How to Distill Disease-Differentiating, Quantitative Phenotypic Patterns from OCT Data

Abstract

Purpose : The widespread adoption of ocular coherence tomography (OCT) has provided retina clinics with large collections of three-dimensional image data; however, identifying clinically relevant, disease-specific phenotypic patterns remains challenging. We present a computational workflow for identifying phenotypic patterns from curated OCT datasets, applied to patients with molecularly confirmed mutations in CRB1 and RS1 (see figure).

Methods : Research-consented subjects were recruited for IRB-approved studies in accordance with the tenets of the Declaration of Helsinki. We retrospectively analyzed OCT data from subjects with CRB1 mutations (n=15), subjects with RS1 mutations (n=12), and control subjects (n=15). We segmented 11 retinal layers using the Iowa Reference Algorithms and computed mean layer thickness values across the 9-region ETDRS grid. Using the 99 layer-by-region values for each patient, we performed principal component analysis to identify major patterns of variation and constructed a set of multinomial logistic regression models to predict the genotype of every patient. To facilitate our computational workflow, we developed ‘heyexr’, a open-source software package for the R programming language with functions for importing, transforming, and visualizing Heidelberg data and Iowa Reference Algorithms segmentation files.

Results : The first principal component captures most of the variation in the outer zone of the outer segment and the ellipsoid zone of the inner segments, whereas the second principal component captures the variation in the thickness of the inner nuclear layer. Of the logistic regression models we tested, the combination of the first four principal components maximized the overall accuracy to predict patient group (accuracy = 93%; 95% confidence interval, 81-99%; p = 1.2×10⁻¹⁴) while minimizing the information lost in modeling.

Conclusions : We successfully distilled disease-differentiating patterns from OCT data, illustrating the utility of our computational workflow. Our ‘heyexr’ package is available on GitHub (https://github.com/barefootbiology/heyexr).

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.