A Neural Spot Counter

In this paper we present an alternative approach to the problem of spot counting in interphase cell nuclei. Spot counting is a technique to detect defects of chromosomes in cells. By counting the number of spots which are labeled (read coloured) chromosomes one is able to indicate whether this cell has abberation that indicates a serious illness. We will not focus on the staining techniques used to colour the diierent chromosomes and how they are processed by the scanning system, but it is assumed that we have some preparated specimen. These specimen are scanned using a microscope system and resulting in computer images of the single cell nuclei (see for instance Netten 1994]). These single cell images are used to detect and count spots. The existing dot counting system, ie. a system consisting of a microscope, CCD camera and computer to scan and process digitized images, extracts from these single cell images the number of dots. We propose a neural network to perform the counting task. This neural network is trained to perform the dot counting on a single cell image. The neural network we use is a standard backpropagation network that is trained using the Matlab neural toolbox, with an adaptive learning scheme and momentum. For training a real-life dataset and an artiicial dataset is used. The artiicial dataset is used to be able to train a neural network without being bothered by the small sample size eeects we have in real-life. This set is generated using standard prooles of cell nuclei and spots. By using such a dataset we are able to get an indication of how the network will perform on the real-life dataset. Using a neural network instead of a traditional image processing approach for the spot counting is that the network nds \soft decision boundaries". This means that the output of the network indicates which is the most probable class a cell belongs to. Classes we distinguish are the 0-spot, 1-spot, 2-spot upto the 5-spot class (ie. 6-classes). The class that has the largest output is usually the winning class, ie. the class which is assigned to, but looking at the other outputs may reveal how certain that classiication is. When the outputs are equally high the network may not be able to decide whereas the current technique would have determined the number of spots without hesitation. Moreover by using our conndence estimating technique ((Tholen …