Multiple skin lesion classification using deep, ensemble, and shallow (DEnSha) neural networks approach

Numerous forms of skin diseases might attack the human body. The appropriate diagnoses and detection of the disease can decrease the severity level. The diseases of other categories are less dangerous when compared to skin cancer. Hence, an automated system is needed to classify skin disease from multiple categories with better accuracy. The paper proposes a classification model for various skin diseases using a four-approach system. To begin with, the use of transfer learning in pre-trained deep neural networks, namely, inceptionresnetv2, inceptionv3, resnet18, resnet101, vgg19, resnet50, and densenet201, are modified and trained for our system. The networks allowed extracting the features and apply using a multi-class support vector machine in the following approach. The former approach features are hybrid into two ensemble networks and applied using a multi-class support vector machine in the third approach. Finally, a shallow neural network is trained on the ensemble approach’s features for classification in the fourth approach. The ISIC 2018 dataset is used for the experiments and evaluation of the proposed model. The results are analyzed with other states of the art classification methods, and our approach proves to be better with global accuracy at 96.2% and balanced accuracy at 89.2%. The receiver operating characteristic curve (ROC) shows the best performance of the proposed classification model.