Introducing a novel method to improve the performance of rare medical conditions image classification using similarity-based transfer learning and data augmentation. In academic literacy, there is a wide survey over the problems of using AI to solve medical field tasks. Two of the greatest problems are:
- Lack of data – the process of getting the private patient’s data and having it labeled is slow and expensive.
- Imbalanced data – due to the major difference between pathological and non-pathological cases, the data is extremely imbalanced which causes mis-accuracy in ML and DL algorithms.
One of the most popular ways to handle the lack of data is to perform transfer learning. [1] Nevertheless, this method does not always perform well in the medical field. This work will show how transferring from a base model which was trained over a small yet similar dataset, would achieve a higher accuracy than transferring from a base model which was trained over a robust natural images datasets e.g. ImageNet. In order to solve the class imbalance problem, the thesis will demonstrate how data augmentation can be a beneficial solution in the field of medical images. The method described in this thesis shows a new way to handle new medical conditions never seen before. This method is not coupled to a single deep learning architecture and can seamlessly be used with any other architecture.