Skin Cancer Classification Through Quantized Color Features and Generative Adversarial Network

Abstract

Early interpretation of skin cancer through computer-aided diagnosis (CAD) tools reduced the intricacy of the treatments as it can attain a 95% recovery rate. To frame up with computer-aided diagnosis system, scientists adopted various artificial intelligence (AI) designed to receive the best classifiers among these diverse features. This investigation covers traditional color-based texture, shape, and statistical features of melanoma skin lesion and contrasted with suggested methods and approaches. The quantized color feature set of 4992 traits were pre-processed before training the model. The experimental images have combined images of naevus (1500), melanoma (1000), and basal cell carcinoma (500). The proposed methods handled issues like class imbalanced with generative adversarial networks (GAN). The recommended color quantization method with synthetic data generation increased the accuracy of the popular machine learning models as it gives an accuracy of 97.08% in random forest. The proposed model preserves a decent accuracy with KNN, adaboost, and gradient boosting.