(A) reinforcement learning-guided self-supervised method and its application to Alzheimer’s disease classification

Abstract

Brain MRI dataset consists of 3D structured data. There is a small number of labeled brain MRI datasets due to privacy issues and difficulty in collecting data. Since there is a limited amount of data in each dataset, there is a high probability of overfitting when training the machine learning model. In general, if the dataset is small, transfer learning is used. However, in the 3D medical domain, there is no large-scale labeled dataset such as ImageNet in the 2D image domain. Therefore, Self-supervised Learning, which can train without labels, is mainly used. Self-supervised Learning is a learning method that generates tasks by itself when there is a dataset without labels. Most approaches to generating tasks are based on random. In this paper, we propose a self-supervised learning method that generates tasks by Reinforcement Learning instead of random-based. We show that replacing random-based with Reinforcement Learning enables control of the difficulty of the generated tasks. Thus, more effective Self- supervised Learning is possible and we confirm its utility by applying it to Alzheimer’s disease patient classification tasks.