Fig 1.
Deep CNN and GANs accurately predict liver tissue structures from cell border images.
(a) Schematic representation of TiMiGNet. (b) 2D sections of 3D fluorescent images of the actin mesh (membranes), Bile Canaliculi and Sinusoids (experimental images) as well as the corresponding predictions by TiMiPNet and TiMiGNet+ in 2D. c) Quantification of the performance of the predictions of the BC network and sinusoids generated by the models. The test images were divided in 128x128x128 cubes and the metrics were estimated independently for each cube, i.e. each dot represents one image cube. The box plots enclose values from the lower to upper quartiles. The middle line represents the median and the whiskers show the data range (N = 1).
Fig 2.
Deep tissue reconstructions using TiMiGNet.
(a) 2D maximum projections of ~20 μm of the axial sections of 3D fluorescent images of the actin mesh (membranes), Bile Canaliculi, Sinusoids (experimental images) as well as the predictions of TiMiPNet and TiMiGNet. (b) Quantification of the BC signal (mean intensity) along the tissue depth, in the corresponding images of panel a. (c) Quantification of the Sinusoids signal (mean intensity) along the tissue depth, in the corresponding images of panel a (N = 1).
Fig 3.
Prediction of tissue microstructure in human liver tissue using TiMiGNet.
b) 2D sections of 3D fluorescent images of the actin mesh (membranes) and Bile Canaliculi (experimental images) of human liver tissue together with the corresponding predictions of the TiMiGNet 2D model trained in mouse and human tissue images. c-d) Quantification of morphological BC parameters for experimental images and TiMiGNet predictions: Radius distribution(b), mean radius (c) and mean branch length (d) along the CV-PV axis. The error bars represent the standard deviations of the values in the region (N = 1).
Fig 4.
Prediction of Kupffer cells shapes in liver tissue using TiMiGNet+.
a) 2D sections of 3D fluorescent images of the actin mesh (membranes), Kupffer cells (experimental images) as well as the corresponding predictions by TiMiPNet, TiMiGNet, TiMiGNet+ in 2D. b-c) Quantification of the performance of the predictions of the Kupffer cells generated by the models using FID (b) as well as Mean Squared Error and b) Structural Similarity Index Measure (c). The test images were divided in 128x128x128 cubes and the metrics were estimated independently for each cube, i.e. each dot represents one image cube. The box plots enclose values from the lower to upper quartiles. The middle line represents the median and the whiskers show the data range (N = 1).