Malignant Melanoma Identification Using Best Visually Imperceptible Features from Dermofit Dataset

In this paper, an analysis of four different feature ranking algorithm is done and a novel approach of hybrid feature selection is proposed for better classification accuracy in malignant melanoma detection. In this work, at first totally 1898 features including geometrical shape, color and texture features are extracted from 1300 melanoma and non-melanoma images archived in Dermofit image library of the University of Edinburgh. Four feature ranking algorithms are used to rank the 1898 features in this work. Three classification algorithms, Support Vector Machine (SVM), K-Nearest Neighbors (KNN) and Ensemble Boosted Tree (EBT) are used as a classifier for the malignant melanoma and nevus lesion. Highest accuracy of 86.1% found with only 700 features selected by minimum correlation based feature ranking algorithm using Support Vector Machine (SVM). In the novel approach, those features are selected which have a higher rank in all the 4 feature ranking algorithms. With only 163 such features, an accuracy of 86.2% is found, which is similar in accuracy level with much less number of features in comparison with the result reported in the related work.