Abstract
Numerous studies have established that estimated brain age, as derived from statistical models trained on healthy populations, constitutes a valuable biomarker that is predictive of cognitive decline and various neurological diseases. In this work, we curate a large-scale heterogeneous dataset (N = 10,158, age range 18 - 97) of structural brain MRIs in a healthy population from multiple publicly-available sources, upon which we train a deep learning model for brain age estimation. The availability of the large-scale dataset enables a more uniform age distribution across adult life-span for effective age estimation with no bias toward certain age groups. We demonstrate that the age estimation accuracy, evaluated with mean absolute error (MAE) and correlation coefficient (r), outperforms previously reported methods in both a hold-out test set reflective of the custom population (MAE = 4.06 years, r = 0.970) and an independent life-span evaluation dataset (MAE = 4.21 years, r = 0.960) on which a previous study has evaluated. We further demonstrate the utility of the estimated age in life-span aging analysis of cognitive functions. Furthermore, we conduct extensive ablation tests and employ feature-attribution techniques to analyze which regions contribute the most predictive value, demonstrating the prominence of the frontal lobe as well as pattern shift across life-span. In summary, we achieve superior age estimation performance confirming the efficacy of deep learning and the added utility of training with data both in larger number and more uniformly distributed than in previous studies. We demonstrate the regional contribution to our brain age predictions through multiple routes and confirm the association of divergence between estimated and chronological brain age with neuropsychological measures.
Introduction
While chronological age is a constant, "brain age"—the biological state of the brain—can vary significantly between individuals. A brain that appears older than its chronological counterpart is a potent biomarker for cognitive decline, Alzheimer’s disease, and other neurological conditions. Historically, brain age estimation models have been limited by small, biased datasets and a lack of interpretability. This study addresses these challenges by utilizing a massive, curated dataset of over 10,000 MRI sessions and applying advanced deep learning techniques to map the neuroanatomical drivers of aging.
Methods
The researchers employed a 3D Convolutional Neural Network (3D-CNN) based on the VGG architecture, specifically adapted for volumetric medical imaging. Scans underwent standard preprocessing, including brain extraction and affine registration to the MNI152 template. To move beyond "black box" predictions, the team used Grad-CAM (Class Activation Mapping) to generate "Age Activation Maps" and conducted ablation studies to determine which brain regions, such as specific lobes, held the most predictive power.
Dataset Curation and Balancing
The study aggregated 10,158 MRI sessions from 6,142 unique healthy subjects across 14 public sources, including ADNI, Cam-CAN, and OASIS. This heterogeneity ensures the model is robust to different scanner types and protocols.
A critical component of the methodology was addressing age-group bias. Raw datasets are often skewed toward specific demographics. By employing a combination of oversampling and undersampling, the researchers created a balanced training set with a uniform age distribution across the adult lifespan.
Key Results
The model achieved state-of-the-art accuracy, with a Mean Absolute Error (MAE) of 4.06 years on a hold-out test set and 4.21 years on the independent Cam-CAN dataset. This performance significantly exceeds previous benchmarks, which reported MAEs around 6.08 years. Furthermore, test-retest experiments demonstrated high reliability, with a standard deviation of only about 1 year across multiple scans of the same subject.
Discussion and Insights
- Frontal Lobe Prominence: Both activation maps and ablation tests identified the frontal lobe as the most critical region for predicting brain age, aligning with clinical observations of its sensitivity to the aging process.
- Uniformity Over Volume: The study found that training on a smaller, age-balanced dataset was more effective than using a larger, unbalanced one. Unbalanced data tends to bias models toward the mean age of the most populous group.
- Cognitive Correlation: The "brain age gap" (estimated minus chronological age) correlated significantly with neuropsychological measures, such as the Benton Face Recognition Test (BFRT), suggesting the model captures meaningful biological variance related to cognitive health.
Summary updated