Data Augmentation Techniques for fMRI Data: A Technical Survey

The application of machine learning to fMRI data classification, prediction, and analysis tasks has experienced rapid growth in recent years. However, its implementation has been limited by the relatively small size of labeled fMRI datasets. Data augmentation, a technique that artificially increases the size and diversity of the training dataset, can help mitigate this issue. In this technical survey, we outline the main existing techniques for augmenting fMRI data, primarily represented as image, time series, and graph data. For each of these representations, we explore specific data augmentation methods. We provide a structured framework for understanding and comparing these principal fMRI data augmentation techniques. We define and summarize examples for each category, including image augmentation, time-series augmentation, graph augmentation, and additionally cover simulation and modeling approaches. For each main technique, we offer a general definition, discuss the technical details, and provide schematic illustrations. The final section provides a critical analysis of all methods, offering recommendations on which methods to use in different scenarios, and includes a comparative table of all techniques. This survey aims to serve as a valuable resource for researchers working on fMRI-based machine learning applications, guiding them in selecting appropriate augmentation techniques and inspiring novel approaches to enhance the performance and generalizability of their models.