Artificial Intelligence for Classifying Oral Lesions
Abstract from thesisThis research, initiated because of concerns about the diagnosis and follow-up of oral lesions in general practice, has been directed towards encouraging their management along "best practice" guidelines by developing an intelligent system as a clinical aid. The hypothesis is that digital images of oral lesions contain information that, using artificial intelligence techniques, can classify the source lesion thus reinforcing diagnosis and increasing the timeliness of referrals to specialist departments for treatment.
The first research target, automatic localization of a candidate lesion in a clinical image, has been met with a feasible method, currently unproven, for automatic derivation of threshold values for segmentation to localize candidate objects. However, practical use of the clinical endoscope requires its direction at the candidate lesion and to emulate this in the research localization has been pragmatic, using manual methods.
The second target, derivation of data suitable for use in artificial neural network techniques, was achieved by devising new sampling techniques to obtain multiple representative vectors from each localized image.
The third and final target, development of an intelligent system for classification of the vector data, was achieved when test vectors were successfully separated into two classes - the definitely benign and the possibly malignant - using a Kohonen and a radial basis function (RBF) network. These successful networks were refined, aiming for more specific identification of individual tissues, and it was found that hierarchical combination of networks was needed for success. Three different combinations were demonstrated, two hierarchies using Kohonen and multilayer perceptron (MLP) networks and one using RBF, Kohonen and MLP networks, each successfully separating the four classes of pathology - Major and Minor Recurrent Aphthous Stomatitis, Reticular Lichen Planus and the ‘rare event’ of Squamous Cell Carcinoma.
These intelligent systems satisfy the objective of the research and provide promising clinical solutions.
Return to Adrian Hopgood's page.
| Adrian Hopgood|