A central aspect of this effort is a detailed 3D map of the brain, derived from many thousands of histological brain slices imaged at ultrahigh resolution with modern microscopes. Mapping brain areas is a very time consuming, semi-automatic process that involves analysing complex patterns of cell distributions in different independent subjects. Scientists at the research centre in Jülich/Germany aim at creating a new generation of brain mapping tools that exploit the most advanced High Throughput Imaging devices, Machine Learning algorithms, and High Performance Computing (HPC) infrastructures available today.
They have trained a deep convolutional neural network (CNN) to classify texture in microscopic scans of brain tissue into different brain areas. The network learns precise texture features from existing annotations in microscopic images, and combines them with information from existing atlases.
The use of this modern HPC infrastructure enables the algorithm to process many large chunks of image data in order to capture both the cellular detail and spatial context. Without HPC, running the network would be almost infeasible. To this end, the neuroscientists and data analysts have worked closely with the JUELICH supercomputing center (JSC) to run the application at scale on the GPU-accelerated clusters JURECA and JURON.