Learning Predictive Cognitive Structure from fMRI Using Supervised Topic Models

We present an experimental study of topic models applied to the analysis of functional magnetic resonance images. This study is motivated by the hypothesis that experimental task contrast images share a common set of mental concepts. We represent the images as documents and the mental concepts as topics, and evaluate the effectiveness of unsupervised topic models for the recovery of the task to mental concept mapping, We also evaluate supervised topic models that explicitly incorporate the experimental task labels. Comparing the quality of the recovered topic assignments to known mental concepts, we find that the supervised models are more effective than unsupervised approaches. The quantitative performance results are supported by a visualization of the recovered topic assignment probabilities. Our results motivate the use of supervised topic models for analyzing cognitive function with fMRI.