Skin Disease Classification using Few-Shot Meta-Transfer Learning

Abstract

Number of annotated images belonging to rare diseases are limited due to the requirement of knowledge in limited clinical domains and the limits in the number of affected patients. Deep networks that are utilized in image based diagnosis usually fail to adapt well to new diseases with only a few annotated data. This is especially more apparent in skin lesion and other disease classification datasets where long-tail data distribution scenarios frequently occur. In the scope of this work, we propose a model that adapts itself very well to rare diseases in SD-198 dataset using meta-transfer learning that is not used in this domain before. The proposed Protonet model, which contains a pretrained deep model in its backbone, could succesfully tune itself with a few-shot meta-learning method without overfitting the limited data. In this scope, we experimented with DenseNet121, ResNet50 and VGG19 as our backbone models and observed that the model with ResNet50 performs better than the recent state-of-the-art meta-learning based models with the same dataset.