2. Related works

Several diagnostic systems for melanoma detection have been proposed. Some systems try to imitate the performance of dermatologists by detecting and extracting several dermoscopic features. These features can then be used to score a lesion in a similar way to the one adopted by dermatologists. In [7], general clinical principles of early melanoma detection are reviewed, providing the clinician with an up-to-date understanding of management strategies for their patients with numerous or atypical nevi. Many researchers have been working on the image processing and computer vision techniques for skin cancer detection. The most probably features to perform skin lesion segmentation used in various papers are shape, color, texture, and luminance.

Three methods of segmentation have been discussed by [8]. The methods are: Otsu's method, gradient vector flow (GVF), and color based using K-mean clustering. Feature extraction is based on the so-called ABCD-rule of dermatoscopy. While [9], a watershed segmentation is the proposed scheme used for image segmentation, border detection and decision related with structural nature of lesion. For more details of study, the segmentation methods used, a survey work on skin lesion segmentation problem implemented by image processing techniques are described by [10–12].

Based on a qualitative assessment of asymmetry (of boundary, color, and mass distribution), size functions (SFs) and support vector machine (SVM) are used to implement a new automatic classifier of melanocytic lesions [13]. An automatic identification of asymmetry in digital images containing melanocytic skin lesion using Stolz strategy, based on the ABCD rule is proposed by [14]. A survey on asymmetry analysis of malignant melanoma using image processing techniques to identify the asymmetricity of the melanoma skin lesions was presented by [15].

Several researchers proposed an image analysis tools to check for the various melanoma parameters like asymmetry, border, color, diameter, in terms of texture, size, and shape analysis for image segmentation and feature stages. The extracted feature parameters are used to classify the image as normal skin and melanoma cancer lesion [16–22].

In [23], they applied a Bag-of-features approach to malignant melanoma detection based on epiluminescence microscopy imaging (low-power microscopy (˜50– 100), commonly a television microscope applied to a glass slide covering mineral oil on the surface of a skin lesion, to determine malignancy in pigmented lesions). Each skin lesion is represented by a histogram of code words or clusters identified from a training data set. Classification results are achieved based on the implementation of naive Bayes and support vector machine classifiers. Other work utilized the Bag-offeature model for the detection of melanomas in dermoscopy images and aimed at identifying the role of different local texture and color descriptors [24]. The reported results show that the sensitivity is 93% and specificity is 85%.

The extracted features of segmented lesions used as inputs to the input layer of the artificial neural network. Different configurations of ANNs were implemented by the researchers for classification [25–28]. In [25], they attached the Dermlite® DL1 dermatoscope to the iPhone. A new method called elliptical symmetry was proposed for quantifying asymmetry. Gaussian smoothing and lacunarity analysis to measure border irregularity were proposed. In Gaussian smoothing, the contour was smoothed and compared with the perimeter of the original lesion. The lacunarity was used to analysis the borders of the image. Finally, the extracted features were fed to input layer of the multi-stage neural network classifier. While [26], 2D-Wavelet transform is the feature extraction method used. These features are given as the input to the artificial neural network classifier. An unsupervised

approach for lesion segmentation is proposed by [27]. Iterative thresholding is applied to initialize level set automatically. The accuracy of detected border is compared with Growcut and mean-shift algorithms. Four features relying on visual diagnosis: asymmetry (A), border (B), color (C), and diameter (D) are computed and used to construct a classification module based on artificial neural network for the recognition of malignant melanoma. The authors of [28] used a hybrid algorithm combining a region-oriented and a thresholding method to segment the lesion. A multilayer perceptron NN model with one hidden layer and one output neuron was chosen as a basis for all the different network configurations examined.

As described by [29], the general approach used by a CAD system consists in describing the skin lesion by means of a set of textural and geometrical shape features known as the ABCD rule (asymmetry, border, color and diameter). Software WEKA was used to apply 13 different techniques and a statistical test K-folds to obtain the classification accuracy.

A different approach proposed by [30] named Modified Texture Distinctiveness Lesion Segmentation algorithm (M-TDLS) to segment the skin lesion. Two steps are involved: TD metric calculation and region classification. The RGB image is converted into XYZ color space and the TD metric is calculated to find dissimilarity between two texture distributions.

In [31], they addressed two different systems for the detection of melanomas in dermoscopy images. The first system used global methods to classify skin lesions, whereas the second system used local features and the bag-of-features classifier.
