Automatic Analysis of Cortical Areas in Whole Brain Histological Sections using Convolutional Neural Networks

The segregation of the human brain in cytoarchitectonic areas is an important prerequisite for the allocation of functional imaging, physiological, connectivity, molecular and genetic data to structurally well-defined entities of the human brain. Cytoarchitecture describes the spatial distribution of cell bodies and their shape and size, and is most appropriately studied at microscopic resolution based on cell-body stained histological sections. To determine boundaries between cytoarchitectonic areas, an observer-independent method that uses image analysis and multivariate statistical tools to capture changes in the distribution of cell bodies is already established. Nowadays, new technologies for high-throughput microscopy allow rapid digitization of histological sections, which increases the need for a fully automatic brain area segmentation method. This task is extremely challenging due to the high interindividual variability in cortical folding, sectioning artifacts, limited labeled training data, and the need for large input sizes for automatic methods. This work shows that convolutional neural networks, a special class of deep artificial neural networks, are suitable for automatic brain area segmentation. It introduces a semantic segmentation model that combines texture input given by high-resolution extracts of the histological sections with prior knowledge given by an existing probabilistic brain area atlas, the JuBrain atlas. This atlas prior helps the model to localize the texture input in the brain and allows it to make topologically correct brain area predictions. To overcome the limited amount of brain area annotations, the model can be pre-trained on a modified task for which training data is easier to obtain. Pre-training the model on a self-supervised task based on predicting the spatial distance between image patches extracted from sections of the same brain significantly increases the segmentation performance and enables the prediction of several brain areas in previously unseen brains. The self-supervised model learns a compact internal feature representation of the input using the inherent structure of the brain, without having explicit access to the concept of brain areas. Extensive evaluations indicate that these features encode cytoarchitectonic properties. This is remarkable result which allows the data-driven analysis of the structure of the entire brain. Although the presented model is not yet robust enough for automatic annotation of all areas in complete human brains, it is already leveraged for practical use by training specialized multi-scale models to propagate brain area labels from annotated sections to spatially close sections. This workflow has the potential to speed up current brain mapping projects by reducing the workload of the neuroscientists and produces previously unattainable high-resolution 3D views of single brain areas