A dictionary learning algorithm for multi-subject fMRI analysis based on a hybrid concatenation scheme

Analysis of functional magnetic resonance imaging (fMRI) data from multiple subjects is at the heart of many medical imaging studies, and recently, the approaches based on dictionary learning (DL) are noted as promising solutions to the problem. However, the DL-based methods for fMRI analysis proposed to date do not naturally extend to multi-subject analysis. In this paper, we propose a DL algorithm for multi-subject fMRI data analysis which is derived using a hybrid (temporal and spatial) concatenation scheme. It differs from existing DL methods in both its sparse coding and dictionary update stages. It has the advantage of learning a dictionary common to all subjects as well as a set of subject-specific dictionaries, as a result, it is able to generate both group-level spatial activation maps as well as group-level temporal dynamics, which are particularly attractive for task-based fMRI studies. In addition, by simultaneously learning multiple sub-specific dictionaries, it also provides us with unique sub-specific features as well. Performance of the proposed DL method is illustrated using simulated and real fMRI datasets. The results show that it can successfully extract common as well as sub-specific latent components.