Predicting brain age and health from MRI scans using AI

Brain age estimation models, or "brain clocks," use MRI scans to estimate the biological age of a brain. Unlike chronological age, brain age provides insights into an individual's brain health, with a higher predicted brain age indicating accelerated ageing. This can serve as a powerful biomarker for detecting conditions like dementia, Alzheimer's, Parkinson's, and multiple sclerosis, as well as identifying individuals at risk of developing neurodegenerative conditions before clinical symptoms appear.

Our research team has recently performed an in-depth analysis of the current trends in artificial intelligence (AI) applied to brain age estimation. Leveraging advanced neuroimaging data from the UK Biobank (UKBB), a comprehensive dataset containing high-quality brain imaging data [ukb_web], and utilising state-of-the-art machine learning (ML) and deep learning (DL) techniques, we have developed robust brain clocks that show significant potential in predicting brain age and detecting early signs of neurodegenerative diseases. Read our latest publication in NeuroImage here!

Our Approach: Cutting-Edge Techniques and Data Integration 

T1-weighted MRI scans from the UKBB were processed via FastSurfer [Henschel, 2020], a state-of-the-art neuroimaging tool, to derive standardised images and to extract detailed image-derived phenotypes for use in our models. Our analysis compared a broad spectrum of ML models, including traditional approaches like LASSO regression, and innovative DL architectures, such as the Simple Fully Convolutional Network.

One of the main challenges in brain age prediction is the systematic bias observed across different age groups. Younger individuals tend to have their brain age overestimated, while the age of older individuals is often underestimated. To address this, we employed advanced correction methods, including Cole’s [Cole, 2020], Lange’s [de Lange, 2020] and Zhang’s age-bias correction [Zhang, 2023], which in general led to an improvement in model stability and performance across all age brackets.

While all considered models were trained exclusively on UKBB data, their generalisability was tested by considering two external datasets (the Alzheimer’s Disease Neuroimaging Initiative (ADNI) [adni_web] and the National Alzheimer’s Coordinating Center (NACC) [nacc_web]).

Key Findings: High Accuracy and Generalisability

• Our models achieved a Mean Absolute Error (MAE) of less than 1 year, showcasing high accuracy in brain age prediction which was maintained across all 5-year-long age bins for individuals between 55 to 85 years of age on all of the databases analysed.

• Additionally, these models show strong potential as biomarkers for neurodegenerative conditions, such as dementia and multiple sclerosis, where brain age prediction achieved an AUROC of up to 0.90 in distinguishing healthy individuals from those with dementia or multiple sclerosis.

• The best-performing models for both age and disease prediction were K-fold cross-validated penalised linear models, typically LASSO, enhanced with Zhang’s age-bias correction. These ML-based models generalised well to the external datasets considered, highlighting the robustness of the approach.  For DL models, generalisation to external datasets was more challenging, underscoring the need for additional fine-tuning to enhance their robustness. Notably, certain architectures, particularly ResNet18, showed significant improvements in age prediction when corrected using the Lange method. However, despite these gains in age prediction, the performance in disease classification remained limited.

• Zhang’s correction emerged as the most effective method for age prediction. However, it may introduce numerical issues, especially in non-linear models prone to overfitting. For disease prediction, Cole’s correction and uncorrected models tend to perform better overall, although there are cases where Zhang- and Lange-corrected models also show strong results. This highlights the need for improved correction techniques that can effectively reduce age bias while preserving the strong disease prediction capabilities of uncorrected models.

Clinical Implications and Future Directions 

The ability of our brain age models to detect deviations from healthy ageing patterns suggests their potential as non-invasive biomarkers for early-stage diagnosis of neurodegenerative conditions. By integrating multimodal data sources and exploring new trends in transfer learning, we aim to further enhance the accuracy and clinical utility of these models. Our future work will focus on incorporating additional MRI modalities to build a more comprehensive predictive framework.

At Oxcitas, we are excited to be at the forefront of this innovative intersection of AI and neuroimaging. Our research highlights the promise of AI-driven brain age estimation as a transformative tool in the early detection and management of neurological diseases.

REFERENCES

[ukb_web] UK Biobank. https://www.ukbiobank.ac.uk/ Last accessed on 18/11/2024.

[Henschel, 2020] Henschel L et al (2020). FastSurfer - A fast and accurate deep learning based neuroimaging pipeline. NeuroImage 219: 117012.

[Cole, 2020] Cole JH (2020). Multimodality neuroimaging brain-age in UK biobank: relationship to biomedical, lifestyle, and cognitive factors. Neurobiology of aging 92: 34-42.

[de Lange, 2020] de Lange A-MG and Cole JH (2020). Commentary: Correction procedures in brain-age prediction. NeuroImage: Clinical 26.

[Zhang, 2023] Zhang B et al (2023). Age-level bias correction in brain age prediction. NeuroImage: Clinical 37: 103319.

[adni_web] Alzheimer’s Disease Neuroimaging Initiative. https://adni.loni.usc.edu/ Last accessed on 18/11/2024.

[nacc_web] National Alzheimer’s Coordinating Center. https://naccdata.org/ Last accessed on 18/11/2024.