Source-linked AI summary
Population Structure Analysis of an Inbred Population using Quantitative Shape Phenotyping from Stereo Retinal Photographs
Li Tang, Michael D Abramoff
TL;DR
Limited phenotypic evidence constrains analysis of population structure and genetic risk factors. This study quantifies three-dimensional optic nerve head shape from retinal photographs, identifies hierarchical features, and finds patterns of relationships within and between groups that may support genetic-risk discovery.
Problem
Limited phenotypic evidence constrains analysis of population structure and discovery of genetic risk factors for disease.
Method
The study reconstructs three-dimensional optic nerve head shape from stereo fundus photographs and quantifies it into hierarchical feature vectors.
Results
The analysis examines relationships within and between groups and reflects patterns of population structure.
Takeaways & Limitations
The approach may potentially lead to new genetic risk factors for glaucoma and other eye diseases.
Takeaways & Limitations
The unlabeled and labeled datasets may not be large enough to fully explore the deep neural network’s expressive power.
Abstract
from arXiv · showhide
The population structure of an inbred population of 781 people on Norfolk Island in the Pacific, 318 of which are descendants of the original Mutineers of the Bounty, is analyzed phenotypically using shape from stereo retinal fundus photographs. Three-dimensional optic nerve head (ONH) shape is reconstructed from stereo pairs by a multi-scale stereo matching algorithm. Using deep neural network, the shape of ONH, which is under genetic control, is decomposed into a set of hierarchical features through self-taught learning. Features captured at different levels are selected according to their discriminant power in identifying the two populations. The prediction accuracy is evaluated with stratified cross validation. Given the selected feature set, individuals are grouped into k hierarchical clusters and cluster membership fractions are determined for k=2,3,4,5,6,7. Population structure analysis on the basis of phenotypes through image analysis allows heritability and linkage analysis, including founder effects from English and Polynesian ancestors, potentially leading to new genetic risk factors for glaucoma and other ONH-related eye diseases.
1 Department of Ophthalmology and Visual Sciences, University of Iowa, Iowa City, IA,
The listed affiliations include the Department of Ophthalmology and Visual Sciences and the Stephen A Wynn Institute for Administration Medical Center in Iowa City, IA, USA.
- The affiliations identify the Department of Ophthalmology and Visual Sciences and the Stephen A Wynn Institute for Administration Medical Center in Iowa City, IA, USA.
1 Introduction
The study addresses the unattempted use of complex retinal image phenotypes for population structure analysis by quantifying heritable three-dimensional optic nerve head shape from stereo photographs. It applies stereo reconstruction and deep neural network features to a Norfolk Island inbred population and validates structure against Mutineer ancestry.
- Motivation: Population structure analysis links individual similarities and dissimilarities to genetic risk factors because gene distributions are associated with population structure.Phenotypic variation can also be associated with putative genotypes and help discover new genes.
- Research gap: High-dimensional population structure analysis from image-based facial or retinal morphology had not previously been attempted.Earlier phenotypic analyses used single or low-dimensional measurements or laborious manual multidimensional traits.
- Research gap: Heritable ONH shape contains complex variations that are difficult for experts to access and quantify, while stereo photography is accessible but traditionally subjective and imprecise.OCT quantifies ONH topography but is unavailable in most clinics and lacks easily collected longitudinal datasets over years or decades.
- Approach: The study reconstructs three-dimensional ONH shape from stereo retinal photographs using robust multi-scale correspondence and condenses disparity maps with a deep neural network into hierarchical features.The approach estimates stereo shape despite spatially varying reflectance, blur, noise, low contrast, and limited illumination.
- Study contribution: The study determines morphological phenotypes and population structure in an inbred Norfolk Island sample, validating the approach by whether individuals trace ancestry to the Mutineers.The population is suited to testing quantitative-genetic expectations for admixture events and gene flow.
2 Subjects and Methods
Three-dimensional ONH shape was reconstructed from stereo retinal photographs and validated against SD-OCT, while twin data indicated strong genetic determination of ONH shape variability. A stacked autoencoder transformed disparity maps into hierarchical features for quantitative shape analysis.
- Stereo reconstruction and validation: 15.9 ± 8.8% of the cup depth was the RMS difference between normalized stereo-reconstructed and SD-OCT ONH structures.This quantitative comparison confirmed that stereo retinal images faithfully reproduced ONH topography.
- Genetic control of ONH shape: Approximately 80% of ONH shape-parameter variability was determined genetically in a sample of 172 subjects including 45 monozygotic and 41 dizygotic twin pairs.The study quantified inherited three-dimensional ONH shape parameters using 344 eyes from the Twins Eye Study in Tasmania.
- Hierarchical feature extraction: Disparity maps were normalized, left-eye maps were flipped to match right-eye ONH shape, and a stacked autoencoder projected them into hierarchical feature layers.The maps were represented as rescaled vectors, then encoded through learned weights and biases in a deep neural network.
- Autoencoder training: The stacked autoencoder was pretrained layer by layer, unrolled into symmetric encoder and decoder networks, and fine-tuned by minimizing reconstruction error.Its cost function included mean squared reconstruction error, weight decay to prevent overfitting, and a sparsity penalty.
- Feature representation: The trained network decomposed ONH variability into basis components whose outputs formed feature vectors characterizing distinct shapes with minimal information loss.These hierarchical features enabled quantitative topographic analysis while preserving information from the original structure.
3 Results
The stacked autoencoder captured ONH shape hierarchically, preserving anatomical variation while removing high-frequency noise. Selected features predicted the two populations modestly, and clustering revealed shared population structures of roughly five clusters.
- SAE feature learning: Recovered maps retained major ONH anatomical shape variations while removing high-frequency noise, indicating that SAE features can also support image denoising.The network progressively encoded higher-level ONH structure in a hierarchical feature space.
- Population discrimination: The selected features came from the first and second autoencoders, combining localized and global representations to distinguish subtle ONH differences between populations.Third-autoencoder features were global and insufficiently discriminative for subtle localized population differences, while most selected features were local.
- Population structure: Clustering by ONH shape proximity identified two clusters and features in superior and inferior sectors that may indicate important genetic compositions.The clustering distribution across the two populations showed shared population structures consisting of roughly five clusters.
- Population structure: Greater within-group variation than expected was attributed to long-range gene flow, whereas less variation than expected was attributed to genetic isolation.These interpretations accompanied the observation that the populations shared common structures.
4 Discussion
The study shows that population structure and complex morphological variation in an inbred population can be analyzed from stereo retinal images using reconstructed ONH shape and learned hierarchical features. The approach supports genetic and disease-related investigations while remaining limited by sample size and the complexity of admixture effects.
- Method: Three-dimensional ONH shape from stereo fundus photographs was quantified as hierarchical feature vectors learned by a deep neural network.A multi-scale stereo matching algorithm produced the disparity-map representation, which was then used for progressively higher-level feature learning.
- Method: Selected features were used to group individuals into continuous hierarchical clusters and determine cluster membership fractions from centroid distances.The analysis assessed differentiation among clusters and patterns of morphological variation within and between groups.
- Limitations: Admixture can produce deviations from expected patterns and integrated shape changes that cannot be explained exclusively by a gene-flow model.Comparison of Pitcairn and Norfolk Island population structures indicated admixture effects on the genetic architecture of traits.
- Implications: The ONH phenotype is largely under genetic control, supporting analysis of isolated-population genetic composition and potential risk factors for glaucoma and other eye diseases.The discussion links phenotypic ONH-shape analysis to developmental mechanisms, gene discovery, and ONH-related disease risk.
- Limitations and future directions: The preliminary approach is constrained by dataset size, but automated three-dimensional ONH estimation and quantitative image phenotyping make larger-scale studies possible.The proposed framework may support population-structure discovery from other stereo images, including faces, and future pedigree analyses require larger samples.