Improving a bag of words approach for skin cancer detection in dermoscopic images

With a rapidly increasing incidence of melanoma skin cancer, there is a need for decision support systems to detect it in its early stages, which would lead to better decisions in treating it successfully. However, developing such systems is still a challenging task for researchers. Several Computer Aided-Diagnosis (CAD) systems have been proposed in the last two decades to increase the accuracy of melanoma detection. Image feature extraction is a critical step in differentiating between melanoma and normal skin lesions. In this paper, we propose to improve a bag-of-words approach by combining features consisting of the color histogram and first order moments with the Histogram of Oriented Gradients (HOG). Experimental results show that the proposed technique significantly improves the detection accuracy, with an average sensitivity of 91% and specificity of 85%. The proposed system was validated on a dataset of 200 medically annotated images (40 melanomas and 160 non-melanomas) obtained from the database of the Hospital Pedro Hispano. [1].