Replicability of unsupervised deep learning derived image phenotypes

Area of Science:

• Neuroimaging
• Genetics
• Artificial Intelligence

Background:

• Unsupervised deep learning generates image phenotypes from brain MRI to advance imaging genetics.
• Replicability of these phenotypes across datasets is crucial but insufficiently evaluated.
• Concerns exist about whether phenotypes reflect robust biology or training artifacts.

Purpose of the Study:

• To assess the replicability and stability of unsupervised deep-learning image phenotypes.
• To evaluate phenotype stability across variations in model initialization, data partitioning, and independent datasets.
• To confirm if these phenotypes capture robust biological structure.

Main Methods:

• Trained multiple Convolutional Neural Network (CNN) and Vision Transformer (ViT) models with varied random seeds and cross-validation splits.
• Derived representations from a separate UK Biobank (UKB) cohort (N = 22,985).
• Assessed representation stability using centered kernel alignment (CKA) and kernel canonical correlation analysis (KCCA), and genetic discovery stability using loci overlap ratio.

Main Results:

• ViT models demonstrated significantly higher representation stability (mean CKA 0.74, KCCA 0.84) compared to random initialized models (CKA 0.27, KCCA 0.60).
• Genetic discovery stability was also significantly higher for ViT models (loci overlap ratio 0.45) versus random models (0.08).
• Unsupervised deep-learning image phenotypes (UDIPs) showed statistically significant stability across training perturbations and preserved biologically meaningful structure across cohorts.

Conclusions:

• Unsupervised deep-learning image phenotypes exhibit significant stability across diverse training conditions and independent datasets.
• These phenotypes preserve biologically meaningful structure, validating their utility in imaging genetics.
• The findings support the robust application of UDIPs for linking brain structure to genetic variation.