Abstract
Brain age prediction from brain MRI scans not only helps improve brain ageing modelling generally, but also provides benchmarks for predictive analysis methods. Brain-age delta, which is the difference between a subject’s predicted age and true age, has become a meaningful biomarker for the health of the brain. Here, we report the details of our brain age prediction models and results in the Predictive Analysis Challenge 2019. The aim of the challenge was to use T1-weighted brain MRIs to predict a subject’s age in multicentre datasets. We apply a lightweight deep convolutional neural network architecture, Simple Fully Convolutional Neural Network (SFCN), and combined several techniques including data augmentation, transfer learning, model ensemble, and bias correction for brain age prediction. The model achieved first places in both of the two objectives in the PAC 2019 brain age prediction challenge: Mean absolute error (MAE) = 2.90 years without bias removal, and MAE = 2.95 years with bias removal.
Introduction
Predicting "brain age" from MRI scans has become a significant area of research in neuroscience. The discrepancy between predicted brain age and chronological age—the brain-age delta—serves as a biomarker for neurodegenerative conditions such as Alzheimer’s disease or schizophrenia. However, developing robust models is challenging due to the scarcity of neuroimaging data and the noise introduced by multi-center collection using different scanners. The Predictive Analysis Challenge (PAC) 2019 aimed to identify robust methods for age prediction across heterogeneous datasets while addressing age-related bias, where models often overestimate the age of younger subjects and underestimate that of older subjects.
Datasets
The study utilized two primary datasets with distinct characteristics:
- PAC 2019: The competition dataset, containing 2,638 training/validation subjects and 660 test subjects from 17 sites, with a wide age range of 17–90 years.
- UK Biobank (UKB): A large cohort used for pretraining, consisting of 5,698 subjects. While larger, the UKB has a narrower age range (44–80 years) compared to the PAC dataset.
Methodology
The core of the approach is the Simple Fully Convolutional Network (SFCN). This lightweight architecture consists of seven convolution blocks designed for rapid downsampling of 3D MRI data, making it more memory-efficient than deeper alternatives.
Key technical strategies included:
- Soft Classification: Instead of direct regression, the model outputs a probability distribution across 38–40 age bins, optimized using Kullback-Leibler (KL) divergence.
- Transfer Learning: Models were pretrained on the UK Biobank and fine-tuned on the PAC 2019 data, improving convergence and final accuracy by approximately 0.22 years MAE.
- Pseudo-Modality Ensembling: An ensemble of 45 models was created using four data versions: linearly registered brains, non-linearly registered brains, grey matter (GM) maps, and white matter (WM) maps.
- Bias Correction: Linear regression was applied to de-bias predictions, ensuring the brain-age delta was not correlated with chronological age.
Key Insights
- Ensemble Performance: Model ensembling significantly outperformed individual models. Power-law fitting suggested that while ensembling is beneficial, performance gains eventually saturate.
- Optimizer Selection: The ADAM optimizer proved more stable and effective for the multi-site PAC 2019 data, contrasting with previous findings where SGD performed better on the more uniform UK Biobank data. This indicates ADAM's suitability for heterogeneous datasets.
Results
The proposed approach, under the team name 'BrainAgeDifference', secured first place in both PAC 2019 objectives:
- Objective 1 (Best MAE): Achieved an MAE of 2.90 years, outperforming the runner-up by 0.18 years.
- Objective 2 (Best MAE with low bias): Achieved an MAE of 2.95 years with minimal age-error correlation ($|r| < 0.1$), winning by a substantial margin over the second-place MAE of 3.799.
Conclusion
While the SFCN-based ensemble achieved state-of-the-art results, limitations remain regarding hyperparameter optimization and the simplicity of the averaging ensemble strategy. Future work should focus on testing generalizability across entirely new scanning sites with different hardware and protocols to ensure the model's robustness in clinical settings.
Summary updated