Click here to


Are you sure ?

Yes, do it No, cancel

Deriving Ventilation Imaging From Free Breathing Proton MRI Via Deep Convolutional Neural Network

D Capaldi1*, F Guo2, L Xing3, G Parraga4, (1) Stanford University, Stanford, CA, (2) Sunnybrook Research Institute, (3) Stanford Univ School of Medicine, Stanford, CA, (4) Western University, London, ON, CA


(Sunday, 7/12/2020) 1:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 2

Purpose: Hyperpolarized noble-gas magnetic-resonance-imaging (MRI) has been optimized to measure lung function, albeit clinical translation of this approach has been limited. As an alternative, free-breathing pulmonary ¹H MRI using clinically available MRI systems and pulse sequences provides non-contrast-enhanced ventilation-weighted maps (Bauman et al., MRM, 2009). Although highly promising, broad research and clinical applications of these methods have been slow, due in part to somewhat complex image processing components needed to generate ventilation-weighted maps. Here, our objective was to explore deep-convolutional-neural-network (DCNN) to learn synthetic MRI ventilation images automatically from free-breathing ¹H MRI as a surrogate of hyperpolarized noble-gas MRI and evaluate the performance for generating imaging biomarkers across a wide range of lung disease.

Methods: Hyperpolarized noble-gas MRI and dynamic free-breathing ¹H MRI were acquired, as previously described (Capaldi et al., Acad. Radiol., 2014), in 72 subjects with different lung diseases (asthma/chronic-obstructive-pulmonary-disease/non-small-cell-lung-cancer) and healthy-volunteers. The DCNN model (Han, Med. Phys., 2017) was trained end-to-end to learn a direct mapping from free-breathing ¹H MRI to their corresponding noble-gas MRI ventilation images. The 72 subjects were used as experimental data for a six-fold cross-validation study. Each learned ventilation map was compared with the noble-gas MRI ventilation image of the same patient using mean-absolute-error (MAE) and Pearson-correlation-coefficient (PCC) on a voxel-by-voxel basis to evaluate the accuracy of the DCNN model.

Results: The DCNN model produced ventilation maps with an average MAE and PCC of 3.1±0.3 and 0.93±0.01, respectively. While training required ~30mins, applying the trained DCNN model to generate a synthetic ventilation map only requires ~1sec, which is much faster than alternative approaches for generating ventilation images from free-breathing ¹H MRI.

Conclusion: In this proof-of-concept study, a DCNN model method was applied to free-breathing ¹H MRI to generate synthetic ventilation-weighted maps that were quantitatively and spatially related to hyperpolarized noble-gas MRI ventilation images.

Funding Support, Disclosures, and Conflict of Interest: Dr. Capaldi received funding support from the Natural Sciences and Engineering Research Council (NSERC) of Canada Postdoctoral Fellowship Program.


Convolution, Ventilation/perfusion, MRI


IM- MRI : Machine learning, computer vision

Contact Email