If you decided to utilize deep learning in any image processing application, you would be faced with the issue, "Which architecture should I use?" due to the proliferation of existing CNN models and their advancements. Unfortunately, your answer will only be partially correct because each alternative has its advantage. The underlying idea of this research is to combine recent CNN models instead of selecting just one for optimal accuracy. Our study applied this idea to color lesion images to diagnose skin diseases. By ensembling, the recent CNNs, over 99% classification accuracy and over 97% sensitivity were achieved for the ISIC-2017 dataset, which contains 2000 lesion images. Our mean sensitivity and AUC values for classifying 10000 color lesion images into seven different skin diseases (ISIC-2018) were 0.825% and 0.922%, respectively.In categorizing over 25000 images from the ISIC 2019 dataset, our suggested technique achieved a mean sensitivity of over 90%.

In this study, the performance of popular convolution architectures against an imbalanced dataset is analyzed in detail with a multi-classing medical image processing application. Our selection for dermoscopic images is a large-scale and imbalanced dataset consisting of 10,015 colored lesion images belonging to 7 different skin diseases, was used as a benchmark. Images without pathological testing are labeled by specialist dermatologists who are members of International Skin Imaging Association. The f1-score was preferred as the measurement metric during the training phase of the convolution networks, which were trained with imbalanced dataset, and the area under the receiver operating characteristic curve and the confusion matrix of each model were calculated at the test phase. In the validation phase of convolution networks, k-fold cross validation technique was used. In addition, the filters obtained from the ImageNet dataset have been imported with the Transfer-Learning option. Fine-tuning was applied to the deepest convolution layers in order for these pre-trained models to develop themselves specific to our application. In order to prevent the overfit problem, the feature extraction outputs of the models were drop-out at a rate of 50% after flattening, and L2-regularization (weigh decay) was applied during the update phase of the weights. Although it is not the main purpose of the study, in order to partially improve the performance of convolution architectures, synthetic lesion images created with data-augmentation for the minor classes in the imbalanced dataset were included in the training process in a way that does not cause information leakage.