Multi-CNN models with Pretraining for Binary Classification in Skin Cancer

In that a myriad of people died because of skin cancer contemporarily and their diversity makes the distinguishing malignant or benign skin cancer even intricate, classifying them with small dataset training would be more challenging. In this paper, obtaining a higher accuracy of the binary classification problem for skin cancer by applying employed models with or without pretraining was tried. 3 proposed models were proposed and mainly popular models such as Inception3, Xception, VGG16, MobileNet, ResNet etc. Moreover, some popular models pretrained by utilizing the ImageNet dataset are worked on this study to test whether pretrained models can do well on this task or not. To achieve this stated goal, popular models are connected with the Dropout layer, the Global Average Pooling layers and the output layer sequentially to accommodate this requirement of the output. By controlling variables, multiple models were compared with different values of hyperparameters, such as number of epochs, batch size and image size. Through experiments, it is conspicuous that simpler models could converges on this small dataset rapidly than more complex models. Plus, pretraining cannot change the convergence of models and it would improve the performance of models which are already converged on this dataset.