Automatic segmentation of biomedical images

A new methodology which combines thresholding and probabilistic relaxation labelling process is proposed for automatic segmentation of biomedical electron micrograph images. The image is first thresholded and initial label probabilities of individual pixels are assigned according to the distance of each pixel to the clusters resulted from thresholding. The label probabilities are then estimated and updated iteratively by employing the relaxation labelling process. To reveal local details, the initial labelling process is localized to sub-images of the original image. A heuristic criterion is defined to determine the number of classes that exist in each sub-image. Computing cost and artifacts are greatly reduced if the process is implemented at multiple levels of resolution. The method has been applied to electron micrograph cell images successfully