Source-linked AI summary
Predicting brain age with deep learning from raw imaging data results in a reliable and heritable biomarker
James H Cole, Rudra PK Poudel, Dimosthenis Tsagkrasoulis, Matthan WA Caan, Claire Steves, Tim D Spector, Giovanni Montana
TL;DR
The paper asks whether brain-predicted age can serve as an accurate, genetically valid, and reliable biomarker of brain ageing. It evaluates CNNs on processed and minimally processed T1-weighted MRI, comparing them with GPR across accuracy, heritability, and reliability samples. CNN predictions were accurate, significantly heritable, and highly reliable, including from raw MRI data.
Problem
The study addresses the need to establish brain-predicted age as a valid biomarker that captures individual differences in brain ageing beyond prediction accuracy.
Method
Three-dimensional CNNs predicted age from processed GM and WM maps or minimally processed raw T1-weighted MRI, with GPR as a comparison across large-adult, twin, and reliability samples.
Results
CNNs accurately predicted chronological age from processed or raw MRI, produced significantly heritable brain-predicted ages, and showed high within-scanner and between-scanner reliability.
Takeaways & Limitations
Brain-predicted age has potential as a biomarker for investigating brain ageing and its relationships with cognitive ageing, neurodegeneration, and age-associated brain disease.
Takeaways & Limitations
Heritability estimates were limited in precision by sample size, and between-scanner reliability was assessed using only two scanners with the same field strength.
Abstract
from arXiv · showhide
Machine learning analysis of neuroimaging data can accurately predict chronological age in healthy people and deviations from healthy brain ageing have been associated with cognitive impairment and disease. Here we sought to further establish the credentials of "brain-predicted age" as a biomarker of individual differences in the brain ageing process, using a predictive modelling approach based on deep learning, and specifically convolutional neural networks (CNN), and applied to both pre-processed and raw T1-weighted MRI data. Firstly, we aimed to demonstrate the accuracy of CNN brain-predicted age using a large dataset of healthy adults (N = 2001). Next, we sought to establish the heritability of brain-predicted age using a sample of monozygotic and dizygotic female twins (N = 62). Thirdly, we examined the test-retest and multi-centre reliability of brain-predicted age using two samples (within-scanner N = 20; between-scanner N = 11). CNN brain-predicted ages were generated and compared to a Gaussian Process Regression (GPR) approach, on all datasets. Input data were grey matter (GM) or white matter (WM) volumetric maps generated by Statistical Parametric Mapping (SPM) or raw data. Brain-predicted age represents an accurate, highly reliable and genetically-valid phenotype, that has potential to be used as a biomarker of brain ageing. Moreover, age predictions can be accurately generated on raw T1-MRI data, substantially reducing computation time for novel data, bringing the process closer to giving real-time information on brain health in clinical settings.
1. Introduction
The study evaluates CNN brain-predicted age as a biomarker of individual differences in brain ageing, focusing on accuracy, genetic validity, and reliability across imaging inputs.
- Brain ageing varies substantially between individuals, motivating biomarkers that may help detect early neurodegeneration and predict age-related cognitive decline.
- Machine learning can learn high-dimensional relationships between neuroimaging patterns and chronological age to generate brain-predicted age.
- Brain-predicted age has been used to quantify added brain ageing by subtracting chronological age from the predicted age.
- Image post-processing can take hours or days, conflicting with clinical decisions that may need to occur within minutes.
- Deep learning and CNNs can infer task-specific representations directly from minimally processed or raw neuroimaging data, potentially reducing reliance on time-consuming preprocessing.
- The study tests CNN age prediction using processed and raw structural data, twin heritability, and within- and between-scanner reliability.
2.1. Datasets
The study combines a large healthy-adult cohort for age prediction with female-twin and repeated-scan cohorts for heritability and reliability analyses.
- All datasets consisted of T1-weighted MRI scans that were processed into normalized brain-volume maps and raw-data inputs.
- N = 2001 healthy adults aged 18–90 formed the Brain-Age Normative Control dataset for evaluating age-modelling accuracy.The cohort included 1016 males and 985 females and combined data from 14 publicly available sources.
- N = 62 healthy female twins, including 27 monozygotic pairs and 4 dizygotic pairs, were used to assess heritability.
- N = 20 participants scanned about 28 days apart provided the within-scanner test-retest reliability sample.
- N = 11 participants scanned at Imperial College London and Amsterdam provided the between-scanner reliability sample.
2.2. Neuroimaging processing
Neuroimaging inputs included normalized GM and WM maps as well as minimally processed T1-weighted MRI, with common spatial orientation and dimensions enabling cross-source analysis.
- T1-weighted MRI data were processed to generate normalized brain-volume maps and raw inputs suitable for analysis.
- SPM12 segmented T1 images into GM, WM, and cerebrospinal-fluid classes before generating normalized three-dimensional GM and WM volume maps.
- Raw inputs underwent DICOM-to-NIfTI conversion, six-degree-of-freedom rigid registration to MNI152 space, and resampling to 1mm3 dimensions of 182x218x182 voxels.
- Figure 1 contrasts processed axial GM and WM volumetric maps with raw or minimally processed T1-weighted MRI registered to a common voxel space.
2.3. Machine learning brain age modelling methods
The paper uses three-dimensional CNNs to predict biological age from processed or raw MRI and compares them with Gaussian Process Regression.
- Convolutional neural networks: CNNs were selected to infer predictive imaging features from both processed and unprocessed brain MRI scans.
- Convolutional neural networks: The 3D CNN maps MRI volumes to a single scalar representing biological age through repeated convolution, normalization, activation, pooling, and fully connected operations.
- Convolutional neural networks: A combined GM-and-WM model joined the highest-level blocks of separate tissue networks before a final fully connected age-prediction layer.
- Model training: Training minimized mean absolute error with stochastic-gradient descent and momentum, while data augmentation was used to reduce over-fitting.
- Computational characteristics: CNN testing required 290–940 milliseconds on a single GPU, despite training times of 18, 42, or 83 hours depending on input type.
- Gaussian Process Regression: GPR provided a comparison method using concatenated GM and WM vectors and a linear-kernel similarity matrix.
2.4. Statistical analysis
The analysis compared CNN and GPR age predictions across four structural-MRI input types, then assessed heritability and Brain-PAD reliability using independent datasets.
- Accuracy analysis: Eight accuracy assessments crossed CNN and GPR with GM, WM, GM+WM, or raw T1 data.Each method was evaluated with all four input forms.
- Accuracy analysis: Accuracy was evaluated on BANC test-set predictions using Pearson’s r, R2, MAE, and RMSE.BANC included 1,601 training, 200 validation, and 200 test participants.
- Heritability analysis: Heritability in 62 TwinsUK participants was estimated with SEM models containing additive genetic, common environmental, and unique environmental components.The models used unbiased predictions generated from models trained on the BANC training set.
- Heritability analysis: AE-model heritability was calculated from the additive-genetic and unique-environmental path coefficients as h2 = a2 + e2.The unique-environment component was retained in all models.
- Reliability analysis: Within- and between-scanner reliability used ICC[2,1] to assess absolute agreement between single scanners from Brain-PAD scores.Brain-PAD was calculated by subtracting chronological age at scanning from brain-predicted age.
3. Results
Across accuracy, heritability, and reliability analyses, brain-predicted age performed well with CNN and GPR, including from raw T1-MRI, while reliability varied by scanner setting and input.
- Accuracy: MAE = 4.16 years was achieved with CNN using GM data, while other CNN and GPR predictions were generally comparable.Using single tissues did not appreciably alter accuracy relative to using combined or raw inputs.
- Heritability: Brain-predicted age estimates were significantly heritable across predictive methods and input types, with the highest estimate for CNN using GM+WM data.Raw-input predictions were the exception, showing considerably lower heritability.
- Reliability: Within-scanner reliability was high for all analyses except CNN using WM, with raw-data CNN achieving ICC = 0.94.Within-scanner reliability was generally higher than between-scanner reliability.
- Conclusion: 3D CNNs accurately estimated chronological age from raw or processed T1-weighted MRI with performance comparable to GPR.The findings support brain-predicted age as a potential biomarker of brain ageing.
- Clinical relevance: Raw-input CNN prediction avoids image pre-processing and could make near-real-time clinical use more feasible.The paper identifies removal of pre-processing as a route to substantially accelerating the imaging pipeline.
- Limitations: Heritability decreased after controlling for chronological age, and the limited twin sample prevented exhaustive analysis of age-related changes in heritability.The authors link this age dependency to increasing unique environmental effects, while noting the sample-size limitation.
- Limitations: Between-scanner reliability was evaluated using only two scanners with the same field strength, requiring further scanner varieties for a comprehensive assessment.The authors state that increasing sample numbers would improve estimate precision.
5. Conclusions
Deep learning models using T1-MRI accurately predict chronological age in healthy individuals, including from minimally processed raw data. Brain-predicted age is significantly heritable and highly reliable, supporting its potential as a brain-ageing biomarker.
- Raw T1-MRI data can support accurate chronological-age prediction with minimal processing.
- Brain-predicted age estimates are significantly heritable, providing external genetic validity to the measure.
- Brain-predicted age is highly reliable and appropriate for longitudinal and multi-centre studies.
- Brain-predicted age has potential for investigating brain ageing and its relationships with cognitive ageing, neurodegeneration, and age-associated brain diseases.