O. Schmitt 1, S. Bethke 2,
P. Sobe 2, E. Maehle 2
1 Institute of Anatomy, University of Rostock, Rostock, Germany
2 Institute of Technical Informatics, University of Lübeck, Germany
SEGMENTATION AND RECONSTRUCTION OF BLOCK FACE IMAGES OF A COMPLETELY SECTIONED HUMAN BRAIN
The investigation of normal brains, resp. the human brain, at microscopic resolutions has become a demand of biological geno- and phenotyping to understand and match genetic changes with changes in the phenotype from the micro- and meso- up to the macroscopic scale. After performing a structural and functional localization study by MRI and fMRI data can be linked with single cell information obtained by analysis of all neurons in high resolution images. This linkage of functional and high resolution structural information offers a promising way for a deeper insight in the organization of neurons and their connectivity in terms of behavior. Recently we have realized (Schmitt et al., 2004, NeuroImage 22 Suppl. 1, TH330), a microregistration of a whole mouse brain and it was shown that efficient algorithms are able to register large neurons and cell clusters in images of whole histologic sections. However, deformations in a reference free registration approach can be reduced but not excluded. Therefore, it is important to perform coregistrations of the deformed histologic sections and an appropriate reference. Usually, such references are derived from MRI. The disadvantage of using MRI reference images of a post mortem brain is the low spatial resolution yielding to interpolation problems when coregistrating histologic sections exhibiting a resolution of about the twentyfold. Therefore, block face (Toga, 1994, J Neurosci Meth, 239-252), i. e. episcopic, images of a human brain were generated before a section of the paraffin embedded brain was achieved. This was performed 6214 times, resp., for each section. Each image has a size of 1352 x 1795 pixel and a dynamic range of 24 Bit. Due to transparency of the wax phase, slightly varying illumination and focussing of the transparent material segmentation of the foreground is a complex task that can not be solved by conventional approaches (segmentation by threshold, adaptive threshold, level sets, cluster techniques, region growing, active contours, Markov random fields). The objective of this work is to determine an optimal preprocessing chain, to implement the adaptive fuzzy c-means (AFCM) (Pham and Price, 1998, IEEE Trans Med Imag, 737-752) and a noise tolerant modification of the AFCM algorithm (Pham and Price, 2001, IEEE CBMS, 127-131) on a Sun Fire parallel computer (sf15k) equipped with a 64 Bit shared memory architecture, 72 GB RAM and 72 CPUs. Using an optimal set of parameters the parallel implementation segments an image in 6 minutes in comparison to 292 minutes of a sequential computation. The noise tolerant AFCM method produces slightly better results within small sulci. Because small changes of the 11 parameters can produce large differences of the segmentation results an automatic parameter optimization procedure was implemented. After segmenting all images this parameter optimization searches for large segmentation differences between segmented images based on sum squared distances (SSD) and performs a minimization of the SSD with respect to the over- or under-segmented images. The segmented images were 3D-reconstructed and the surface has been visualized. We found only small resting volumes of under- or over-segmentations which can be targeted now with the 3D-implementation of the noise tolerant AFCM technique (Pham and Price, 1999, IPMI, 140-153). The current segmented dataset is a promising starting point for a high resolution coregistration of the digitized histologic sections (Schmitt et al., 1999, NeuroImage 6: S22).