The broad objective of this research is to develop a powerful deep-learning based multiple testing approach for high-dimensional spatial data that arise commonly in biomedical imaging studies, in particular, brain imaging studies. The motivating problem is to detect the cerebral metabolic abnormalities in Alzheimer?s disease (AD) from Fluorine-18 fluorodeoxyglucose positron emission tomography (FDG-PET) data. Existing multiple testing approaches in solving this problem often ignore or inadequately capture the spatial dependence among the test statistics obtained from brain voxels and thus lose substantial power for the detection. We will develop a novel spatial multiple testing method that utilizes the deep convolutional neural network (DCNN), a key deep- learning technique, to well capture the spatial dependence among test statistics and thus to achieve the optimal power in the sense of minimizing the false nondiscovery rate (FNR) while correctly controlling the false discovery rate (FDR) at a given level. The proposed DCNN-based FDR controlling method has enhanced power to discover new AD-related brain regions that are missed by conventional methods, thereby leading to novel clinical and pathological studies.
The specific aims of this proposal include: 1. To develop an optimal spatial FDR controlling approach by connecting the unsupervised local-significance-index based multiple testing with the supervised DCNN-based image segmentation; 2. To evaluate the proposed spatial FDR controlling approach via extensive simulations under various three-dimensional spatial dependence structures, in comparison with multiple classical and state-of-the-art methods; 3. To apply proposed spatial FDR controlling approach to detect AD-related brain regions using the FDG-PET datasets from the Alzheimer?s Disease Neuroimaging Initiative and the Weill Cornell Brain Health Imaging Institute; 4. To develop a user- friendly and publicly available software package with versions in both Python and R to implement the proposed spatial FDR controlling approach. The proposed DCNN-based approach will also be widely applicable to large- scale multiple testing problems in other fields of biomedical research that involve spatial dependence.
This project will exploit recent advances in deep learning to efficiently solve the large-scale spatial multiple testing problems that arise commonly in biomedical imaging studies. The proposed powerful deep-learning based spatial multiple testing approach will be particularly useful in brain imaging studies on neurodegenerative disorders such as Alzheimer?s disease and age-related cognitive impairment.
