Replicability of unsupervised deep learning derived image phenotypes

Unsupervised deep-learning image phenotypes derived from brain MRI are propelling imaging genetics to link brain structure to genetic variation. However, their replicability across data sets has not been sufficiently evaluated, raising questions about whether they capture robust biological structure or reflect training-specific artifacts. Here, we assess the replicability of unsupervised deep-learning image phenotypes under variation in model initialization, data partitioning, and cohort, directly evaluating their stability across experimental conditions. We trained multiple models under (i) different training batch random seeds, (ii) cross-validation splits, and (iii) independent datasets (UKB and ADNI), across CNN and ViT architectures. We then derived representations from a separate UKB discovery cohort (N = 22,985) for both trained models and random initialized models without training. The representation stability was assessed using centered kernel alignment (CKA; mean ViT 0.74 vs random 0.27) and kernel canonical correlation analysis (KCCA; mean ViT 0.84 vs random 0.60), as well as genetic discovery stability using loci overlap ratio (mean ViT 0.45 vs random 0.08). We further applied weighted MAXVAR generalized CCA to 12 embeddings to extract a shared 30-dimensional subspace. Our result showed that UDIPs exhibit statistically significant stability (CKA, KCCA t test p < 0.001) across training perturbations and preserve biologically meaningful structure (loci overlap ratio t test p <0.001) across cohorts, supporting their use in imaging genetics.