The artefacts and non-uniform resolution accompanied with slow reconstruction speed in 13 light-field microscopy compromises its full capability for intoto observing fast biological dynamics.14 Here we demonstrate that combining a view-channel-depth (VCD) neural network with light-field 15 microscopy can mitigate these limitations, yielding artefact-free 3D image sequences with uniform 16 spatial resolution and three-orders higher, video-rate reconstruction throughput. We image neuronal 17 activities across moving C. elegans and pumping blood flow in beating zebrafish heart at single-cell 18 resolution and volume rate up to 200 Hz. 20A recurring challenge in biology is the quest to extract ever more spatiotemporal information from the targets 21 since many millisecond-transient cellular processes occur in three-dimensional tissues and across long time 22 scale. Several imaging techniques, including epifluorescence and plane illumination methods, can image live 23 samples in three dimensions at high spatial resolution 1-4 . Meanwhile, they need to record a number of 2D 24 images to comprise a 3D volume, in which the transient temporal profiles are compromised due to extended 25 acquisition time. 26The recent advent of Light field microscopy (LFM) has become a unique technique of choice for rapid and 27 instantaneous volumetric imaging 5-9 . It particularly permits the retrieval of transient 3D signal distribution 28 through post-processing of 4D light field recorded by single camera snapshot. As LFM provides high-speed 29 volumetric imaging only limited by the camera frame rate, it has delivered promising results for various 30 applications, such as recording of neuronal activities and visualization of cardiac dynamics in model 31 organisms 9-11 . Despite this advancement, its generally low spatial resolution, especially non-uniform axial 32 resolution, and presence of reconstruction artefacts by traditional light-field recovery methods yet prevents 33 its more widespread applications to capture millisecond time-scale biological processes at cellular resolution. 34Though these problems can be mitigated through optimizing the way light field being recorded 9,12 , the extra 35 system complexity could impede the wide dissemination of LFM technique. Furthermore, current LFMs 36 heavily rely on computationally demanding, iterative recovery process that intrinsically limits the overall 37 throughput of LFM reconstruction, compromising its potentials for long time-scale applications. 38Here we propose a novel LFM strategy based on a View-Channel-Depth neural network processing of light 39 field data 13 , termed VCD-LFM. By developing a light-field projection based on wave optics model, we 40 generated plenty of synthetic light-field images from high-resolution 3D images experimentally obtained 41 beforehand, and readily paired them as target and ground-truth data, respectively, for network training. The 42VCD network (VCD-Net) procedure was designed to enable the extraction of multiple views from these 2D 43 light fields...