Residual motion artifact removal enables dynamic μMRI of a behaving Pachnoda marginata

Microscopic magnetic resonance imaging, also referred to as $\mu$MRI, is a non-invasive imaging modality ideal for studying small live model organisms. However, $\mu$MRI raw data acquisition is inherently sequential and slow in comparison to the biomechanics timescale of the behaving organism, leading to motion artifacts upon image reconstruction. Recently, we have developed an integrated spherical treadmill with a prospectively triggered k-space acquisition technique to provide position consistency for studying live, behaving insect using $\mu$MRI. Despite this advancement, behaving insects on the treadmill still exhibited motion artifacts due to tethered locomotion being coupled with internal organ dynamics. Here, we are addressing the large-scale non-rigid nature of the abdominal motion of the behaving insect by developing a fully retrospective gating strategy using the motion information obtained from an in-situ computer vision system. Residual motion artifacts persisting after gating are effectively managed through a deep learning technique. We trained a U-Net-based deep convolutional neural network using pairs of simulated motion-corrupted and motion-free images as a supervised image-to-image translation problem. Our results demonstrate that combining retrospective gated $\mu$MRI reconstruction with a deep learning residual motion compensation technique can significantly reduce the motional artifacts, thereby paving the way for the non-invasive dynamic imaging studies of behaving organisms with 117 $\mu$m in-plane resolution.