Please use this identifier to cite or link to this item:
|Title:||Investigating subtypes of alzheimer's disease using graph neural networks on functional MRI scans||Authors:||Ang, Jun Liang||Keywords:||Engineering::Computer science and engineering::Computer applications::Life and medical sciences||Issue Date:||2021||Publisher:||Nanyang Technological University||Source:||Ang, J. L. (2021). Investigating subtypes of alzheimer's disease using graph neural networks on functional MRI scans. Final Year Project (FYP), Nanyang Technological University, Singapore. https://hdl.handle.net/10356/148087||Project:||SCSE20-0230||Abstract:||While advances in Deep Learning have improved the state-of-the-art for classification of neurological disorders such as Alzheimer’s Disease (AD), their effectiveness is limited because they have failed to consider disease heterogeneity, an important characteristic of neurological disorders. Due to the presence of disease heterogeneity, we cannot achieve effective diagnosis and therapy if we consider the entire AD group as homogenous. In this study, a deep learning approach is proposed to identify subtypes of AD from resting-state functional MRI (rs-fMRI) scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI). We first build a Population Scan Network (PSN) from connectivity matrices derived from the rs-fMRI scans and phenotypic information of corresponding subjects. Next, we train a Graph Convolutional Network (GCN) with multitask learning to perform both disease classification and memory score prediction simultaneously on the PSN. More specifically, disease classification refers to the differentiation between cognitively normal (CN) subjects and AD patients. Memory score prediction refers to the prediction of Mini-Mental State Examination (MMSE) scores of subjects, which are also taken from ADNI. GCN was used so we can incorporate not only rs-fMRI data but also phenotypic metadata such as age and gender in our model. Additionally, multitask learning forces the model to learn robust generalizable features useful for both tasks. Subtypes are then obtained by clustering the representations learned by the GCN. Four subtypes of AD were identified. Gene association analysis was performed with Fisher’s Exact Test to find genes significantly different between AD subtypes and CN group but not significantly different between the entire AD cohort and CN group. The model was also decoded to find subtype-specific biomarkers in terms of anatomical locations. By finding subtypes of AD, we may facilitate clinical trials and precision medicine. Our proposed approach is applicable to other diseases that are heterogenous and for other modalities beyond fMRI.||URI:||https://hdl.handle.net/10356/148087||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Student Reports (FYP/IA/PA/PI)|
Updated on May 5, 2021
Updated on May 5, 2021
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.