• DL ventilation MRI accurately reflected ventilation defect percentages, correlating well with ³He MRI (rS = 0.83) and pulmonary function (FEV1, rS = -0.51).

Conclusions:

• Deep convolutional neural networks can generate accurate lung ventilation maps from free-breathing proton MRI.
• This deep learning approach serves as a viable surrogate for noble gas MRI, applicable across various lung pathologies.
• The method correlates well with established ventilation imaging and pulmonary function tests, enhancing clinical utility.