Investigating brain aging with Magnetic Resonance Imaging and Convolutional Neural Networks

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

Throughout human lifetime there are various ageing-related changes occurring. It has been demonstrated that departures from the healthy ageing trajectory can be used as a biomarker for several neurodegenerative conditions. Previous neuroimaging studies have used deep learning algorithms to investigate structural biomarkers including volumetric, microstructural and focal ones, using both cross-sectional and longitudinal cohorts. In this thesis, I investigate deep learning methods for analysing structural Magnetic Resonance Imaging (MRI) data. The methods developed are used to understand brain ageing throughout human adulthood, by identifying population-wide brain ageing proles. In this thesis such proles are dened by determining periods of the human lifetime which have similar brain age-related features and studying the differences between the features characterising each period. Brain ageing proles covering human adulthood are investigated as brain ageing occurs during adulthood. Training dataset used in this thesis to train brain age prediction algorithms is the largest up-to-date and, therefore, the algorithms can be considered as generalisable as possible. Therefore, brain ageing proles are called "population wide". Population-wide brain ageing proles are derived and described by developing 5 main contributions in this thesis:

1. A dataset comprising of 10,878 MRI T1-weighted scans acquired in healthy subjects (17 96 years of age) was assembled from publicly available data.

2. The dataset was used to develop a convolutional neural network-based ordinal regression model for brain age prediction in order to reflect accumulative nature of brain ageing, using the imaging data with minimal preprocessing (rigid transformation and resampling onto a template).

3. The distribution of predicted ages from both ordinal and metric models were leveraged to propose ageing prole by assuming that the distribution of age predictions represents the distribution of age-related structural features. Ageing profiles were obtained using 4 different methods 2 methods were applied for this purpose (Deep Embedded Clustering (DEC) and Preference Ranking Organization METHod for Enrichment of Evaluations (PROMETHEE) II) and 2 methods were developed for this thesis (the method considering the distribution of predictions of a deep learning model for brain age prediction and ordinal DEC). Ordinal DEC was proposed to introduce the concept of ordinality into the standard DEC.

4. The issues of extracting brain ageing-related features describing brain ageing profiles from MRI data were addressed. In order to understand which features of the original inputs which correspond to ageing, ve existing saliency mapping methods (vanilla backpropagation, guided backpropagation, the Smoothgrad method, Gradient Class Activation Mapping (Grad-CAM), guided Grad-CAM) were compared, and a methodology for applying such methods to ordinal methods was proposed. As all the existing saliency mapping methods used were originally developed to work on natural images, modifications were needed to adapt them to work on 3D input MRI data.

5. Generalisability of the models presented in this thesis was examined. For this purpose their performance was assessed on an independent clinical dataset - the Institute of Psychiatry, Psychology & Neuroscience (IoPPN) dataset. Further, the performance of brain age prediction models was observed on the Healthy Brain Ageing from Public Sources (HBAPS) dataset with reduced resolution.

Accuracy of the ordinal model implemented for the task of brain age prediction was compared to the deep learning model with metric regression. The ordinal and metric models achieved Mean Absolute Error (MAE) of 3.62 and 3.87 years, respectively, on the test data. On the same test data with repeat scans eliminated, i.e. excluding repeat samples acquired from the subjects contained in the training data, ordinal and metric models achieved MAE of 4.10 and 4.37 years. Both metric and ordinal models achieved better results than previously published experiments using input data with the same level of pre-processing. The performance of both models was shown not to be affected by the fact that the dataset contained data collected using different scanners and protocols. The models' performance was, however, affected for the samples with labels for which the number of samples in the training data was small.

The proles obtained using the four methods, and both metric and ordinal models, all agreed within models' accuracy dened by MAE achieved. This suggests that they are all being driven by the same underlying brain features, and that these features may be biologically meaningful.

In order to aid interpretation of the age-related features defining the ageing proles both subject-specific and averaged importance maps were obtained. Further, the importance maps were compared between the ones obtained using ordinal and metric models for brain age prediction. All the results were also interpreted for biological significance. The results agree with existing knowledge on the biology of brain ageing. Among existing saliency mapping methods the methods of guided backpropagation, Grad-CAM and guided Grad-CAM are shown to be usable for studying brain ageing features from MRI data. Saliency maps produced using proposed methodology for saliency mapping for ordinal models also highlighted biologically interpretable features.

The overall aim of this work is not just to develop the methodologies and obtain corresponding results, but also to consider the feasibility of such methods for future use in a clinical setting, in light of the challenges outlined above.
Date of Award1 Sept 2021
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorGareth Barker (Supervisor), Robert Leech (Supervisor) & Giovanni Montana (Supervisor)

Cite this

'