Towards Higher Accuracy in Diagnosis of Skin Cancer: An Adaptive CNN-RF Model for Diagnosis of Skin Cancer based-on Different Oversampling Methods

Abstract

One of the most hazardous diseases in the world is skin cancer. Convolutional neural networks have recently received further interest for their use in spotting skin malignancies in dermoscopy images. In this paper, a new hybrid model based on CNN and random forest algorithm is proposed for detecting and classifying skin cancer images. The dataset used for this study is based on the HAM10000 dataset. Dataset was first preprocessed and different oversampling methods were applied to overcome class imbalance. Then, features of the preprocessed dataset were extracted using customized CNN. Finally, these features were classified using RF algorithm. The effective hyper-parameters for CNN and RF were detected besides different batch sizes and image sizes were implemented to ensure consistency of the proposed model. The proposed model has the ability to achieve 98.88 % for accuracy, 0.99 for precision, 0.99 for recall and 0.9999 for AUC.