Machine Learning Approaches for Multi-Class Classification of Skin Cancer

Skin Cancer is the abnormal growth of skin cells that are usually exposed to the sun. The chances of receiving successful skin cancer therapy are highest when the disease is discovered early. Hence, to aid the health workers Machine Learning (ML) and Deep Learning (DL) techniques are useful for medical diagnostics. These Models could help in diagnosing benign and malignant cells and also predict the different classes of cancer with high accuracy. This study compares such two effective ML techniques to classify different cancerous skin lesions, namely (i) Support Vector Machine (SVM) classification with Grey Level Co-occurrence Matrix (GLCM) feature input, and (ii) A CNN architecture. Before classification using either method, the images are preprocessed to achieve the balance in the imbalanced dataset used. For SVM classification, GLCM features are extracted from the skin patches. But, before that, the skin patch images were subjected to the Dull Razor algorithm to remove the hair image present to enhance the image quality followed by augmentation to increase the image numbers for training. The CNN architecture used the Synthetic minority over-sampling technique (SMOTE) technique combined with the Edited Nearest Neighbor(ENN) i.e. SMOTE-ENN algorithm to achieve the balance in the number of images. From the study, the GLCM feature extraction technique showed an accuracy of 98 % with SVM, and an accuracy of 99.30% through the CNN architecture. The lowering of SVM accuracy is due to insufficient training for fewer images in certain classes, which was overcome in the preprocessing method applied before classification using CNN.