Neurodegenerative Disease Prediction via Transferable Deep Networks

Advances in machine learning (ML) techniques allow broader and more accurate predictions of neurodegenerative conditions, and brain networks extracted from magnetic resonance imaging (MRI) play an important role in medical image analysis. However, data scarcity is common in brain datasets, due to the high expenses, ethical regulations, and missing visits of subjects. Data scarcity makes it harder to apply many ML techniques, particularly deep learning methods, to brain network analysis. To address this challenging problem, we propose a new semi-supervised domain-adaptation method to integrate heterogeneous brain network datasets for neurodegenerative condition prediction. We use a transferable batch normalization approach for deep neural networks to avoid problems related to domain shift in the data distributions. Guided by medical knowledge, we not only use labeled data from heterogeneous datasets but also employ unlabeled data via virtual adversarial training. Experimental results show that our approach outperforms baseline and related methods for predicting important clinical scores and biological brain ages. In particular, our approach enables the transfer of domain knowledge across brain network datasets, and the transferability can boost the prediction ability of ML models.