End-to-End Nonrigid MOtion COrrected DEep Learning Spiral Image REconstruction (MOCO-DESIRE) for free breathing perfusion imaging

Motivation: Traditional MOCO myocardial perfusion imaging reconstructions require multiple iteration and time-consuming inverse deformation field estimations, posing challenges for rapid on-line reconstruction. Goal(s): To develop an end-to-end deep learning-based motion corrected reconstruction network for fast cardiac perfusion imaging. Approach: The end-to-end Nonrigid Motion-Corrected Deep Learning Spiral Image Reconstruction (MOCO-DESIRE) network integrates the VoxelMorph network for estimating deformation fields from temporal basis informed NUFFT images and employs the DESIRE network with 3D convolutional filters to reconstruct high-quality perfusion images from undersampled navigator guided spiral perfusion datasets. Results: The end-to-end MOCO-DESIRE network achieves motion-corrected, high-resolution myocardial perfusion reconstructions (1.3 mm²) in just 4.5 seconds. Impact: This approach obtains high-quality motion corrected perfusion images within seconds, facilitating the possibility of on-line imaging.