Fig. 2: Deep learning enables single-volume super-resolution for volumetric CFM.

Cortical regions of a Thy 1-eYFP mouse brain were imaged. a Near isotropic resolution was achieved by the proposed method in a CFM volume. Axial images were blindly enhanced by the generative neural network, which is trained and tested on a single CFM volume that spans 1-2 gigabytes in memory. The resolution improvement in the axial planes was global and consistent throughout the image space (also see Supplementary Fig. 6). The converging cross-sectional intensity profiles of a cylindrical dendrite in the XY plane (yellow line) and XZ plane (blue line) indicate a near-perfect isotropic resolution in comparison to the input in XZ (dotted blue line), with XY FWHM of 1.71 µm and XZ FWHM of 1.76 µm, greatly reduced from input XZ FWHM of 6.04 µm. Scale bars: 100 µm, 50 µm, 20 µm (2D), and 20 µm (3D) in the progressively zooming order. b Image restoration results (“Network”) show the upper cortical regions of the mouse cortex, as MIP images of 150-µm thickness. The results were compared to the original axial imaging (“Input”) and reference lateral imaging (“90°-rotated”). Zoomed-in ROIs are marked as yellow boxes. Suppressed or blurred details were recovered in the network output images and matched the lateral imaging. Scale bars: 50 µm (top ROIs), 10 µm (middle ROIs), 20 µm (bottom ROIs). c 3D reconstruction of pyramidal neurons in the upper cortical layer before and after restoration, with neuronal tracings. Resolution improvement in the axial plane allows a more precise and detailed reconstruction of 3D neuronal morphology. We verified the additional neuronal tracings by substantiating in the corresponding locations from the lateral imaging. The under-sampling and the Z-blurring made it more difficult to trace neurites that run perpendicularly to the scanning direction (arrows in 2D ROIs), whereas more accurate tracing was possible from the network output. Scale bars: 50 µm (3D) and 10 µm (2D ROI). d PSNR distribution of MIP images as a distance metric to the reference image, with pairwise improvements. MIP images (n = 31 independent images) were from 140 × 140 µm² cross sections with depths of 150 µm. For the box plots, the box shows the IQR between Q1 and Q3 of the dataset, with the central mark showing the median and the whiskers indicating the minimum (Q1-1.5*IQR) and the maximum (Q3+1.5*IQR). e Cross-sectional intensity profiles from marked lines in zoomed-in ROIs from (b). The 90°-rotated line is registered to the input. In (a) and (c) panels, the color bars represent the signal intensity normalized between 0 and 1.