LAPNet: Deep-learning based non-rigid motion estimation in k-space from highly undersampled respiratory and cardiac resolved acquisitions
Thomas Küstner1,2, Jiazhen Pan3, Haikun Qi2, Gastao Cruz2, Kerstin Hammernik3,4, Christopher Gilliam5, Thierry Blu6, Sergios Gatidis1, Daniel Rueckert3,4, René Botnar2, and Claudia Prieto2
1Department of Radiology, Medical Image and Data Analysis (MIDAS), University Hospital of Tübingen, Tübingen, Germany, 2School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom, 3AI in Medicine and Healthcare, Klinikum rechts der Isar, Technical University of Munich, München, Germany, 4Department of Computing, Imperial College London, London, United Kingdom, 5RMIT, University of Melbourne, Melbourne, Australia, 6Chinese University of Hong Kong, Hong Kong, Hong Kong
A novel deep learning non-rigid registration in
k-space inspired by optical flow is proposed. For highly accelerated
acquisitions of respiratory and cardiac motion, this enables aliasing-free
motion estimation which shows superior accuracy to conventional image-based
registrations.
Fig. 1: Proposed LAPNet to perform non-rigid registration in k-space.
Moving νm and reference νr k-spaces are tapered to a smaller support W. The bundle of
k-space patches is processed in a succession of convolutional filters (kernel
sizes and channels are stated) to estimate the in-plane flows u1,u2 at the central voxel location determined by
the tapering T/window W for size 33x33x33. Overall 3D
deformation field u is obtained from a sliding window over all
voxels in orthogonal directions.
Fig. 2: Respiratory non-rigid
motion estimation in a patient with a liver metastasis in segment VIII. Motion
displacement is estimated by the proposed LAPNet in k-space in comparison to
image-based non-rigid registration by FlowNet-S (neural network) and NiftyReg
(cubic B-Spline). Estimated flow displacement are depicted in coronal and
sagittal orientation. Undersampling was performed prospectively with a 3D
Cartesian random undersampling for 8x and 30x acceleration.