Classification of similar dermatological diseases from skin lesion images with ensemble learning

Dermatological disorders are one of the most common medical problems today. Some of these diseases have a structure that challenges image-oriented prediction systems by showing similar lesions on the skin. In the present study, it is aimed to predict the disease correctly under these challenging conditions and to increase the prediction success with data augmentation techniques. For these purposes, Urticaria, Actinic Keratosis, Acne Rosacea and Eczema diseases were subjected to classification experiments with 4 different deep learning models. In the experiments conducted with Resnet50, VGG19, EfficientNetV2B3 and InceptionV3 models, the highest classification success was 82.72% with EfficientNetV2B3 model. In the next stage, the models with the highest success were run together with Ensemble Learning and the overall success increased to 87.94%. In addition, it was observed that the classification success increased by an average of 3% with various data augmentation techniques. It was found that the prediction success was higher in images where healthy and diseased skin surfaces were seen together. It was observed that the success decreased when cross-sectional images were taken from diseases such as urticaria that spread over a large surface.