Generalizable and explainable deep learning for brain MRI: a multi-cohort evaluation of 3D architectures for age and sex prediction.

Abstract

Deep learning (DL) methods increasingly outperform classical approaches in brain MRI analysis, yet their generalizability across independent imaging cohorts remains insufficiently evaluated. Because age and sex are fundamental neurobiological factors influencing brain structure and disease risk, this study systematically compares three three-dimensional architectures-Simple Fully Convolutional Network (SFCN), DenseNet121, and Swin Transformer-for age and sex prediction using T1-weighted MRI from four independent cohorts: UK Biobank (UKB, n = 47,390), Dallas Lifespan Brain Study (DLBS, n = 132), Parkinson's Progression Markers Initiative (PPMI, n = 108 controls), and Information eXtraction from Images (IXI, n = 319). SFCN consistently demonstrated the most robust performance. For sex classification, it achieved an AUC of 1.00 [1.00-1.00] in the UKB internal test set and 0.85-0.91 across external cohorts. For age prediction, SFCN achieved a mean absolute error (MAE) of 2.66 years (r = 0.89) internally and 4.98-5.81 years (r = 0.55-0.70) externally. Pairwise DeLong and Wilcoxon tests with Bonferroni correction confirmed significantly better performance of SFCN compared with Swin Transformer in most cohorts (p < 0.017). No significant demographic subgroup biases were observed. Explainability analyses further showed task-specific and spatially consistent attention patterns across cohorts. These findings demonstrate that simpler convolutional architectures can generalize more reliably than more complex attention-based models in multi-cohort settings. The study highlights the importance of external validation and emphasizes potential trade-offs between model complexity, robustness, and interpretability for clinically relevant neuroimaging applications.