Logo Logo
Help
Contact
Switch language to German
Geometric deep learning for Alzheimer's disease analysis
Geometric deep learning for Alzheimer's disease analysis
Alzheimer’s Disease (AD) represents between 50-70% of the cases of dementia, which translates in around 25-35 million people affected by this disease. During its development, patients suffering from AD experience an irreversible cognitive decline, which limits their autonomy on their daily lives. While many of the causes of AD are still unknown, researchers have noticed a abnormal amyloid deposition and neurofibrillary tangles that will start affecting the short-term memory of the patient, together with other cognitive functions. In fact, these pathophysiological changes start taking place even before the patient experiences the first symptoms. One of the structures that is first affected by the disease is the hippocampus. During the development of AD, this part of the brain experiences an irregular deformation that affects its capabilities of forming new memories. Therefore, many clinical work has set a focus on studying this structure and its evolution along the disease. Identifying the changes it suffers can help us understand better the causes of the patient's cognitive decline. Given the complexity that characterizes AD, identifying patterns during its development is still a cumbersome task for physicians.Thus, aiding the diagnosis and prognosis of the disease using Deep Learning methods can be highly beneficial, as seen for other medical applications. In particular, if the focus is set on single structures (e.g. the hippocampus) Geometric Deep Learning offers a set of models that are best suited for 3D shape representations. We believe these methods can help doctors identify abnormalities in the structure that can lead to AD in the future. In this work, we first study the capabilities of current Geometric Deep Learning methods in diagnosing patients suffering from AD, by only looking at the hippocampus. We start by studying one of the simplest 3d representations, point clouds. We continue by comparing this representation to other non-Euclidean representations, such as meshes, and also Euclidean ones (e.g. 3d masks). We observe that meshes are one of the optimal ways of representing 3d structures for capturing fine-grained changes, but they carry additional pre-processing steps that Euclidean representations do not require. Finally, once we have confirmed that Geometric Deep Learning, particularly mesh neural networks, can properly capture the effects of AD on the hippocampus, we expand their application to longitudinal analysis of the structure. We propose a new temporal model based on Spiral Resnet and Transformers that sets a new state-of-the-art for the task of predicting longitudinal trajectories of the hippocampus. We also evaluated the effect that imputing missing longitudinal data has on detecting subjects that are developping to AD. Our experiments show an increase of a 3% in distinguishing between converting and stable trajectories.
Not available
Sarasúa Cañedo-Argüelles, Ignacio
2023
English
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Sarasúa Cañedo-Argüelles, Ignacio (2023): Geometric deep learning for Alzheimer's disease analysis. Dissertation, LMU München: Faculty of Medicine
[thumbnail of Sarasua_Ignacio.pdf]
Preview
PDF
Sarasua_Ignacio.pdf

3MB

Abstract

Alzheimer’s Disease (AD) represents between 50-70% of the cases of dementia, which translates in around 25-35 million people affected by this disease. During its development, patients suffering from AD experience an irreversible cognitive decline, which limits their autonomy on their daily lives. While many of the causes of AD are still unknown, researchers have noticed a abnormal amyloid deposition and neurofibrillary tangles that will start affecting the short-term memory of the patient, together with other cognitive functions. In fact, these pathophysiological changes start taking place even before the patient experiences the first symptoms. One of the structures that is first affected by the disease is the hippocampus. During the development of AD, this part of the brain experiences an irregular deformation that affects its capabilities of forming new memories. Therefore, many clinical work has set a focus on studying this structure and its evolution along the disease. Identifying the changes it suffers can help us understand better the causes of the patient's cognitive decline. Given the complexity that characterizes AD, identifying patterns during its development is still a cumbersome task for physicians.Thus, aiding the diagnosis and prognosis of the disease using Deep Learning methods can be highly beneficial, as seen for other medical applications. In particular, if the focus is set on single structures (e.g. the hippocampus) Geometric Deep Learning offers a set of models that are best suited for 3D shape representations. We believe these methods can help doctors identify abnormalities in the structure that can lead to AD in the future. In this work, we first study the capabilities of current Geometric Deep Learning methods in diagnosing patients suffering from AD, by only looking at the hippocampus. We start by studying one of the simplest 3d representations, point clouds. We continue by comparing this representation to other non-Euclidean representations, such as meshes, and also Euclidean ones (e.g. 3d masks). We observe that meshes are one of the optimal ways of representing 3d structures for capturing fine-grained changes, but they carry additional pre-processing steps that Euclidean representations do not require. Finally, once we have confirmed that Geometric Deep Learning, particularly mesh neural networks, can properly capture the effects of AD on the hippocampus, we expand their application to longitudinal analysis of the structure. We propose a new temporal model based on Spiral Resnet and Transformers that sets a new state-of-the-art for the task of predicting longitudinal trajectories of the hippocampus. We also evaluated the effect that imputing missing longitudinal data has on detecting subjects that are developping to AD. Our experiments show an increase of a 3% in distinguishing between converting and stable trajectories.