There are many active research projects accessing and applying shared ADNI data. Use the search above to find specific research focuses on the active ADNI investigations. This information is requested annually as a requirement for data access.
| Principal Investigator | |
| Principal Investigator's Name: | Mazen Fouad A-wali Mahdi |
| Institution: | Daskell |
| Department: | Research & Development |
| Country: | |
| Proposed Analysis: | Proposed Analysis 1: Magnetic Resonance Imaging (MRI) of Alzheimer's disease shows regional atrophy as the disease progresses. Alongside atrophy, functional and structural connectivity network alterations are witnessed as the brain tries to cope with the disease. Most of the work developed around graphs revolves around high-quality diffusion imaging and functional MRI. Only a handful of studies examined the potential of gray matter networks derived from T1-weighted structural imaging as disease biomarkers for Alzheimer's disease. While recent work has suggested that individual gray matter networks are useful in a variety of neurological diseases, the various options for graph construction and representation have not been thoroughly benchmarked using state-of-the-art methods. Graph construction methods usually fall into two sides. Intensity-based gray matter networks and morphometric similarity gray matter networks. In this analysis, we would like to use the ADNI dataset to compare between the intensity-based and morphometric similarity gray matter networks with the ultimate goal of applying a proof-of-concept diagnostic task between controls and patients using a: 1- Network science approach (graph metrics such as small-worldness, analysis of hubs / rich-club coefficient, the evolution of metrics on longitudinal subjects) 2- An embedding approach where we transform the constructed graphs into vectors or points in space using Graph2vec + UMAP. 3- An end-to-end learning approach with graph convolutional neural networks. Finally, we would like to test how the addition of side information (via late or early fusion in the embedding or end-to-end learning approaches) such as the patient’s sex, age, or more interesting and clinically relevant data such as tau protein levels, amyloid-beta protein levels, or the presence of APOE e4 gene might affect the analysis. Proposed Analysis 2: To complement the proposed analysis 1, we want to investigate the effect of graph data augmentation on the end-to-end learning approach. Data augmentation techniques have been commonly used to improve the performances and generalisability of machine learning models. It is a very common practice when your data is formed of images. Specifically, data augmentation for images includes but is not limited to image translation, rotation, crops, brightness manipulation, and noise injection. However, very few studies have looked into graph data augmentation due to its’ complex nature. In addition, studies that have investigated graph data augmentation were focused on node-classification tasks rather than graph-classification tasks. In our analysis we aim to test various graph data augmentation techniques for graph classification tasks by manipulating the adjacency matrix (extracted from proposed analysis 1) using : 1- Edge density thresholding 2- Non-linear graph filtering methods 3- Removing edges based on graph centrality criterions We will use standard benchmark graph classification datasets such as Proteins or Letter graphs to test various graph neural network methods (Graph Attention Networks, Graph SAGE, GIN Conv). Finally, the same data augmentation techniques will be applied to the gray matter networks (from proposed analysis 1) in the end-to-end learning approach to test whether an improvement in performance is noticed. |
| Additional Investigators | |
| Investigator's Name: | Christian Scuito |
| Proposed Analysis: | Proposed Analysis 1: Magnetic Resonance Imaging (MRI) of Alzheimer's disease shows regional atrophy as the disease progresses. Alongside atrophy, functional and structural connectivity network alterations are witnessed as the brain tries to cope with the disease. Most of the work developed around graphs revolves around high-quality diffusion imaging and functional MRI. Only a handful of studies examined the potential of gray matter networks derived from T1-weighted structural imaging as disease biomarkers for Alzheimer's disease. While recent work has suggested that individual gray matter networks are useful in variety of neurological diseases, the various options for graph construction and representation have not been thoroughly benchmarked using state-of-the-art methods. Graph construction methods usually fall into two sides. Intensity-based gray matter networks and morphometric similarity gray matter networks. In this analysis, we would like to use the ADNI dataset to compare between the intensity-based and morphometric similarity gray matter networks with the ultimate goal of applying a proof-of-concept diagnostic task between controls and patients using a: 1- Network science approach (graph metrics such as small-worldness, analysis of hubs / rich-club coefficient, the evolution of metrics on longitudinal subjects) 2- An embedding approach where we transform the constructed graphs into vectors or points in space using Graph2vec + UMAP. 3- An end-to-end learning approach with graph convolutional neural networks. Finally, we would like to test how the addition of side information (via late or early fusion in the embedding or end-to-end learning approaches) such as the patient’s sex, age, or more interesting and clinically relevant data such as tau protein levels, amyloid-beta protein levels, or the presence of APOE e4 gene might affect the analysis. Proposed Analysis 2: To complement the proposed analysis 1, we want to investigate the effect of graph data augmentation on the end-to-end learning approach. Data augmentation techniques have been commonly used to improve the performances and generalisability of machine learning models. It is a very common practice when your data is formed of images. Specifically, data augmentation for images includes but is not limited to image translation, rotation, crops, brightness manipulation, and noise injection. However very few studies have looked into graph data augmentation due to its’ complex nature. In addition, studies that have investigated graph data augmentation were focused on node-classification tasks rather than graph-classification tasks. In our analysis we aim to test various graph data augmentation techniques for graph classification tasks by manipulating the adjacency matrix (extracted from proposed analysis 1) using : 1- Edge density thresholding 2- Non-linear graph filtering methods 3- Removing edges based on graph centrality criterions We will use standard benchmark graph classification datasets such as Proteins or Letter graphs to test various graph neural network methods (Graph Attention Networks, Graph SAGE, GIN Conv). Finally, the same data augmentation techniques will be applied to the gray matter networks (from proposed analysis 1) in the end-to-end learning approach to test whether an improvement in performance is noticed. |
