Source-linked AI summary
Brain-PACE: A Deep Siamese MRI Framework for Modelling Longitudinal Brain Acceleration
Samuel Maddox, Jacob Newman, Saber Sami, Michal Mackiewicz, for the Alzheimer's Disease Neuroimaging Initiative, the Australian Imaging Biomarkers, Lifestyle flagship study of ageing
TL;DR
Cross-sectional brain-age measures provide limited information about longitudinal ageing trajectories, motivating a direct measure of person-specific structural change. Brain-PACE uses paired T1w MRI in a Siamese probabilistic framework and identified accelerated ageing in 42.6% of MCI participants, with faster pace associated with clinical impairment and regional tau burden.
Problem
Most machine-learning measures of biological brain age are cross-sectional and do not directly quantify person-specific rates of ongoing structural change.
Method
Brain-PACE uses paired T1w MRI in a Siamese framework combining difference embeddings, spatial attention, soft label distributions, Cramér distance, and uncertainty estimation.
Results
42.6% of MCI participants were classified as showing accelerated Brain-PACE, which was associated with greater clinical impairment and regional tau burden.
Takeaways & Limitations
Brain-PACE is supported as a complementary longitudinal imaging phenotype sensitive to clinical and biological changes in early neurodegeneration.
Takeaways & Limitations
The cohorts primarily represented older adults, and biomarker analyses were restricted to the ADNI MCI group, limiting generalization across the lifespan and settings.
Abstract
from arXiv · showhide
Brain age estimation has become a popular research proxy for assessing brain health and disease, yet longitudinal trajectories of brain ageing are still poorly defined, and clinical use is limited. Building on existing Siamese longitudinal frameworks, we develop Brain-Predicted Age Acceleration (Brain-PACE) to directly estimate the pace of structural brain ageing from paired T1-weighted MRI. Brain-PACE identified accelerated ageing in $42.6$% of participants with mild cognitive impairment. Faster Brain-PACE was associated with greater functional and cognitive impairment (FAQ; $r=0.35$, ADAS13; $r=0.30$, CDR-SB; $r=0.32$) and greater regional tau burden in the posterior cingulate ($r=0.59$), precuneus ($r=0.47$), and entorhinal cortex ($r=0.37$). These associations were stronger than those observed when pace was calculated indirectly from repeated cross-sectional brain age estimates, suggesting that direct longitudinal modelling captures complementary information relevant to ongoing pathological change. Methodologically, Brain-PACE extends the LILAC framework by combining spatial attention with soft label distribution learning and a Cramér distance objective, improving probabilistic performance and reducing prediction bias while providing measures of predictive uncertainty. Together, these findings support Brain-PACE as a complementary longitudinal imaging phenotype with sensitivity to relevant clinical and biological changes in early neurodegeneration.
1 Introduction
Brain-PAD captures cross-sectional brain-age deviation, but longitudinal ageing rates remain insufficiently characterized for early disease. Brain-PACE addresses this gap by directly modelling change between paired MRI scans with probabilistic longitudinal learning.
- 1 Introduction: Cross-sectional brain-age measures do not necessarily quantify person-specific rates of structural change, motivating direct longitudinal modelling.A persistently positive brain-PAD can occur without ongoing structural change, while healthy ageing and early AD-related changes may overlap.
- 1 Introduction: Existing longitudinal approaches capture progression rates, but remain limited by prediction accuracy, short intervals, and validation against neurodegenerative disease markers.Prior work includes longitudinal neighbourhood embeddings, LILAC, LILAC+, and MRI-derived pace measures linked to functional change.
- 1 Introduction: Soft label distribution learning represents age as an ordered target, whereas Cramér distance penalises errors according to their distance from the target.This addresses limitations of regression and KL-divergence for ordinal temporal prediction, including instability near certainty.
- 1 Introduction: Brain-PACE directly estimates longitudinal brain-age pace from paired T1w MRI using difference embeddings, spatial attention, distributional targets, and predictive uncertainty.The framework evaluates generalisation, direct-versus-indirect pace estimation, and relationships with clinical decline and regional tau burden in MCI.
2 Methods
The study combines standardized multi-cohort MRI preprocessing with a Siamese network that predicts longitudinal interval distributions from paired scans. Evaluation uses probabilistic, point-estimation, calibration, uncertainty, and attribution analyses.
- 2.1 Datasets and Preprocessing: MRI scans undergo brain extraction, intensity normalization, denoising, bias-field correction, and affine registration to 1mm MNI152 space.The pipeline is applied to longitudinal imaging datasets including ADNI, OASIS-3, AIBL, WRAP, and MCSA.
- 2.2 Network Architecture: The Siamese architecture processes baseline and follow-up scans through a shared SFCN backbone, subtracts their embeddings, and applies spatial attention before prediction.The backbone uses five convolutional blocks, while attention replaces the standard downstream pooling operation.
- 2.2 Network Architecture: Temporal gaps are represented across 80 bins spanning 0–8 years at 0.1-year resolution, enabling distributional rather than scalar prediction.Cramér loss compares predicted and target cumulative distributions and provides a smoother landscape than KL-divergence.
- 2.3 Model Performance and Uncertainty Evaluation: Five-fold ensemble averaging produces final pace estimates, while participant-level medians and bootstrap comparisons support model evaluation.Models are trained in PyTorch for 50 epochs, with paired differences assessed using 10,000 bootstrap resamples and FDR correction.
- 2.3 Model Performance and Uncertainty Evaluation: Performance is assessed with CRPS, MAE, MSE, APE, signed error, pace ratio, entropy-based uncertainty, and empirical prediction-interval coverage.CRPS evaluates the full predictive distribution, and coverage is measured at 50%, 80%, 90%, and 95% central intervals.
3 Results
Brain-PACE consistently favored Cramér loss with spatial attention, improving probabilistic performance, bias, uncertainty, and benchmark error relative to alternatives. Directly estimated pace identified accelerated ageing in 42.6% of ADNI MCI participants and related to tau burden, cognition, and functional impairment more strongly than indirect pace measures.
- Brain-PACE Performance Evaluation and Loss Comparison: CRPS was 0.471 years for Cramér-attention versus 0.508 years for KL-attention internally, with significant gains in APE, signed error, and pace P.MAE was numerically lower but not significant after FDR correction, whereas APE, signed error, and P remained significant.
- Brain-PACE Performance Evaluation and Loss Comparison: Cramér-attention generalized to external cohorts, with significant improvements across WRAP CRPS, MAE, MSE, APE, and pace ratio, while MCSA gains were strongest for bias and pace calibration.In MCSA, signed error was −0.012 versus 0.302 years and P was 1.074 versus 1.179; WRAP also showed lower CRPS, MAE, MSE, and APE.
- Brain-PACE Performance Evaluation and Loss Comparison: MAE decreased from 0.99 years for LILAC+ to 0.66 years for Brain-PACE, while MSE decreased from 1.97 to 0.76 years^2; locally, MAE and MSE improved by 14.9% and 28.6%.Brain-PACE also improved model fit to r = 0.88 and R^2 = 0.77 in the pair-level comparison.
- Uncertainty Evaluation: Cramér-attention produced lower normalized entropy in all groups, while WRAP 90% interval coverage was 82.7% versus 73.8% for KL-attention.All reported entropy differences were FDR-adjusted q < 0.001; the WRAP 90% coverage difference remained significant at q = 0.044.
- Healthy and Pathological Ageing: Brain-PACE classified 42.6% of ADNI MCI participants as accelerated, compared with 20.6% using indirectly derived brain-age pace.The MCI group had mean adjusted direct pace 1.33 ± 0.62, versus 1.17 ± 0.58 for indirect pace.
- Healthy and Pathological Ageing: Faster Brain-PACE correlated with regional tau burden and clinical impairment, whereas indirect pace showed no feature correlation surviving FDR correction.Tau associations were strongest in posterior cingulate, precuneus, and entorhinal cortex; FAQ, ADAS13, and CDR-SB were also associated with faster pace.
4 Discussion
Brain-PACE directly models longitudinal brain-age change and showed distinct clinical and biological associations from cross-sectional or indirectly derived pace measures. Its findings support use as a complementary longitudinal imaging feature, while important limits remain around calibration, cohort diversity, lifespan scope, and biomarker interpretation.
- Methodological contribution: Brain-PACE extends LILAC by combining spatial attention, soft label distribution learning, a Cramér distance objective, and probabilistic uncertainty estimation.Relative to LILAC+, pair-level MAE decreased from 0.779 to 0.663 years and MSE from 1.059 to 0.756 years^2.
- Methodological contribution: Brain-PACE generalised across independent healthy cohorts, but uncertainty calibration was reduced in MCI and WRAP when structural change became pathological.The authors attribute this sensitivity to cohort and pathological differences.
- Clinical and biological profile: 42.6% of MCI participants showed accelerated Brain-PACE, compared with 5–10% in healthy external groups and 20.6% using indirectly derived pace.Direct Brain-PACE was more closely coupled to selected clinical and pathological measures than cross-sectional brain-PAD or indirectly derived pace.
- Clinical and biological profile: Faster Brain-PACE was associated with greater functional and cognitive impairment and higher tau burden in the posterior cingulate, precuneus, and entorhinal cortex.The same tau comparisons were not statistically significant for indirectly derived pace or cross-sectional brain-PAD.
- Limitations and scope: The model should be interpreted primarily as a measure of later-life brain ageing because training and evaluation cohorts mainly represented older adults.The biomarker analysis was restricted to the ADNI MCI group, so the observed associations do not establish prediction of subsequent decline or specificity to AD pathology.
- Limitations and scope: Brain-PACE is intended as a complementary longitudinal imaging feature rather than a replacement for brain-PAD, volumetric, or disease-specific biomarkers.Future work should test more diverse healthcare settings and assess clinical progression, treatment response, and multimodal longitudinal models.