Pattern Recognition Methods for Oral Lesion Classification using Digital Color Images
The present thesis addresses the development and application of pattern recognition methods for classification of oral lesions using digital color images as input. The human oral cavity is a site of numerous diseases and two of the common and visually similar lesions are oral leukoplakia and oral lichenoid reactions. The lichenoid reactions, which may occur in different subclasses, are usually harmless lesions while leukoplakia can develop into cancer. An automatic detection of potentially precancerous lesions can enhance the diagnostic process and reduce the need for biopsy.
The problems studied represent a two-class classification problem (potentially precancerous vs. harmless lesions) and a four-class problem (complete classification into leukoplakia, atrophic, plaqueformed and reticular lichenoid reactions). Different classifiers are investigated, from classical Fisher linear discriminant to the novel Support Vector Machines. Different shape and color features are extracted from color images and evaluated as to their discrimination power. Using morphological and color features around 90% classification accuracy has been obtained on the two-class problem and 75% on the four-class problem. This corresponds to the performance of a very experienced oral specialist. The obtained 100% sensitivity for leukoplakia corresponded to 68% specificity for lichenoid reactions. Detection of the lesion boundaries was performed both manually and using Active Contour Models (snakes). The conclusion is that the system can be used as a decision support system and an educational tool in odontological practice. The methods developed have been applied to patients' images acquired from the Department of Oral Medicine, Faculty of Odontology, Göteborg University and Karlstad Central Hospital.
support vector machines