A deep learning phenotyping method for genetic analysis of 3D micro-CT data

The number of Genome-Wide Association Studies (GWAS) has been growing rapidly in recent years due to developments in genotyping and sequencing platforms. When applied to quantitative traits, these and other statistical genetics approaches require large amounts of consistently and accurately measured phenotypes. Here, we introduce a computational toolbox based on deep convolutional neural networks that we have developed to phenotype quantitative traits describing morphology from micro-CT-scan image datasets. We illustrate the use of this Deep Learning Phenotyper (DLP) on a sample set of craniofacial CT scans of 118 samples from two very closely related species of Lake Malawi cichlid fish, Maylandia zebra and Cynotilapia zebroides. We show that the pipeline constructed and implemented here is capable of measuring morphological skeletal phenotypes with high accuracy. We also demonstrate how this pipeline can be integrated with existing GWAS frameworks to identify candidate association loci. We believe the methods we present here will be valuable for groups studying quantitative morphological traits not only in fishes, but in other vertebrates using CT scan datasets.