- This event has passed.
Neuroimaging Initiative webinar: Bayesian scalar-on-image neural networks with application to neuroimaging data
October 2 @ 9:00 am - 10:00 am
About Dr. Kang: Jian Kang is a Professor in the Department of Biostatistics and is a faculty member of the Kidney Epidemiology and Cost Center (KECC) at the University of Michigan. He received his PhD in Biostatistics from the University of Michigan in 2011. He was an Assistant Professor in the Department of Biostatistics and Bioinformatics and the Department of Radiology and Imaging Sciences at Emory University from 2011 – 2015. He was a core faculty member in the Center for Biomedical Imaging Statistics (CBIS) at Emory University. His primary research interests are in developing statistical methods for large-scale complex biomedical data with application in precision medicine, imaging, epidemiology and genetics.
BAYESIAN SCALAR-ON-IMAGE NEURAL NETWORKS WITH APPLICATION TO NEUROIMAGING DATA
Deep neural networks have been adopted in the scalar-on-image regression which predicts the outcome variable using image predictors. However, training DNN often requires a large sample size to achieve a good prediction accuracy and the model fitting results can be difficult to interpret. In this work, we construct a novel single-layer Bayesian neural network (BNN) with spatially-varying coefficients (SVC) for the scalar-on-image regression. Our goal is to select interpretable image features and to achieve the high prediction accuracy with limited training samples. We assign the soft-thresholded Gaussian process (STGP) prior to the SVCs and develop an efficient posterior computation algorithm based on stochastic gradient Langevin Dynamics (SGLD). The BNN-STGP provides a large prior support for sparse, piecewise-smooth and continuous SVCs, enabling efficient posterior inference on image feature selection and automatically determining the network structures. We establish the posterior consistency of estimating the SVCs in the model and image feature selection consistency when the number of voxels/pixels grows much faster than the sample size. We compared our methods with state-of-the-art deep learning methods via extensive simulations and analyses of multiple real datasets including the task fMRI data from the ABCD study.
Connect to the webinar via this link. Meeting ID: 923 3875 2870. Meeting Passcode: 149254