Deep Heterogeneous Neural Network for Alzheimer's Disease Classification

Deep Heterogeneous Neural Network for Alzheimer’s Disease Classification

Title: Deep Heterogeneous Neural Network for Alzheimer’s Disease Classification


Type: Master Thesis

Student: Tom Nuno Wolf

Supervisor: Sebastian Pölsterl, Christian Wachinger

Status: Finished, January 12, 2021



Alzheimer’s Disease (AD) is a neurodegenerative disease, which is characterized by symptoms like dementia and an inability to form new memories. Currently, clinical staging of the disease is based on symptoms rather than its underlying biochemical processes, which take place decades before symptoms arise and are not fully understood yet. Recent trends in medical imaging allow for in-vivo assessment of the disease, with flagship studies like Alzheimer’s Disease Neuroimaging Initiative (ADNI) and Australian Imaging, Biomarker and Lifestyle Flagship Study of Ageing (AIBL) gathering huge amounts of data. This enables researchers to train machine learning models to classify AD, with Convolutional Neural Networks (CNNs) a popular choice. Ultimately, those models could help to gain new insights into the interaction of biomarkers related to the disease. As relevant biomarkers are retrieved from different sources like imaging and cerebrospinal fluid (CSF) analysis, models need to be able to handle heterogeneous data. Although other fields of study working with heterogeneous data have proposed sophisticated architectures for deep learning models, they have not been applied to AD classification. Based on the literature review, the standard concept is concatenation of high-level features. Therefore, we evaluate new CNN- based architectures, that are able to incorporate heterogeneous data obtained by ADNI. Models are trained on the hippocampal formation of T1-weighted MRI, aggregates of Positron Emission Tomography (PET) imaging, CSF analysis, genetic analysis and demographic data. We carry out a benchmark of different architectures and show that merging of heterogeneous features on a channel level, based on FiLM [40], improves the current state-of-the-art and we carry out an exhaustive ablation study on FiLM. Lastly, we propose a new model that takes into account the distribution of an image feature map when merging heterogeneous features and outperforms all models of the literature.