Search

Abstract Understanding the genetic architecture of brain structure is challenging, partly due to difficulties in designing robust, non-biased descriptors of brain morphology. Until recently, brain measures for genome-wide association studies (GWAS) consisted of traditionally expert-defined or software-derived image-derived phenotypes (IDPs) that are often based on theoretical preconceptions or computed from limited amounts of data. Here, we present an approach to derive brain imaging phenotypes using unsupervised deep representation learning. We train a 3-D convolutional autoencoder model with reconstruction loss on 6130 UK Biobank (UKBB) participants’ T1 or T2-FLAIR (T2) brain MRIs to create a 128-dimensional representation known as Unsupervised Deep learning derived Imaging Phenotypes (UDIPs). GWAS of these UDIPs in held-out UKBB subjects (n = 22,880 discovery and n = 12,359/11,265 replication cohorts for T1/T2) identified 9457 significant SNPs organized into 97 independent genetic loci of which 60 loci were replicated. Twenty-six loci were not reported in earlier T1 and T2 IDP-based UK Biobank GWAS. We developed a perturbation-based decoder interpretation approach to show that these loci are associated with UDIPs mapped to multiple relevant brain regions. Our results established unsupervised deep learning can derive robust, unbiased, heritable, and interpretable brain imaging phenotypes.

Deep learning implementations using convolutional neural nets have recently demonstrated promise in many areas of medical imaging. In this article we lay out the methods by which we have achieved consistently high quality, high throughput computation of intra-cranial segmentation from whole head magnetic resonance images, an essential but typically time-consuming bottleneck for brain image analysis. We refer to this output as “production-level” because it is suitable for routine use in processing pipelines. Training and testing with an extremely large archive of structural images, our segmentation algorithm performs uniformly well over a wide variety of separate national imaging cohorts, giving Dice metric scores exceeding those of other recent deep learning brain extractions. We describe the components involved to achieve this performance, including size, variety and quality of ground truth, and appropriate neural net architecture. We demonstrate the crucial role of appropriately large and varied datasets, suggesting a less prominent role for algorithm development beyond a threshold of capability.