Training a learning vector quantization network for biomedical classification

Abstract

A competitive learning vector quantization (LVQ) neural net (ANN) was trained to identify third stage parasitic strongyle larvae from domestic animals on the basis of quantitative data obtained from processed digital images of larvae. Various novel quantitative features obtained from processed digital images of larvae were tested whether they are variant or invariant to the shape taken by the motile larvae during image recording. A total of 255 images of 57 larvae of 5 genera were recorded. 16 novel features were measured of which 7 were selected as invariant to larva shape. By trial and error two of those features, area and perimeter, along with the quantitative features used in conventional identification, overall body length, width, and tail of sheath, were used as an effective training data set for the ANN. This ANN coupled with an image analysis facility and a knowledge relational database became the basis for developing a larva identification system whose overall identification performance was 91.9%. It is fast and objective. Its objectivity is based on the fact that it is not subject to inter- and intra-observer variability arising from the user's profile of competency in interpreting subjective and nonquantifiable descriptions. The limitations of the system are that it cannot handle raw images but only data extracted from images, its performance depends on the reliability of the input vectors used as training data, and its use is restricted to well equipped laboratories since its uses expensive instrumentation.