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: Running



Alzheimer’s disease (AD) is an incurable, progressive neurological brain disorder. Earlier detection of AD is essential for effective treatment, for which researcher proposed a wide range of statistical and machine learning techniques in the past. A large part of previous work focused on analysing structural magnetic resonance images (MRI) to derive biomarkers for diagnosis. However, AD pathology is complex and not fully captured by MRI, which is why approaches based on non-imaging data often outperform sophisticated deep learning models that work exclusively on MRI. A wide range of non-imaging biomarkers are recorded routinely to measure amyloid and tau deposition, genetic alterations, and cognitive abilities. Different biomarkers are sensitive during different stages of the disease, thus combining imaging and non-imaging information is important to capture the full extent of the disease state. Modern image-based deep learning architectures can have millions of weights to extract high-level features, whereas models on non-imaging biomarkers are often shallow. Therefore, great care must be taken when combining both types of information in one network such that both are used effectively. The focus of this thesis is on the development of heterogeneous neural networks that will utilise non-imaging features to find distinct kernels for the convolutional layers of the network, thus, achieving a tight integration of image and non-image data.