Real‐time 3D motion estimation from undersampled MRI using multi‐resolution neural networks

Purpose: To enable real‐time adaptive magnetic resonance imaging–guided radiotherapy (MRIgRT) by obtaining time‐resolved three‐dimensional (3D) deformation vector fields (DVFs) with high spatiotemporal resolution and low latency ([Formula: see text] ms). Theory and Methods: Respiratory‐resolved [Formula: see text] ‐weighted 4D‐MRI of 27 patients with lung cancer were acquired using a golden‐angle radial stack‐of‐stars readout. A multiresolution convolutional neural network (CNN) called TEMPEST was trained on up to 32 [Formula: see text] retrospectively undersampled MRI of 17 patients, reconstructed with a nonuniform fast Fourier transform, to learn optical flow DVFs. TEMPEST was validated using 4D respiratory‐resolved MRI, a digital phantom, and a physical motion phantom. The time‐resolved motion estimation was evaluated in‐vivo using two volunteer scans, acquired on a hybrid MR‐scanner with integrated linear accelerator. Finally, we evaluated the model robustness on a publicly‐available four‐dimensional computed tomography (4D‐CT) dataset. Results: TEMPEST produced accurate DVFs on respiratory‐resolved MRI at 20‐fold acceleration, with the average end‐point‐error [Formula: see text] mm, both on respiratory‐sorted MRI and on a digital phantom. TEMPEST estimated accurate time‐resolved DVFs on MRI of a motion phantom, with an error [Formula: see text] mm at 28 [Formula: see text] undersampling. On two volunteer scans, TEMPEST accurately estimated motion compared to the self‐navigation signal using 50 spokes per dynamic (366 [Formula: see text] undersampling). At this undersampling factor, DVFs were estimated within 200 ms, including MRI acquisition. On fully sampled CT data, we achieved a target registration error of [Formula: see text] mm without retraining the model. Conclusion: A CNN trained on undersampled MRI produced accurate 3D DVFs with high spatiotemporal resolution for MRIgRT.