TL;DR
This paper introduces a robust framework for estimating brain age from T1-weighted MRI scans by integrating multiple types of regional brain features. Their approach achieves a Mean Absolute Error (MAE) of 3.25 years, outperforming previous methods and offering a more reliable biomarker for assessing neurological health and age-related brain changes.
What Problem Does This Paper Address?
Estimating a person's "brain age" – the biological age of their brain as opposed to their chronological age – is a valuable tool for understanding neurological disorders and the aging process. Magnetic Resonance Imaging (MRI) is a key technology for this, but developing accurate and robust brain age estimation (BAE) models has been challenging. Two main hurdles are: 1) identifying the most informative MRI-derived features that capture subtle age-related changes, and 2) the high cost and complexity of acquiring MRI data. Previous BAE approaches often relied on individual feature sets (like global volumes or specific cortical measurements), which sometimes led to less accurate predictions. This study aims to overcome these limitations by creating a more comprehensive BAE framework that leverages a fusion of diverse regional brain features.
Methods
The core of this research is a novel BAE framework that integrates three distinct types of MRI-derived regional features:
- CAT12 ROI: This involves volumetric measurements of Gray Matter (GM), White Matter (WM), and Cerebrospinal Fluid (CSF) across 142 cortical and sub-cortical regions, derived using the CAT12 Voxel-Based Morphometry pipeline.
- Desikan ROI: This utilizes surface-based cortical measurements from the Desikan-Killiany atlas, including surface area, GM volume, cortical thickness, and curvature metrics for 68 cortical regions.
- Destrieux ROI: Similar to Desikan ROI, this uses cortical measurements but from the Destrieux atlas, covering 74 global regions per hemisphere.
These features are then fed into various regression models, including Linear Regression (LR), Support Vector Regression (SVR), Relevance Vector Regression (RVR), and Generalized Linear Model (GLM). The researchers found that the Generalized Linear Model (GLM) generally performed best, especially when all region-wise features were combined. The model was trained on an 80% subset of the data and validated using 10-fold cross-validation, with the remaining 20% used as an independent test set.
Datasets
The study utilizes the Open Big Healthy Brain (OpenBHB) dataset. This is a newly introduced, multi-site, publicly available benchmark dataset comprising T1-weighted brain MRI scans from 3965 healthy controls. The dataset covers a wide age range (6 to 86 years) and includes region-wise feature metrics derived from the MRI data. The dataset was curated from various existing sources, including ABIDE-1/2, GSP, CoRR, IXI, brainomics/Localizer, MPI-Leipzig, NAR, NPC, and RBP.
Key Insights and Ideas
- Fusion is Key: Integrating multiple types of regional brain features (volumetric, surface area, thickness, curvature) significantly improves brain age estimation accuracy compared to using individual feature sets. This suggests that different regional metrics capture complementary information about brain aging.
- GLM Shines: The Generalized Linear Model (GLM) proved to be a robust choice for brain age estimation, particularly when combined with the comprehensive set of integrated features.
- Gender Differences Exist: While the model performs well overall, there are slight differences in performance between male and female participants, potentially reflecting underlying sex-based differences in brain anatomy and aging trajectories.
- Age-Related Patterns: The study observed that the BAE model tends to overestimate brain age in younger subjects and underestimate it in older subjects, a pattern consistent with previous research.
Key Results
The proposed BAE framework, by integrating the three region-wise feature sets (CAT12 ROI, Desikan ROI, and Destrieux ROI) and using a GLM, achieved a Mean Absolute Error (MAE) of 3.25 years. This represents a significant improvement over previously reported MAEs for similar T1-weighted MRI-based BAE models, which ranged from 3.7 to 4.98 years in comparable studies. The study also reported an R² score of 0.90 for the combined features. When tested on gender-specific hold-out sets, the model achieved an MAE of 3.19 years for males and 3.32 years for females.
Limitations and Future Work
The authors acknowledge a few limitations:
- Age Distribution Skew: While the dataset is extensive, the age distribution is right-skewed, with fewer participants in older age groups. This could impact the model's performance for very elderly individuals.
- Limited Atlas Coverage: The study primarily used volumetric features from the Neuromorphometrics atlas and did not explore other voxel-based atlases like Suit and Cobra.
- Unimodal Data: The current approach relies solely on T1-weighted MRI data and does not incorporate multimodal data such as functional MRI (fMRI), which has been shown to be beneficial in other studies.
Future research directions suggested include:
- Expanding the training dataset with more adult and elder healthy individuals to improve performance across the entire age spectrum.
- Investigating the use of other voxel-based atlases.
- Exploring multimodal data integration (e.g., fMRI) to further enhance BAE accuracy.
Summary updated