Distribution of calbindin-positive neurons across areas and layers of the marmoset cerebral cortex
Fig 2
Training and evaluation of the U-Net Convolutional Neural Network (U-Net CNN).
(A) Examples of counting boxes manually annotated by neuroanatomists (red circles of various sizes). The entire counting box is 150 × 150 μm while the area annotated with the red square is a 256 × 256 px (127 μm × 127 μm) image patch used for training the CNN. A total of 4,072 counting boxes were defined, including boxes representing various densities of neurons (top left and top middle), parts of the tissue containing blood vessels or artifacts (top right and bottom left), as well as samples of white matter (bottom center) or samples that do not depict the brain tissue (bottom right). (B) Density map generated based on the counting box indicated by the gray arrow. A single Gaussian blob corresponds to a single CB+ neuron, and 43.89 neurons are located within the indicated image patch (red square). Note that non-integer cell number estimates are possible, as explained in Automated detection of CB+ neurons Methods section. Upon training the U-Net CNN counting boxes not previously presented to the network can be turned into a density maps, and the total number of neurons within an image patch can be computed. (C) Locations of nine areas selected for a comparison between the CB+ densities estimated by the U-Net CNN and multiple human annotators (V1, V2, MT, AuA1, A3b, A3a, A4ab, A8C, and A13M, see S1 File for a list of areas, color coding and abbreviations). (D) Per-area (i.e. average values for all boxes sampled from a given area) densities for the three analyzed hemispheres. Different symbols show the results for individual human experts (dM1 to dM3), an average of the three expert observers (), and the densities obtained with the U-Net CNN (dCNN). The mean of the differences between dCNN and the dM densities is statistically indifferent from zero (see S3F and S3G Fig for statistical details), indicating that the automatic and the average manual counts are indistinguishable. (E) The results obtained by the U-Net CNN reflect the consensus between the expert annotators. (rows, top to bottom) Examples of individual CB+ neurons marked by all, two, and only one expert, respectively, and a proportional density estimate by the U-Net CNN. The proposed method helps alleviate the interindividual variability of manual cell counting. Box size: 50 μm. (F) We applied the procedures for estimating the density of CB+ neurons to all CB-stained sections in all three hemispheres studied. Here, results for an example section (CJ1741-r16c) are presented. From left to right: section image, segmentation of the cortex into individual areas based on manual parcellation and coregistration to the reference template [35] (see Materials and Methods for details), and density map of CB+ neurons (see G for scale). Note that the quantification of the results is performed only in the cortical areas, while the subcortical regions are not considered. (G) Example three-dimensional reconstruction of a CB+ density map constituting the basis of the flat maps illustrated in other figures. The black line indicates the location of the section presented in panel F, and the black patches correspond to the parts of cortex excluded from the analyses due to staining artifacts or corrupted tissue. The datasets are available for download from https://www.marmosetbrain.org/whole_brain_cb_maps.