High-order concept discovery in functional brain imaging

Many spatiotemporal medical image datasets exhibit “high-order” structure, in which many independent variables exist (e.g. space and time) or features are not scalar at all. We analyze these datasets as tensors (high-order generalizations of matrices), preprocessing our dataset using wavelets to improve efficiency and performing latent concept discovery using parallel factor analysis. Both our method and naive tensor approaches discovered concepts representing handedness in an 11 subject motor task fMRI dataset. However, our method compressed the dataset by 98% and completed in 2 hours vs. 8 days, suggesting that a wavelet and tensor approach gains the benefits of high-order analysis while preserving the efficiency of low-order techniques.