A BORDER IRREGULARITY MEASURE USING HIDDEN MARKOV MODELS AS A MALIGNANT MELANOMA PREDICTOR B. S. Aribisala and Ela Claridge School of Computer Sciences The University or Birmingham Birmingham B15 2TT, U.K. ABSTRACT Malignant melanoma, a skin cancer, manifests itself as a dark lesion, most often with an irregular boundary. The degree of irregularity is an important diagnostic indica- tor. This paper presents a new measure of irregularity us- ing Hidden Markov Models (HMMs) based on the Weibull probability distribution. The measure was tested on 98 skin lesions of which 16 were malignant melanoma. The ROC analysis showed that the measure is 82% sensitive and 82% specific in discriminating the malignant and benign lesions. These results compare favourably with other measures and indicate that HMM captures some distinguishing features in the boundary of malignant lesions. KEY WORDS Irregularity measure, Hidden Markov Models, Malignant melanoma, Weibull distribution, Skin lesions 1 Introduction Melanoma is a malignant tumour of melanocytes. The tu- mour initially starts from the upper skin layer (epidermis) and later invades the dermis below. The survival rate for pa- tients is inversely proportional to the depth of the tumour. Early detection of melanoma is the most important factor affecting the survival of a patient. Malignant melanoma can be characterised using some physical features such as shape, edge, colour and surface texture. The border irregularity of pigmented skin lesions was identified by Keefe et al.[1] as the most significant diagnostic factor in clinical diagnosis of melanoma. Re- search by Morris Smith [2] revealed that “irregularity” is one of the major vocabulary terms used for describing bor- der of the malignant melanoma in medical textbooks. The same research also showed that the clinicians place a sig- nificant emphasis on border irregularity when describing malignant melanoma. These findings coupled with the fact that border irregularity is one of the major features in the seven point checklist used for computing a “suspicious- ness” score for skin lesions [3] indicate that border irreg- ularity is a very important factor in the diagnosis of malig- nant melanoma. Email: B.S.Aribisala@cs.bham.ac.uk Email: E.Claridge@cs.bham.ac.uk It has been empirically discovered that clinicians have difficulties in visually assessing border irregularity of skin lesion outlines and that their assessments are not invari- ant to reflection and rotation [2, 4]. Much research on quantitative measures of irregularity has been carried out to overcome these shortcomings[2]. The most common ap- proaches include the Compactness Index (e.g. [5]), Fractal Dimension (e.g. [6]) and measures based on radial distance (e.g. [7]). These methods are critically reviewed in a recent paper by Lee et al [8]. Claridge et al. [4, 9] proposed the use of fractal dimension (FD) and a modified fractal dimension called Structural Fractal Dimension( SFD). FD is not sensitive to the structural features such as indentations and protrusions [8]. SFD was designed to remove this drawback but its re- sponse to single indentations and protrusions, which can be indicative or early melanomas, was too week to enable reliable classification. [8, 10]. Sigma Ratio and Indentation Irregularity Index, pro- posed by Lee and Atkins [8, 10], are non-linear measures derived by a curvature-scale filtering. They have been shown to be successful in detection of indentations and pro- trusions in the lesion border and have been the most suc- cessful algorithm to date for classifying skin lesions on the basis of their border irregularity. The term “Irregularity” is intuitive and can express different meanings. If irregularity is to be quantified, it is necessary first to develop its formal definition, or at least provide its formal description. Five attributes of irregular- ity have been proposed [11]. One of these attributes which is of interest here is lack of predictability. The elements of a sequence corresponding to a regular shape or pattern are predictable, whereas in an irregular sequence they cannot be easily predicted. That is, the extent to which a sequence can be predicted may help us to determine how regular the sequence is. In this paper we present a new measure of irregularity based on Hidden Markov Model. In contrast to the existing measures, the proposed measure is based on a formal cri- terion of irregularity outlined above. The model is trained on idealised regular lesion shapes represented as ellipses. If a given lesion is regular, its shape will be represented by the model. The degree to which a given lesion boundary conforms to the model is evaluated by computing the log- likelihood of the outline given the model. This measure