A Survey on Feature Extraction Methods for Skin Cancer

Cancer is one of the most lethal diseases in the world by which many people lost their lives. Cancer is caused by a gene mutation in which the cell grows abnormally. There are different types of cancer like blood cancer, lung cancer, brain cancer, skin cancer and much more [1, 2]. In a human body, the skin is the biggest organ. It covers the bones, muscles and all the other parts of the body. Skin elements in the human body have more noteworthy significance in light of the fact that; a little change in its working may affect other parts and organs in the human body [1,31]. Skin is presented to external surroundings. Consequently, illnesses and contamination happen more to the skin. In this way, we need to give a more noteworthy consideration regarding skin disease. In the last few decades, skin cancer is growing rapidly around the world [1,5,33].

Melanoma is the most precarious type of skin cancer in the world. According to the American cancer society in the year 2018, the estimated new cases of melanoma are 91,270 and the death rate is estimated to be 9,320[2]. Melanoma is caused by the unusual growth of skin cells called melanocytes [1]. One of the main reasons for this cancer is UV radiation and tanning beds. Melanocytes produce melanin which gives colour to our skin and also protects from harmful UV radiations [31]. The majority of melanomas are black or brown, but they can also be skin-coloured, pink, red, purple, and blue or white [31]. If melanoma is recognized and treated early, it is almost always curable, but if it is not, cancer can advance and spread to other parts of the body, where it becomes hard to treat and can be fatal [1,31].

In recent years, the fusion of dermoscopy techniques with the CAD system has become an important research field as it helps medical practitioners in gaining meaningful information from images. Later, they can be used to accurately identify melanoma [19]. Although in recent years the use of multiple methods in the diagnosis of melanoma disease has been shown to be rapid, the evaluation of the regional structures, such as region borders, pigment network, dots/globules, and streaks, can be complicated and require more time and attention [4,33]. 

It is believed that accurate and effective identification of these structures may help in early identification of melanoma [1,32]. Due to the poor quality of images and artefact present in images many segmentation and feature extraction approaches fail to extract the whole and complete structures [25,32]. Therefore, effective segmentation and feature extraction methods are very much required for these structures in order to improve the diagnosis and reduce the time required in the diagnosis process.

Feature extraction is one of the important steps in image processing. Hundreds of features can be acquired from each dermoscopy image and used as an image descriptor. However, not all are appropriate for lesion classification. Too many irrelevant features make the classifier complicated and require more computational time, which also reduces the classification accuracy. The best features have to be able to represent the characteristics of the regions in skin cancer images. Therefore, a suitable number of features should be extracted, with the best way possible to distinguish between images. Accordingly, melanoma images can be discriminated significantly from benign images by a classifier. On this basis, the segmented lesion images can be used to extract several numbers of features as the best way to address the region in isolation.

These features can be divided into four classes: i) handcrafted features, which are the most popular ones and comprise global image descriptors of shape, symmetry, colour, and texture; ii) dictionary-based features, where methods such as bag-of-features or sparse coding are used to obtain local descriptions of the skin lesions; iii) deep learning features that use convolutional neural networks to automatically learn good image representations, and iv) clinically inspired features that aim at attributing a medical meaning to the features used by the CAD system.

The Pattern analysis method was introduced by Pehamberger et. al in 1987 and then updated in 2000 in Consensus Meeting of Dermoscopy (CNMD (Pathan). Pattern analysis method has been used for feature extraction in the automatic diagnosis of skin lesion. (Olivia). This method is based on finding a specific pattern in the dermoscopic images and the pattern can be either global or local. In lesion images, global patterns are identified by textured shapes present in them. It includes reticular, globular, cobblestone, homogeneous, stardust, parallel, multicomponent, lacunar and unspecific patterns (Pathan Olivia).

Local patterns of dermoscopic structures include pigment systems, dabs and globules, lines, blue-white shroud, recessive structures, hypopigmentation, and vascular structures located in a specific region of the lesion. Pattern analysis consists of size checking, Uniformity, and distribution of the above patterns. The benign lesions are usually uniform. In other words, lesions do not present many patterns in their structure. Therefore, the presence of at least three (Multi-component), parallel or non-specific global patterns Indicates a high probability of melanoma (malignant). In addition, the presence of local patterns, Such as blue curtains and regression structures, or even Some patterns are considered atypical, irregular, or disproportionate can identify melanoma.

Menzies’s Method: This method is based on nine positive (Blue-white veil, Multiple brown dots, Pseudopods, Radial streaming, Scar-like depigmentation, Peripheral black dots-globules ,Multiple color, Multiple blue/grey dots, Broadened network) and two negative features (Point and axial symmetry of pigmentation, Presence of a single color) present in structure. A lesion is said to be melanoma when there is at least one or more positive features are present and none of the negative feature is present.