Multi-classifier decision fusion for enhancing melanoma recognition accuracy

This paper proposes an automated non-invasive multi-classifier system for skin cancer (melanoma) detection. The proposed system fuses the results obtained from three classification systems to enhance the melanoma detection rate. All of the classification systems use Support Vector Machine classifier. However, the image feature sets used in each classification system are different. The features sets used are Wavelets and Color features, Curvelets features and Grey Level Co-occurrence Matrices features. The output class labels or class probabilities of the three classification systems are combined using Majority Voting or Averaging Fusion to obtain enhanced classification rates. The dataset used include digital images for benign and malignant skin lesions. Experimental results show that the proposed multi-classifier fusion method outperforms standalone Skin Lesion classification systems in terms of recognition accuracy. Consequently, this can increase the chances of non-invasive melanoma detection from digital images.