Neural Comput & Applic (2001)10:39–47  2001 Springer-Verlag London Limited A Comparison of State-of-the-Art Classification Techniques with Application to Cytogenetics Boaz Lerner and Neil D. Lawrence Computer Laboratory, University of Cambridge, Cambridge, UK Several state-of-the-art techniques – a neural net- work, Bayesian neural network, support vector machine and naive Bayesian classifier – are exper- imentally evaluated in discriminating fluorescence in situ hybridisation (FISH) signals. Highly-accurate classification of valid signals and artifacts of several cytogenetic probes (colours) is required for detecting abnormalities in FISH images. More than 3100 FISH signals are classified by each of the techniques into colour and as real or artifact with accuracies of around 98% and 88%, respectively. The results of the comparison also show a trade-off between simplicity represented by the naive Bayesian clas- sifier, and high classification performance rep- resented by the other techniques. Keywords: Bayesian neural network; Fluorescence in situ hybridisation (FISH); Multilayer perceptron; Naive Bayesian classifier; Signal classification; Support vector machine 1. Introduction In recent years, fluorescence in situ hybridisation (FISH) has emerged as one of the most significant new developments in the analysis of human chromo- somes. FISH offers numerous advantages compared with conventional cytogenetic techniques, since it allows numerical chromosome abnormalities to be detected during normal cell interphase. One of the most important applications of FISH is dot counting, i.e. the enumeration of signals (also called dots) within the nuclei, as the dots in the image represent Correspondence and offprint requests to: B. Lerner, Computer Laboratory, University of Cambridge, Cambridge CB2 3QG, UK. Email: boaz.lernercl.cam.ac.uk the inspected chromosomes. Dot counting is used for studying numerical chromosomal aberrations in, for example, haematopoietic neoplasia, various solid tumours, prenatal diagnosis and for demonstrating disease-related chromosomal translocations [1]. However, a major limitation of the FISH tech- nique for dot counting is the need to examine large numbers of cells. This is required for an accurate estimation of the distribution of chromosomes over cell population, especially in applications involving a relatively low frequency of abnormal cells. As visual evaluation by a trained cytogeneticist of large numbers of cells and enumeration of hybridisation signals is expensive and time-consuming, FISH analysis for dot counting can be expedited by using an automatic procedure [2]. One approach to dot counting relies on an auto- focusing microscope to select the ‘clearest’ image for the analysis [2]. However, basing dot counting on auto-focusing has some shortcomings [3]. Instead, it has been recently proposed [3] to base FISH dot counting on a Neural Network (NN) classifier, discriminating between in and out-of-focus images taken at different focal planes of the same Field- Of-View (FOV), as an alternative to the use of auto-focusing mechanism. Images at different focal planes of a specific FOV compose a stack of images that represents this FOV. Each stack image is ana- lysed, and its signals are classified by the NN as valid data or artifacts, which are the result of out- of-focusing. Following the discrimination of valid signals and artifacts in each stack image, the image that contains no artifacts is selected as the in-focus image to represent the stack (FOV), whereas the other stack out-of-focus images are rejected. The procedure is then repeats itself for other FOVs until the entire slide is covered or the required number of (in-focus) images (or nuclei) are collected. Pro-