Deep learning for interpretable brain age estimation
File(s)
Author(s)
Bintsi, Kyriaki Margarita
Type
Thesis
Abstract
Brain ageing, and more specifically the difference between the chronological and the biological age of a person, may be a promising biomarker for identifying neurodegenerative diseases, such as Alzheimer’s Disease. In healthcare, predictive tools must not only be accurate but also easily understood and trusted by doctors. In this thesis, we leverage deep learning for the estimation of brain age from MR images, while also enhancing the interpretability of the employed architectures.
Firstly, we explore how Convolutional Neural Networks (CNNs) perform on the task of brain age regression. We leverage 3D patches of the brain to develop a localised brain age estimator.
Subsequently, we delve deeper into interpretability, constructing a pipeline that generates importance maps for brain age regression using noise masks. These maps discover the parts of the brain that are the most important for brain age.
A way to seamlessly integrate multi-modal information and capture the relationships among individuals is using population graphs along with Graph Neural Networks (GNNs).
Here, we highlight the importance of a meaningful graph construction, which is a non-trivial task, and experiment with different static population-graph construction methods and their effect on GNN performance for brain age estimation. The findings indicate that static graph construction approaches are potentially insufficient for the task of brain age estimation.
Thus, we explore adaptive graph learning and we propose framework that learns a population graph structure optimised for the downstream task. An attention mechanism assigns weights to a set of imaging and non-imaging features (phenotypes), which are then used for edge extraction. Additionally, by visualising the attention weights that were the most important for the graph construction, we increase the interpretability of the graph.
Overall, this dissertation focuses on the application of deep learning techniques, particularly CNNs and GNNs, for brain age estimation, while concurrently increasing interpretability.
Firstly, we explore how Convolutional Neural Networks (CNNs) perform on the task of brain age regression. We leverage 3D patches of the brain to develop a localised brain age estimator.
Subsequently, we delve deeper into interpretability, constructing a pipeline that generates importance maps for brain age regression using noise masks. These maps discover the parts of the brain that are the most important for brain age.
A way to seamlessly integrate multi-modal information and capture the relationships among individuals is using population graphs along with Graph Neural Networks (GNNs).
Here, we highlight the importance of a meaningful graph construction, which is a non-trivial task, and experiment with different static population-graph construction methods and their effect on GNN performance for brain age estimation. The findings indicate that static graph construction approaches are potentially insufficient for the task of brain age estimation.
Thus, we explore adaptive graph learning and we propose framework that learns a population graph structure optimised for the downstream task. An attention mechanism assigns weights to a set of imaging and non-imaging features (phenotypes), which are then used for edge extraction. Additionally, by visualising the attention weights that were the most important for the graph construction, we increase the interpretability of the graph.
Overall, this dissertation focuses on the application of deep learning techniques, particularly CNNs and GNNs, for brain age estimation, while concurrently increasing interpretability.
Version
Open Access
Date Issued
2024-04
Date Awarded
2024-09
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Rueckert, Daniel
Hammers, Alexander
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/L015226/1
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)