Determining Important Features for Melanoma Classification Through Feature Selection

Abstract: Skin cancer is a common disease and malignant melanoma is the most dangerous form of it. Although dangerous, the survival rate of melanoma patients is high if the diagnosis is made at an early stage. Computer aided diagnostics has been shown to have potential in accurately diagnosing the disease utilizing machine learning. Thus, machine learning algorithms can be used to effectively classify a skin lesions as either benign or malignant. These algorithms can be made more accurate and efficient by applying feature selection since it decreases the dimensionality of the feature space. The aim of this study is to apply feature selection on four different classifiers to compare morphological and SIFT features in order to determine which features are important for classifying melanoma. The results show that morphological features in general had a higher importance than the SIFT features, although this varied between different classifiers. Furthermore, forward selection was more effective than backward selection in terms of accuracy for three out of the four classifiers. Lastly, two morphological features were significantly more important than the other features. The most effective feature measured the compactness of the lesion and the second most described the contrast between the lesion and the surrounding skin in terms of the color red.