Three-Dimensional Vessel Reconstruction from Microscopic Image Sequence

Image registration is vital to medical image analysis. It is frequently used in 2D mosaicing to construct the whole image of a biological specimen and in 3D reconstruction to build up the structure of a specimen from a series of microscopic images. Nevertheless, many factors, including microscopic optics, mechanical factors, sensors, and manipulation may contribute to great differences even between adjacent image slices. Cuts, tears, folds, and deformation can cause chromatic aberration as well as geometric discrepancies. These flaws can make the alignment of all image slices very difficult. In this paper, we adopt a featurebased registration method, called robust point matching (RPM), to reconstruct 3D vessels automatically from rat brains using a series of microscopic images. The registration algorithm simultaneously evaluates spatial correspondence and geometric transformation between two point sets corresponding to the extracted feature points from two adjacent slices. Using this robust statistical mechanism, the deterministic annealing method automatically retains matched feature point pairs and removes unmatched ones as outliers. Experimental studies have shown promising results on 3D vessel reconstruction of rat brains and demonstrate accurate registration from microscopic image sequences.