Abstract
Hyperpolarized gas MRI is a reference standard for imaging regional lung ventilation, but its reliance on specialized equipment and contrast agents limits availability. Synthesizing ventilation from structural proton (¹H) MRI, which is widely available and non-ionizing, could broaden access to functional lung imaging. Previous synthesis work has primarily relied on convolutional neural networks (CNNs); mechanisms such as adversarial training and attention-based architectures remain unexplored for this task. We synthesized three-dimensional hyperpolarized ¹²⁹Xe-MRI ventilation from ¹H-MRI together with a derived specific-ventilation (SV) surrogate, and systematically compared the following architectures: CNN, CNN-Generative Adversarial Networks (GAN), Vision Transformer (ViT), ViT-GAN, alongside an investigation of the loss functions most appropriate for the task. Using five-fold cross-validation on 150 paired scans, synthetic ventilation was evaluated against ¹²⁹Xe-MRI using mean absolute error, structural similarity index measure and Spearman’s rank correlation. Transformer-based models outperformed their convolutional counterparts. The strongest models, a ViT trained with a Huber loss and a ViTGAN trained with an LPIPS generator and least-squares adversarial loss, attained voxel-wise mean Spearman correlations with ¹²⁹Xe-MRI of 0.711 (median 0.713, Q1 0.655, Q3 0.753) and 0.688 (median 0.693, Q1 0.640, Q3 0.736), significantly exceeding the physiologically informed PhysVENeT baseline (mean 0.627, median 0.675, Q1 0.551, Q3 0.745). Attention-based synthesis can therefore recover accurate hyperpolarized gas ventilation from routinely available proton MRI, supporting more widely accessible functional lung imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/TIA_032.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{LiuChe_Deep_MICCAISAT2026,
author = { Liu, Chensheng AND Astley, Joshua R. AND Biancardi, Alberto M. AND Wild, Jim M. AND Tahir, Bilal A.},
title = { { Deep Learning-Based Synthesis of Hyperpolarized Gas MRI Lung Ventilation from Multi-Inflation Proton MRI: A Comparison of CNN, GAN and Transformer Architectures } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17269},
month = {pending},
page = {pending}
}
