Approaches towards advanced brain age prediction models

Wang, Hanzhi ORCID: https://orcid.org/0000-0002-7714-4606 2024. Approaches towards advanced brain age prediction models. PhD Thesis, Cardiff University. Item availability restricted.

Preview	PDF - Accepted Post-Print Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (7MB) | Preview
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (170kB) | Request a copy

Abstract

As the global population ages, it becomes crucial for the early detection and prevention of neurological aspects of ageing, such as cognitive decline. The human brain ageing process is biologically complex and could be affected by various factors. Therefore, the determination of a person’s brain biological ageing process holds important clinical implications, which reflect that person’s brain health and may indicate the risk of age-associated brain diseases. To quantitatively measure the brain biological ageing process, brain age, defined as the biological age of the brain, has been proposed, which has demonstrated huge potential in clinical diagnosis. Brain age can be estimated by brain age prediction models, which take in brain-ageing related information, such as brain MR scans, and adopt machine learning models to make predictions. Despite the rapid research advancements in brain age prediction over the past decade, brain age prediction framework has yet to mature before the implementation in clinical practice. In this thesis, we propose three novel approaches, each focused on a distinct perspective, to make the brain age prediction model a more reliable, practical, and accurate tool for clinical diagnosis. Our contribution is three-fold: firstly, we propose a skewed loss function to correct a commonly-observed regression bias in brain age prediction models. The skewed loss function unifies the model training and bias correction stages, achieving improved accuracy compared with the state-of-the-art practices in the literature. A dynamic training algorithm is further proposed for the skewed loss function. It adopts a heuristic approach to iteratively tune the hyperparameters of the skewed loss function, which has been proven to be robust to different datasets, model architectures and problem domains. The proposed skewed loss function makes the model produce unbiased estimations of brain age, which makes brain age prediction a reliable and trustworthy tool for clinical use. Secondly, we generalise brain age prediction for clinical-grade low-resolution MR images, which makes it a practical and accessible tool for hospital settings. We propose an integrated workflow by combining brain super-resolution and age prediction models. Clinical MR images are firstly super-resolved, before being fed into a pre-trained age prediction model, which have been proven to achieve negligible differences in predicting age compared with high-resolution images. A non-uniform sampling strategy is also demonstrated to improve the image reconstruction quality especially in the high-frequency regions of the brain. Lastly, we demonstrate the strength of adopting a multi-modal approach for predicting brain age more accurately compared with uni-modal models. Structural T1-weighted MR images and diffusion MRI measures of the brain microstructures are adopted to provide the model with a more complete picture of brain ageing. A tract-wise training approach is also proposed for predicting brain age from diffusion MRI measures, which performs feature selections in the model training process. It prioritises more age-sensitive features and discards less useful ones for brain age prediction, which achieves an improved accuracy on a relatively small dataset.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Computer Science & Informatics
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software
Funders:	China Scholarship Council
Date of First Compliant Deposit:	10 September 2024
Date of Acceptance:	6 September 2024
Last Modified:	13 Sep 2024 09:28
URI:	https://orca.cardiff.ac.uk/id/eprint/171973

Actions (repository staff only)

Edit Item