Purpose: Hyperpolarized gas MRI can visualize and quantify regional lung ventilation with exquisite detail. However, clinical uptake is limited to a few centres worldwide due to the requirement of highly specialized equipment and exogenous contrast. Alternative, non-contrast techniques have been proposed to model ventilation, including multi-inflation CT-based models of ventilation that have shown moderate spatial correlations with hyperpolarized gas MRI. Recent advances in deep learning (DL) using convolutional neural networks (CNNs) have shown great promise for numerous medical imaging applications, including image synthesis. Here, we propose a multi-channel 3D CNN approach for synthesizing hyperpolarized gas MRI scans using multi-inflation CT and a conventional density-based CT ventilation imaging (CTVI) model as inputs. We compare the accuracy of the CNN-based ventilation images with those generated by the conventional CTVI model.
Methods: The dataset comprises paired inspiratory and expiratory CT and helium-3 hyperpolarized gas MRI for 47 patients with a range of pulmonary pathologies. A VNet CNN with a root mean square error loss function was used for image synthesis. We performed 6-fold cross-validation on the dataset and evaluated the correlation of the synthetic ventilation and hyperpolarized gas MRI scans at full resolution using Spearman’s ρ on all lung voxels. The combined DL and CTVI model was compared to the conventional CTVI model.
Results: Qualitatively, there are numerous examples of the CNN-generated synthetic ventilation images accurately replicating ventilation defects in the hyperpolarized gas MRI scans. Quantitatively, the CNN approach yielded statistically significant improvements in Spearman’s correlation compared to the CTVI model (p<0.0001) with mean±SD ρ of 0.5±0.2 vs 0.4±0.2.
Conclusion: We propose a multi-channel CNN-based approach to generate synthetic ventilation images from routinely acquired multi-inflation CT and CTVI models. We show that a synergy between conventional model-based CTVI and DL yields statistically significant improvements in spatial correlation compared with conventional CTVI modelling.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by Yorkshire Cancer Research, Weston Park Cancer Charity, National Institute of Health Research and the Medical Research Council.