Abstract

Malignant melanoma is the deadliest form of skin cancer and must be diagnosed and excised during its earliest stages. The development of computerised systems which accurately quantify features representative of this cancer aims to assist diagnosis and improve preoperative diagnostic accuracy. One clinical feature suggestive of malignancy is asymmetry, which considers lesion shape, colour distribution and texture. In this paper techniques for the detection of colour asymmetry are evaluated and a new method for visually displaying and quantifying colour asymmetry is proposed. Automatic induction methods and a neural network model are utilised to evaluate the diagnostic capability of our features and identify those of greatest relative importance. Results indicate that those features quantifying possible areas of regression are most indicative of colour asymmetry.