Classification of Melanoma (Skin Cancer) using Convolutional Neural Network

Background and Aim: The current state of art solution for detecting melanoma using Convolutional Neural network has not considered selection of only affected areas from the input images of skin lesion which has resulted in the unnecessary processing of non-affected skin parts and thus less accuracy. The aim of this research is to propose a new solution to solve the above issue by creating a bounding box around the affected areas and decrease the search space by regression technique which results in more accuracy for classification. Methodology: The proposed system consists of three parts. i) data augmentation ii) boundary extraction and iii) DCNN feature extraction and selection. In the boundary extraction part, exclusive or (XOR) is used with regression technique which creates the bounding box around the affected areas of skin lesion. It helps to reduce search space, improve the accuracy in terms of classification and reduce the processing time to extract the features. Results: The proposed system here is tested on PH2, ISBI 2016 and 2017 datasets which has increased approx. 1.2 % of accuracy compared to state-of-art solution.Conclusions: The proposed system has outperformed the current best solution. Whereas, the difference is quite low, so can be further improve by testing other type of CNN network and classifiers.