Academic

Since its advent in the early 1990s, functional magnetic resonance imaging (fMRI) has revolutionized cognitive neuroscience by providing a non-invasive window into brain function. However, its utility for clinical practice has thus far been limited. Unlike structural MRI, which provides a three-dimensional image that can be visually inspected by an expert radiologist, functional MRI is a four-dimensional modality that requires advanced image analysis algorithms to interpret. A major challenge is the difficulty of extracting accurate subject-level functional brain measures from fMRI data, which is large, noisy, and complex. Computationally convenient analysis approaches, like massive univariate models, have become popular in practice for group-average analysis but are suboptimal for individual-level analysis. Bayesian techniques are well-suited to enhance statistical efficiency and power by leveraging sources of shared information, e.g. across subjects in a population or neighboring locations in the brain. In this talk, I will present pragmatic Bayesian modeling techniques for fMRI data and illustrate their benefits for individual-level fMRI analysis. First, I will present a hierarchical Bayesian independent component analysis (ICA) model employing population-derived priors. Compared to traditional multi-subject hierarchical models, this approach is fast, easily extendable to more complex formulations, and applicable to single-subject analysis. Second, I will present the use of surface-based and volumetric spatial priors, designed to respect neuroanatomy, to leverage spatial dependencies for greater accuracy and power. The proposed models are validated through simulated and real fMRI data.