Fairness-aware AI reduces racial bias in brain MRI dementia classification

A domain-adaptation method (RegAlign) substantially reduced cross-group performance gaps in MRI-based dementia classification between Non-Hispanic White and Hispanic White populations in a multi-site study, while standard models showed significant bias when tested across racial…

• Retrospective study of 8,560 subjects across three racial/ethnic groups using T1-weighted brain MRI; external or prospective validation not reported.
• Standard models demonstrated significant cross-group bias: performance dropped when a model trained on one group was tested on another, particularly for underrepresented populations.
• The proposed RegAlign framework uses few-shot domain adaptation with source-side focal learning, target-side class-weighted supervision, and class-conditional alignment.

RadPigeon summaries are original and for information only. They are not clinical advice.