Semi unsupervised M-FISH chromosome image classification

Abstract

M-FISH (Multiplex Fluorescent In Situ Hybridization) is a multichannel chromosome imaging technique that allows the color discrimination of human chromosomes. Although M-FISH facilitates the visual detection of chromosome rearrangements, the success of this technique largely depends on the accuracy of the pixel-by-pixel classification. In this study, we present a semi unsupervised method for M-FISH chromosome image classification. First, we segment the chromosome pixels using an automated thresholding procedure. Five features for each pixel are extracted, describing the intensity of the five channels. These features are then normalized. Second, we employed the K-means algorithm in order to cluster the chromosome pixels into the 24 chromosome classes. We have used the emission information for each chromosome class in order to initialize the cluster centers. Our method has been tested on the ADIR M-FISH database and an overall accuracy of 72.48% is achieved. This methods that use training set to build the classifiers, while our method does not use a training set.