Machine Learning based Identification of Melanoma Skin Cancer using Fractional Coefficients of Cosine Transformed Dermoscopy Images
Melanoma skin cancer has proven deadly for human beings. Early detection of this cancer can be very helpful in treating patients; however, it requires doctors with great expertise for accurate identification. Computers supported diagnostics is now being used to assist doctors on a large scale for automated cancer detection due to the lack of experts. With the use of machine learning algorithms in image processing it is possible to speed up the process of early detection with accurate identification to a great extent. This paper attempts to use machine learning algorithms for identification of melanoma skin cancer using features of dermoscopy skin images expressed in form of fractional coefficients of transformed dermoscopy images through Cosine transformation of the affected area. The consideration of fractional coefficients helps in accelerating the identification process. The paper explores the features generated with the variations of 4*4, 8*8, 16*16, 32*32 and 64*64 high energy fractions of Cosine transformed dermoscopy images. The performance improves from 4*4 fraction and then gradually improve in performance is observed till the size of fraction increased up to 32*32, after which the marginal improvement is observed in fraction 64*64 of Cosine transformed dermoscopic skin images. The ensemble combination of machine learning algorithms is experimented to boost the performance than the individual machine learning algorithms with the exception of SVM classifier which performs better than the ensemble of classifiers in sensitivity, accuracy and specificity metrics. Overall, it is observed that from the average of the performance metrics as well as of all the variations of fractions, the combination of SVM-AD Tree-Random Forest classifiers give the best performance proving worth of proposed technique.