Abstract:
|
Identifying time-varying brain regions exhibiting distinct measurements between patients and the healthy controls is of scientific importance for early disease diagnosis. For this purpose, we consider a novel dynamic tensor response regression model, where the longitudinal image tensor is treated as a response, while the disease status and other covariates are treated as predictors. Our model exploits the intrinsic tensor structures and temporal smoothness effect to complete the missing MRI imaging data and identify significant brain regions unique to AD patients. In theory, we establish the non-asymptotic error bound of our iterative estimator, which reveals an interesting trade-off between statistical and computational errors and provides the theoretical limit of the missing percentage. The application to the ADNI data further demonstrates the efficacy of our method.
|