Machine and deep learning for finding multimodal imaging biomarkers in prodromal AD

我们提出的研究重点是发展深度神经网络和复杂的多变量分析

Our proposed study focuses on developing deep neural networks and sophisticated multivariate analysis methods for studying episodic memory activations in prodromal AD subjects and age-matched normal controls. We are particularly interested in investigating the effects of spatial and object pattern-separation in subfields of the hippocampus, nearby regions of the medial temporal lobe, and functional whole-brain connections. In order to acquire a fuller understanding of the underlying physiological processes driving AD pathology, activations in hippocampal subfields must be investigated in further depth and with methodologies that exceed the current limitations of fMRI at 3T. Acquiring data that will yield a more accurate view of these hippocampal interactions is far more easily facilitated with 7T technology, although barriers complicate such a study even at 7T. Our proposed study centers on circumventing these barriers, particularly data contamination from excessive system noise, head-motion noise, and physiological noise which is proportional to the field strength, to develop methods that will allow investigators to work within the parameters of 7T at its fullest capacity toward the development of more powerful imaging biomarkers for diagnosing AD. To increase the likelihood of the successful completion of our study, we consider it imperative to develop better task fMRI designs and imaging protocols (Aim 1), automatic segmentation methods (Aim 2), noise-reduction methods (Aim 3), and multivariate analysis methods, such as novel algorithms and software tools based on constrained canonical correlation analysis (constrained CCA), kernel CCA, and deep CCA and relevant group-level analysis using fusion CCA, multiset CCA, and machine learning and deep learning techniques (Aim 4) for studying memory function to obtain novel imaging biomarkers (Aim 5) to identify individuals at risk for AD. This study will enable the creation of clearer and more detailed brain activation maps and thus promote the discovery of currently unknown aspects of brain function in prodromal AD. The successful completion of our objectives could lead to more effective diagnostic tools for AD, including an fMRI-based diagnostic test for memory impairment to characterize abnormal memory function in people at risk for AD. Our advanced methodology, combining 7T high-resolution fMRI, automatic segmentation, data denoising and multivariate analysis, will be essential for detecting subtle functional changes in subfields of the hippocampus and its connections to other cortical regions. Results from this study are expected to broadly impact scientific understanding of brain function beyond only enhancing current understanding of memory function in AD. We anticipate that the methods developed from findings acquired in our proposed study will have a far-reaching influence on improving fMRI data quality, enable more accurate detection of brain activation, open a path toward better automated and instantaneous hippocampal subfield segmentation for many other MRI/fMRI applications of neurodegenerative diseases, and contribute new and vital discoveries to the field of neuroscience in general.