Fig. 6

a Schematic and photograph of the optimized Low-LightField microscope. (1) Sample, (2) Objective, (3) Tube lens, (4) Microlens array, (5) Relay lens, (6) camera. b–d Training pipeline to obtain fast deconvolution of 2D experimental lightfield data into 3D image stacks. b A 3D image stack was acquired on fluorescent samples representative of the bioluminescent signal in the final experiment. The stack was convolved with the lightfield PSF, to obtain a synthetic lightfield image, which was subsequently used to map onto the 3D ground-truth stack (see also Supplementary Fig. 7). The training quality of the individual models (c) was tested against unseen samples by calculating the difference and similarity to the ground truth (N = 70 images/model). The best model (indicated with an asterisk, p < 1.2e-9 compared to the 2nd best model, Wilcoxon test) with the lowest error and highest similarity was used to d reconstruct experimental bioluminescence images. Scale bars = 50 μm. Blue dots in c indicated median ± 95% confidence interval. e Pipeline for bioluminescence reconstruction. An initial (i) CARE denoising step is used to increase the SNR of (ii) noisy bioluminescent lightfield images. The individual layers are color coded according to their function. The clean images (iii) are fed into the VCD network²⁴ (iv) after perspective extraction to reconstruct the 3D information. Scale bar = 50 μm.