Accurate Segmentation of Single Isolated Human Insulin Crystals for In-Situ Microscopy

An Algorithm is presented for accurate segmentation of single isolated human insulin crystals in an image captured by an in situ microscope inside of a bioreactor. It consists of three major steps. First, the foreground regions are extracted by thresholding the image. Then, the regions which correspond to the single isolated human insulin crystals are detected among the previously segmented foreground regions using a single nearest prototype rule, where each segmented foreground region class prototype represents a 7-dimensional mean vector of rotation, translation and scale invariant shape characteristics of several class members, which were extracted a priori from a training set of images. Finally, the contour accuracy of the detected regions is improved by moving each contour point to that image position where the weighted sum of the image intensity and the first and the second contour derivatives is minimal. The search of the minimum is carried out only along the line segment that goes from the region contour point position to the position of the region center of gravity. Experiments with 60 real images revealed very accurate segmentation results with an average contour accuracy of 1.61 +/- 2.53 pixel.