Melanoma is the most deathful of all skin cancers and the number of cases grows every year. The extirpation in early phases implies a high degree of survival so it is fundamental to diagnose it as soon as possible. In this paper we present a clinical decision support system for melanoma diagnosis using as input an image set of the skin lesion to be diagnosed. The system analyses the image sequence to extract the affected area, determinates the characteristics which indicate the degree of damage and, according to them, it makes a decision. Several methods of classification are proposed: a multilayered perceptron, a Bayesian classifier and the algorithm of the K nearest neighbours. These methods work independently and also in combination making a collaborative decision support system. The classification rates obtained are around 87%.
The incidence of melanoma skin cancer has been increasing over the past decades. Since the early 1970s, melanoma incidence has increased significantly, for example an average 4% every year in the United States. Currently, 132,000 melanoma skin cancers occur globally each year (World Health Organization, 2010).
Melanoma is the most deathful of all skin cancers. The main cause of melanoma is due to a long exposition to ultraviolet radiations, although skin type or other genetic factors can influence too. The most effective treatment is an immediate extirpation, but just when the melanoma had been detected in early phases (Geller, Swetter, Brooks, Demierre, & Yaroch, 2007). In other cases, if it is not diagnosed in time, the life expectancy is reduced up to less than one year. Therefore, it is fundamental to distinguish as soon as possible between benign lesions (as a simple spot or a mole) and melanomas. Dermatologists employ for their diagnosis several techniques which have been developed based on experience, among which it is emphasized to obtain the total dermatoscopy score based on the mnemonic ABCD (Stolz, Rieman, & Cognetta, 1994), the rule of 7 points (Argenziano et al., 1998) and the method of Menzies (Menzies, Ingvar, Crotty, & McCarthy, 1996). All these techniques allow identifying symptom of a malignant lesion based on the observation of a set of characteristics using dermoscopy images. Even so, in some cases it could be a hard task, the interpretation of these properties visually, and therefore, to make a right diagnosis.
We propose a clinical decision support system that classifies images with suspicious skin lesions in order to manage a referral list to the specialist. The patients whose images present a high probability of being a melanoma will be referred to the dermatologist as soon as possible. There are other clinical decision support systems implemented with this idea of managing a referral list; one example is the ERA (Early Referral Application), a system to support family doctors in identifying patients with suspected cancer that should be referred to a specialist in a short time period (Coiera, 2003). The system can also be used by a dermatologist as a second expert opinion to complement and compare his decision. It is important to bear in mind that the clinical diagnosis of melanoma depends on the dermatologist’s experience (Kittler, Pehamberger, Wolf, & Binder, 2002) so a second opinion can be important for a dermatologist in a training period or at the beginning of the professional activity.
The system bases its automatic diagnosis in three phases: detection, description and classification of the lesion (Fig. 1). In the first phase a preprocessing of image is done which allows us to identify the affected area. Afterwards, in the description phase, the image to determinate the optimum set of characteristics which indicate the degree of malignant tissue is analysed. Finally, this characteristics vector is used as the input of an artificial entity that is able to offer a diagnosis of the lesion, classifying it as a melanoma or a benign lesion. The entities proposed here are the method of the K nearest neighbours, a parametric classifier based on the decision theory of Bayes, a multilayer perceptron and the combination of these three methods in a voting collaborative system.
We have presented in this paper a clinical decision support system for melanoma diagnosis using several classification methods working individually and in a collaborative way. We base its logic in a first stage of preprocessing and segmentation by means of Otsu thresholding method. Afterwards, from a set of 24 characteristics that define the injury of the lesion, we obtain an optimum subgroup of six attributes. We have implemented three methods in order to do a classification of the images of skin lesions: the algorithm of the K nearest neighbours, a statistics classifier (Bayesian classifier) and a multilayer perceptron. The system, from the subgroup of six descriptors, makes the diagnosis of the input image in two possible classes: benign lesion or melanoma.
In order to improve the classification rates, we apply the algorithms individually and in a collaborative way thanks to a voting system: each algorithm gives a possible diagnosis decision with a reliability percentage and the system reaches a consensus. The classification rate (87.76%) obtained in the collaborative method is higher than in each method individually.
