Computer recognition of skin structures using discriminant and cluster analysis.

Abstract

BACKGROUND/AIMS: Automated image analysis of complex tissues is usually limited by the difficulty of recognizing special structures by computer. The aim of this study was to test the applicability of discriminant and cluster analysis to the interpretation of skin images. METHODS: Digital images from microscopic, dermatoscopic and clinical views of skin specimens were electronically dissected into elements of equal size and shape, and a set of grey level, colour and texture features was assessed for each element. Elements were classified interactively and submitted to discriminant analysis. Furthermore, hierarchical cluster analysis was used to enable the system to classify the tissue elements automatically, based on the available digital information. The classification results were relocated to the original image in order to evaluate the performance of the procedure. RESULTS: The system performs well in reproducibly detecting different skin structures in digital images. Discriminant analysis of interactively classified elements yielded a correct reclassification in 98 to 100% of tissue elements. Among the cluster analysis procedures, the conservative Ward method after removal of all highly correlated features produced the best results. The method turned out to be applicable irrespective of the image source used. CONCLUSIONS: Discriminant and cluster analysis may be helpful techniques for a user-independent, subjectively unbiased measurement system of skin structures.