MeshMonk: Open-source large-scale intensive 3D phenotyping

Abstract

In the post-genomics era, an emphasis has been placed on disentangling 'genotype-phenotype' 3 connections so that the biological basis of complex phenotypes can be understood. However, 4 our ability to efficiently and comprehensively characterize phenotypes lags behind our ability to 5 characterize genomes. Here, we report a toolbox for fast and reproducible high-throughput 6 dense phenotyping of 3D images. Given a target image, a rigid registration is first used to orient 7 a template to the target surface, then the template is transformed further to fit the specific shape 8 of the target using a non-rigid transformation model. As validation, we used N = 41 3D facial 9images registered with MeshMonk and manually landmarked at 19 locations. We demonstrate 10 that the MeshMonk registration is accurate, with 0.62 mm as the average root mean squared 11 error between the manual and automatic placements and no variation in landmark position or 12 centroid size significantly attributable to landmarking method used. Though validated using 19 13 landmarks for comparison with traditional methods, MeshMonk allows for automatic dense 14phenotyping, thus facilitating more comprehensive investigations of 3D shape variation. This 15 expansion opens up an exciting avenue of study in assessing genomic and phenomic data to 16better understand the genetic contributions to complex morphological traits. 18 19The phenotypic complement to genomics is phenomics, which aims to obtain high-throughput 20and high-dimensional phenotyping in line with our ability to characterize genomes 1 . The 21 paradigm shift is simple and similar to the one made in the Human Genome Project: instead of 22'phenotyping as usual' or measuring a limited set of simplified features that seem relevant, why 23 not measure it all? In contrast to genomic technologies, which successfully measure and 24 characterize complete genomes, the scientific development of phenomics lags behind. 25However, with the advent of new technologies, hardware exists for extensively and intensively 26collecting quantitative phenotypic data. For example, 3D image surface and/or medical 27 scanners provide the optimal means to capture information of biological morphology and 28appearance. Today, the challenge is to establish standardized and comprehensive phenotypic 29representations from large scale image data that can be used to study phenotypic variation in 30 the context of genetic variation 2 . This is a challenge that we address with the development of 31the MeshMonk toolbox, which enables fast and reproducible high-throughput phenotyping of 3D 32images, or quasi-landmark indication, which can be applied to 3D facial images as well as 3D 33scans of other complex morphological structures. 35Dense correspondence phenotyping is important beyond genomics and could be employed by 36anthropologists, biologists, and medical clinicians to accurately and reproducibly characterize 37anatomical structures such that underlying qualities about the structure can be understood. The 38 study of va...

Syndrome-informed phenotyping identifies a polygenic background for achondroplasia-like facial variation in the general population

“…1a). We obtained homologous facial configurations by non-rigidly mapping an atlas (composed of 7160 points) to each individual image 16 . Controls were Procrustes aligned to a common coordinate system and principal component analysis (PCA) was applied to capture the major axes of normal-range facial variation.…”

“…For both US and UK datasets separately, we combined the ACH-derived trait scores across the 58 significant segments into a single phenotype matrix ([n x m] with n US = 4680 controls, n UK = 3566 controls, and m = 58 facial segments). This phenotype matrix was tested for genome-wide SNPassociations in a multivariate association framework using canonical correlation analysis (CCA) following White et al 16 . However, instead of performing a separate GWAS per facial segment, scores generated across multiple segments were now combined into a single multivariate GWAS.…”

“…For 6,772 multi-ancestry participants of the Adolescent Brain Cognitive Development (ABCD) study 69 , the outer head surface was extracted from magnetic resonance images as described by Goovaerts et al 29 . The MeshMonk toolbox 16 was used to perform rigid and subsequently nonrigid surface registration using a full-head template comprising 28,218 quasi landmarks. From this, we cropped out the area covering the cranial vault (n = 11,410 quasi landmarks), encompassing the supraorbital ridge, and extending towards the occipital bone.…”

“…Facial phenotyping and segmentation 3D facial images were acquired using three digital stereophotogrammetry systems (3dMDface, Vectra H1, Creaform Gemini) and one laser scanning system (Konica Minolta VI-900). We non-rigidly registered an average facial atlas to each 3D image using the MeshMonk toolbox 16 to obtain a standard facial representation defined by 7160 homologous quasi-landmarks. Due to the bilaterally paired construction of the quasi-landmarks constituting the atlas, registered images were symmetrized by averaging the original configuration and its horizontally reflected copy following Procrustes superimposition.…”

Robust genome-wide ancestry inference for heterogeneous datasets: illustrated using the 1,000 genome project with 3D facial images

“…Height and weight were measured using an Accustat stadiometer (Genentech, San Francisco, CA), a clinical scale (Tanita, Arlington Heights, IL), or by self-report. 3D facial images were imported into matlab 2016b in .obj wavefront format to perform spatially dense registration (MeshMonk 45 ). This resulted in homologous spatially dense configurations of 7,160 quasi-landmarks per facial image.…”