fMRI Brain Image Retrieval Based on ICA Components

This manuscript proposes a retrieval system for fMRI brain images. Our goal is to find a similarity-metric to enable us to support queries for "similar tasks" for retrieval on a large collection of brain experiments. The system uses a novel similarity measure between the result of probabilistic independent component analysis (PICA) of brain images. Specifically, the times series of an fMRI dataset will be represented using a number of ICA components as high level task- related features. The similarity between two datasets is the value of the maximum weight bipartite matching defined on the component-wise similarities. The component-wise similarities are calculated based on the size of the overlap between the "highly activated" regions in the corresponding activation maps. We evaluated the performance of the proposed method on a moderate size fMRI image database with considerable variety. The ICA-based component selection in combination with bipartite matching similarity measure outperforms several other component selection methods and similarity measurements. The results also suggest that there is a direct correlation between the involvement of ICA components in cognitive processes and their time course spectrum. Along with other heuristics, this property can be for fMRI image retrieval and classification.