Melanoma Literature Review

Automated Melanoma Classification: A Literature Review
Introduction
A Melanoma is a skin cancer with the highest mortality rate and one of the most common cancers in young adults. Since 1973 the incidences of melanomas has increased by 150% and the mortality rate by 44%. Although melanoma survival rates are poorer than those of other skin cancers, if detected earlier their treatment is more effective (Jain & Jain, 2012). However there is a significant number of cases where melanomas, some of which could have been removed though excisions, were not diagnosed correctly. The diagnostic sensitivity ranges from 66% to 81% for dermatologist and it is believed to be lower for nondermatologists (Stanley et al., 2008).
A lot of research conducted over

This task is even often difficult for trained dermatologists, because of the fact that the lesion often transition smoothly to the skin. In addition some dermoscopy images contain artifacts such as uneven lumination, gel, black, frames, ink markings, measuring devices, air bubbles, hair, vessels etc. These elements complicate segmenting the lesion from the skin. Preprocessing is therefore usually required to remove most of these elements (Masood & Ali Al-Jumaily, 2013).
Various techniques have been developed to facilitate this process such as image resizing, masking, cropping, attenuation, and conversion to grayscale. A straightforward general purpose method is to apply a filter such as peer group filter (Celebi et al., 2005), mean filter (Silveira & Marques, 2008), media filter (Celebi et al., 2007), Gaussian filter (Maglogiannis et al., 2006), or anisotropic diffusion filters. These filters are however not guaranteed to remove all artefacts. An alternative is to use a specialized method for each approach. Many methods have been suggested, but none deal with all the artefacts involved (Abbas et al., 2011) (Celebi et al., 2009) (Rashid et al.,

More advanced methods have been used where global thresholding are combined with adaptive thresholding along with colour clustering (Ganster et al., 2001). A double thresholding techniques has also been proposed that claims to be the simplest, most accurate technique to date (Jain & Jain, 2012). The key parameters can be fixed by a fitted-curve of the RGB component histogram.
Edge-based methods: This methods focusses on detecting edges on the contours of images. It however struggles when blurring and smooth transitions between skin and lesion leads to broken contours. It also have difficulty when in the presence of artefacts or even irregular skin textures (Erkol et al., 2005).
Region-based methods: In this method performs in a 2 phase approach. Firstly the image is divided in many regions based on similar intensity levels. Then regions are merged on some criteria such as similarity in hue. It is difficult to determine the appropriate parameters for grouping and merging. Strict parameters can cause fragmentations while relax parameters can cause overmerging (Celebi et al., 2005). Examples of region based methods include multiscale region growing, the modified fuzzy c-means algorithm which is orientation