A multi-channel deep learning approach for lung cavity estimation from hyperpolarized gas and proton MRI
Joshua R Astley1,2, Alberto M Biancardi2, Helen Marshall2, Paul JC Hughes2, Guilhem J Collier2, Laurie J Smith2, James Eaden2, François-Xavier Blé3, Rod Hughes4, Jim M Wild2,5, and Bilal A Tahir1,2
1Oncology and Metabolism, University of Sheffield, Sheffield, United Kingdom, 2POLARIS, Sheffield, United Kingdom, 3Translational Science and Experimental Medicine, Research and Early Development, Respiratory & Immunology, BioPharmaceuticals R&D, AstraZeneca, Cambridge, United Kingdom, 4Clinical Development, Research and Early Development, Respiratory & Immunology, AstraZeneca, Cambridge, United Kingdom, 5Insigneo Institute for in silico medicine, University of Sheffield, Sheffield, United Kingdom
Functional lung imaging biomarkers, such as the ventilated defect percentage, are computed from segmentations derived from spatially co-registered functional hyperpolarized gas MRI and structural 1H-MRI. Although hyperpolarized gas and 1H-MRI can be acquired in the same or similar breaths, the acquired scans are frequently misaligned. Here, we propose a multi-channel deep learning-based approach to generate accurate lung cavity estimations (LCEs). Across all evaluation metrics, the multi-channel approach significantly outperformed single-channel approaches and generated plausible LCEs across a wide range of pulmonary pathologies. In addition, correlation and Bland-Altman analyses of lung volumes demonstrated strong correlation and minimal bias with expert LCEs.
This abstract and the presentation materials are available to members only;
a login is required.