Simple 3D CNN outperforms Swin Transformer for age and sex prediction on multi-cohort brain MRI

Multi-cohort brain MRI: a simple 3D CNN (SFCN) achieved AUC 1.00 for sex, MAE 2.66y for age (internal); external AUC 0.85-0.91, MAE 4.98-5.81y. Simpler architectures generalized better than Swin Transformer.

• Retrospective multi-cohort study: trained on UK Biobank (n=47,390), externally validated on DLBS (n=132), PPMI (n=108 controls), IXI (n=319).
• SFCN significantly outperformed Swin Transformer (p<0.017, Bonferroni-corrected) with no demographic subgroup biases; explainability showed task-specific attention patterns.
• Limitation: small external cohorts (total 559), age/sex prediction as proxy tasks, only T1-weighted MRI.

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