Abstract
Addressing the challenges of rare diseases is difficult, especially with thelimited number of reference images and a small patient population. This is moreevident in rare skin diseases, where we encounter long-tailed datadistributions that make it difficult to develop unbiased and broadly effectivemodels. The diverse ways in which image datasets are gathered and theirdistinct purposes also add to these challenges. Our study conducts a detailedexamination of the benefits and drawbacks of episodic and conventional trainingmethodologies, adopting a few-shot learning approach alongside transferlearning. We evaluated our models using the ISIC2018, Derm7pt, and SD-198datasets. With minimal labeled examples, our models showed substantialinformation gains and better performance compared to previously trained models.Our research emphasizes the improved ability to represent features inDenseNet121 and MobileNetV2 models, achieved by using pre-trained models onImageNet to increase similarities within classes. Moreover, our experiments,ranging from 2-way to 5-way classifications with up to 10 examples, showed agrowing success rate for traditional transfer learning methods as the number ofexamples increased. The addition of data augmentation techniques significantlyimproved our transfer learning based model performance, leading to higherperformances than existing methods, especially in the SD-198 and ISIC2018datasets. All source code related to this work will be made publicly availablesoon at the provided URL.