Transfer Learning for Automated Diagnosis of Skin Conditions using Existing Convolutional Neural Network (CNN) Architectures

Abstract

Biomedical image analysis has been exploited considerably by recent technology involvements, carrying about a pattern shift towards‘automation’ and ‘error free diagnosis’ classification methods with markedly improved accuratediagnosisproductivityandcosteffectiveness.Thispaperproposesan automated deep learning model to diagnose skin disease at an early stage by usingDermoscopyimages.Theproposedmodelhasfourconvolutionallayers, twomaxpoollayers,onefullyconnectedlayerandthreedenselayers.Allthe

convolutional layers are using the kernel size of 3 ∗3 whereas the maxpool layer is using the kernel size of 2 ∗ 2. The dermoscopy images are taken fromtheHAM10000dataset.Theproposedmodeliscomparedwiththethree

different models of ResNet that are ResNet18, ResNet50 and ResNet101. The models are simulated with 32 batch size and Adadelta optimizer. The proposed model has obtained the best accuracy value of 0.96 whereas the ResNet101 model has obtained 0.90, the ResNet50 has obtained 0.89 andtheResNet18modelhasobtainedvalueas0.86.Therefore,featuresobtained from the proposed model are more capable for improving the classification performance of multiple skin disease classes. This model can be used for early diagnosis of skin disease and can also act as a second opinion tool for dermatologists.