Biomedical imaging technology has undergone rapid advancements over the last several decades, producing large volumes of multimodal imaging data that hold great promise as biomarkers for aging-related diseases. Current imaging biomarkers are primarily based on specific extracted 1D measures that may not fully capture the richness of imaging data. Utilizing 3D or higher imaging information directly may facilitate the identification of more effective disease biomarkers to inform diagnosis, prognosis, and treatment. However, this also brings significant challenges, such as analyzing irregularly shaped 3D objects, managing high-dimensional and high-resolution data, addressing noisiness and complexity, quantifying uncertainty, and ensuring the interpretability of the results. We propose developing efficient statistical learning approaches and scalable computing tools to extract and assess biomarkers from large-scale brain imaging studies. We will also incorporate genetic and clinical information in constructing the biomarkers. Specifically, our proposal comprises five interrelated research aims carried out by investigators with complementary expertise from three institutions. Aim 1 focuses on developing an interpretable model for genome-wide association studies with brain imaging phenotypes and non-visual contextual information. Aim 2 targets to develop novel distributed learning methods for analyzing 3D brain imaging data using an innovative domain decomposition strategy to improve computing performance. Aim 3 quantifies the bias effect in image processing and develops inference methods to reveal the underlying signal from brain imaging data and identify significant brain regions among different diagnosis groups. Aims 4-5 aim to develop statistical methods for obtaining and evaluating imaging-adjusted biomarkers for disease diagnosis and prognosis and assess the incremental value of imaging information over genetic biomarkers on diagnosis and prediction accuracy.
