Alexander M Matheson1, Rachel L Eddy1, Jonathan L MacNeil2, Marrissa J McIntosh1, and Grace M Parraga1,2
1Medical Biophysics, Robarts Research Institute, Western University, London, ON, Canada, 2School of Biomedical Engineering, Robarts Research Institute, Western University, London, ON, Canada
1Medical Biophysics, Robarts Research Institute, Western University, London, ON, Canada, 2School of Biomedical Engineering, Robarts Research Institute, Western University, London, ON, Canada
Co-registered hyperpolarized gas and proton MRI are used to calculate lung function biomarkers but registration is challenging due to different contrast and imaging features. Neural networks generated multi-spectral registration transforms with an average error of less than one pixel.
Five representative
examples of CNN correction performance. Images show cyan 3He
images overlaying 1H images. Top
row: images with random affine transformations applied. Middle row: proposed
registration generated by multi-input UNet. Bottom row: original co-registered
images obtained through semi-automated, landmark based registration.
Participant
demographics, pulmonary function tests and imaging measurements. Values
reported as mean (standard deviation). BMI = body mass index, PFT = pulmonary
function tests, FEV1 = forced expiratory volume in 1 second, FVC =
functional vital capacity, RV = residual volume, TLV = total lung volume, VDP =
ventilation defect percent