CAREER: Efficient Statistical Inference using Neuroimaging data for Sample Enrichment and Optimizing Power

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

Hypothesis testing on neuroimaging data traditionally has made use of classical statistical tests (on uni-variate response variables). This makes sub-optimal use of the structure of images, particularly problematic if the two groups being tested have weak differences to begin with. Failure to detect statistically significant differences may imply failure of the experiment itself. Acquiring more images is expensive but also occasionally infeasible. This project develops technologies to address these problems (particularly those dealing with differential analysis of brain images) via the lens of computer vision and machine learning. The algorithmic component of this project is (1) a suite of convex optimization based multi-modal learning schemes to seamlessly leverage a spectrum of brain imaging data, (2) new multi-resolution representations for inference with surface/network based signals (data derived from structural/functional brain images), and (3) using these mechanisms for boosting statistical power even in experiments with small sample sizes. The project has broad scientific impact. Extending the operating range of statistical image analysis methods for neuroimaging will foster a new inter-disciplinary area at the interface of computer vision, biostatistics, and machine learning, which is highly intellectually stimulating. The research team brings real neuroimaging research data for undergraduate/graduate students to explore and study. The project goals also include training and mentoring of students, increased involvement of under-represented groups, seminars, and an extensive set of outreach activities. In addition, the resultant software tools drive the analysis of neuroscience studies, which has clear broad societal impact.