Few-shot Learning for Medical Image Classification

Medical Imaging (MI) is essential in clinical diagnosis because it improves the information availability of the body which benefits the diagnostic accuracy and therapeutic procedures. Collecting medical image data is complex and time consuming. As a result, the sizes of medical image datasets are usually much smaller than the sizes of natural image datasets. It hinders the application of deep learning methods in medical image field. Few-shot Learning (FSL) is a type of machine learning problem that has only a limited number of training examples with supervised information. It aims to reduce data gathering effort and computational cost. Therefore, it has potential to solve the lack of data issue in MI field. Common FSL methods include Data Augmentation Meta-learning and Transfer Learning. In this study, we propose a model integrates cross-domain transfer learning and different levels of augmentation to address the Few-shot classification problem for medical image. The experiments are done on the subsets of Optical Coherence Tomography (OCT) 2017 dataset. We pre-train the classifier by ImageNet dataset and use the synthetic images to support the fine-tuning. The synthetic images are generated by the integration of pixel-level augmentation and Cycle-GAN augmentation methods. The result shows that the integrated method outperforms each single method and the combination the augmentation has potential to further enrich the feature representation of the train data. Low generalization ability is another issue for MI. Few-shot MI tasks could make the issue worse. One main reason is that few-shot MI tasks are highly rely on human design. We introduce automatic augmentation and weight-sharing to our previous framework to test if it is able to improve the generalization ability of the framework. The experiments are conducted on four different medical image datasets. Discussions on open issues and future research directions are also included.