A new SVM solver applied to Skin Lesion Classification

Abstract

We present a unified framework for solving the nonlinear Support Vector Machines (SVM) training problems. The framework is based on an objective function approximation so that the Problem becomes separable, with low computational cost root-finding methods to solve the resulting subproblems. Because of the diagonalization of the objective function in the first stage of the framework, we named the new SVM solver DiagSVM. To test the performance of the DiagSVM, we reported preliminary numerical experiments with benchmark datasets. From the results, we chose the best combination used in the framework to solve the Skin Lesion Classification (SLC) problem. Since melanoma (skin cancer) is the most dangerous and deadliest disease that affects the skin, the application of the DiagSVM can be integrated into several Computer-Aided Diagnosis (CAD) systems to help them detect skin cancer and significantly reduce both morbidity and mortality associated with this disease. Machine learning (ML) and deep learning (DL) based approaches have been widely used to develop robust skin lesion classification systems. For the SLC problem, three pre-trained convolutional neural networks (CNN), Xception, InceptionResNetV2 and DenseNet201, were employed as feature extractors and their dimension was reduced using Principal Component Analysis (PCA), Kernel Principal Component Analysis (KPCA) and Independent Component Analysis (ICA). Finally, the samples were fed into two SVM solvers: DiagSVM and Libsvm. The experiment shows that using PCA, KPCA, or ICA, the SVM can perform better than without feature reduction. The classification performance of the proposed methodology is analyzed on the ISIC2017 and PH2 datasets. The benchmark and SLC results indicate a promising proposal for accuracy, specificity and sensitivity metrics.