Parallelized segmentation of a serially sectioned whole human brain

Building a digital three dimensional representation of a human brain is a challenging task. Such a model provides insights into the microstructure of cortical layering and columns. The presented work is based on a complete dissected and preserved human brain that has been serially sectioned at a coronal resolution that is suitable for single cell detection. More than 6000 sections have been generated and exist as digital images. To obtain a valuable three dimensional representation, morphology preserving affine linear and nonlinear registration schemes are necessary steps. To rebuild a serially sectioned brain, reference images derived from a non deformed object, e.g., MRI or block face images, are necessary for a faithful affine linear and nonlinear registration. In the case of block face images the brain regions must be separated from highly variable background regions to obtain a suitable stack of segmentation images. Among the image segmentation algorithms we found fuzzy c-means techniques as a promising starting point for a sophisticated segmentation framework of either gray level or color images within 2- and 3-dimensions.With respect to algorithmic complexity and computation cost, two fuzzy c-means algorithms were implemented. A proper image preprocessing strategy turned out to be necessary for accurate and robust segmentation results. Primarily, the algorithms work in a parametric resp. supervised mode. Additionally, an automatic mode helps to explore the parameter space within a reasonable range and to compare the segmentation result with an optimal one, provided by an expert. By minimizing the differences we can set up parameters that are used for series of adjacent images. So, it is possible to obtain optimal segmentations independent of illumination disturbances, artifacts and defocusing.We present a complete high resolution and accurate segmentation of the first complete human brain that was sectioned, photographed and digitized at histologic resolution. Based on these images, a succeeding 3D representation is presented. Finally, a segmented and spatially correct straightened data set is available now for coregistration tasks together with the high resolution histologic data set.