| Sažetak: | <p>Brain ageing remains an intricate, multifaceted process, marked not just by chronological time but by a myriad of structural, functional, and microstructural changes that often lead to discrepancies between actual age and the age inferred from neuroimaging. Machine learning methods, and especially Convolutional Neural Networks (CNNs), have proven adept in capturing patterns relating to ageing induced changes in the brain. The differences between the predicted and chronological ages, referred to as brain age deltas, have emerged as useful biomarkers for exploring those factors which promote accelerated ageing or resilience, such as pathologies or lifestyle factors. However, previous studies relied overwhelmingly on structural neuroimaging for predictions, overlooking rich details inherent in other MRI modalities, such as potentially informative functional and microstructural changes. This research, utilising the extensive UK Biobank dataset, reveals that 57 different maps spanning structural, susceptibility-weighted, diffusion, and functional MRI modalities can not only predict an individual's chronological age, but also encode unique ageing-related details. Through the use of both 3D CNNs and the novel 3D Shifted Window (SWIN) Transformers, this work uncovered associations between brain age deltas and 191 different non-imaging derived phenotypes (nIDPs), offering a valuable insight into factors influencing brain ageing. Moreover, this work found that ensembling data from multiple maps results in higher prediction accuracies. After a thorough comparison of both linear and non-linear multi-modal ensembling methods, including deep fusion networks, it was found that linear methods, such as ElasticNet, generally outperform their more complex non-linear counterparts. In addition, while ensembling was found to strengthen age prediction accuracies, it was found to weaken nIDP associations in certain circumstances where ensembled maps might have opposing sensitivities to a particular nIDP, thus reinforcing the need for guided selections of the ensemble components. Finally, while both CNNs and SWINs show comparable brain age prediction precision, SWIN networks stand out for their robustness against data corruption, while also proving a degree of inherent explainability. Overall, the results presented herein demonstrate that other 3D maps and modalities, which have not been considered previously for the task of brain age prediction, encode different information about the ageing brain. This research lays the foundation for further explorations into how different factors, such as off-target drug effects, impact brain ageing. It also ushers in possibilities for enhanced clinical trial design, diagnostic approaches, and therapeutic monitoring grounded in refined brain age prediction models.</p>
|
|---|