Improving racial fairness in brain age models using style-transfer synthesis
- 1ETH Zurich
Abstract
Brain age models can systematically mispredict age for specific demographic groups, risking biased estimates of neurological health. We present an approach to improve accuracy and reduce racial disparities using synthetic T1 images generated via style-transfer with SuperSynth, an open-source FreeSurfer tool that produces intensity harmonized isotropic images regardless of the input's contrast or resolution. We refer to the original scans as the real domain and their SuperSynth-derived counterparts as the synthetic domain. Because each synthetic image derives from the participant's own scan, comparisons between domains hold anatomy and demographic composition constant. We audited fairness by training a neural network across thirteen racial compositions on a diverse cohort (683 White, 605 Black, 431 Asian) in both domains. Our findings highlight three key insights. First, the synthetic domain enhanced both fairness and accuracy, and even models trained on a single demographic showed reduced racial disparity purely from switching domain. The same pattern appeared in two independently developed pretrained models. Second, models trained on synthetic data demonstrated superior out-of-distribution robustness in external clinical testing. Third, augmenting imbalanced datasets with synthetic minority images closed 61% of the fairness gap without requiring new data collection, and unlike explicitly supplying race labels, does not require race as a model input at deployment. Representational analyses indicated that appearance standardization changes which features drive age prediction rather than removing demographic information from the images. By utilizing synthetic data, our approach offers a robust pathway to more equitable and generalizable brain age models.
Related
- Improving Brain-Behaviour Causality with Neurofeedback
- Piezo1 knockdown alleviates high-altitude cerebral edema by improving brain microvascular endothelial cell tight junctions via blocking the calcium influx-mediated Calpain-1 activity
- Single-cell analysis of West Nile virus infection in brain organoids identified a serine synthesis enzyme as a determinant of viral neural tropism and a druggable antiviral target
- High-Dimensional Information Transfer from the Spinal Cord to the Brain using Spinal Cord Stimulation
- The Score Is Not the Structure: Brain Alignment and Cross-Lingual Transfer
BibTeX
@misc{banchieri2026improving,
title = {{Improving racial fairness in brain age models using style-transfer synthesis}},
author = {V. Banchieri and D. Zemlyanker and J. E. Iglesias},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.16.751796},
url = {https://www.biorxiv.org/content/10.64898/2026.09.16.751796v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Improving racial fairness in brain age models using style-transfer synthesis
AU - V. Banchieri
AU - D. Zemlyanker
AU - J. E. Iglesias
PY - 2026
DA - 2026/09/16
PB - bioRxiv
DO - 10.64898/2026.09.16.751796
UR - https://www.biorxiv.org/content/10.64898/2026.09.16.751796v1
AB - Brain age models can systematically mispredict age for specific demographic groups, risking biased estimates of neurological health. We present an approach to improve accuracy and reduce racial disparities using synthetic T1 images generated via style-transfer with SuperSynth, an open-source FreeSurfer tool that produces intensity harmonized isotropic images regardless of the input's contrast or resolution. We refer to the original scans as the real domain and their SuperSynth-derived counterparts as the synthetic domain. Because each synthetic image derives from the participant's own scan, comparisons between domains hold anatomy and demographic composition constant. We audited fairness by training a neural network across thirteen racial compositions on a diverse cohort (683 White, 605 Black, 431 Asian) in both domains. Our findings highlight three key insights. First, the synthetic domain enhanced both fairness and accuracy, and even models trained on a single demographic showed reduced racial disparity purely from switching domain. The same pattern appeared in two independently developed pretrained models. Second, models trained on synthetic data demonstrated superior out-of-distribution robustness in external clinical testing. Third, augmenting imbalanced datasets with synthetic minority images closed 61% of the fairness gap without requiring new data collection, and unlike explicitly supplying race labels, does not require race as a model input at deployment. Representational analyses indicated that appearance standardization changes which features drive age prediction rather than removing demographic information from the images. By utilizing synthetic data, our approach offers a robust pathway to more equitable and generalizable brain age models.
ER -