ABSTRACT
Analysis of chromosomes is an important and time-consuming task in the diagnosis of inherited or acquired genetic abnormality. Machine vision systems can contribute to the visual inspection of microscope images and the assignment of chromosomes to 24 classes is a critical stage in this analysis. A multilayer perceptron classifier has been developed for use in an automated chromosome analysis system. The inputs to the classifier are chromosome size, centromere position and a representation of the banding pattern measured from microscope images of dividing cells. The outputs are likelihoods of class membership. Optimum performance was obtained by factoring the classifier into two networks, one using size and centromere position alone to provide a first assignment into seven groups, followed by a second step in which the banding information was incorporated to give a final classification. The network is trained by backpropagation and considerable advantage is obtained by using a strategy of gain reduction using both total error and classification accuracy as network monitoring parameters. Classifier performance was tested on fairly large sets of chromosome measurements covering a representative range of data quality. Overall classification accuracy was found to equal or exceed that of a well developed statistical classifier applied to the same data.
