Regional Spectral-Domain Optical Coherence Tomography Features Better Predict Frisén Scale Grades than Total Volume Alone in Papilledema

Abstract

 

Purpose

 

We have recently demonstrated a high correlation between total retinal volume measured from SD-OCT volumes of the optic nerve head (ONH) and expert-defined Frisén scale grades in papilledema subjects (Wang et al., IOVS, June 2012). However, the relationship between region-based volumetric features and Frisén scale grades was not assessed. The purpose of this work is to determine the extent that a machine-learned combination of region-based volumetric SD-OCT features can predict Frisén scale grades in papilledema.

 

Methods

 

Seventy ONH-centered SD-OCT volumetric scans (Cirrus, Carl Zeiss Meditec, Inc., Dublin, CA; 200×200×1024 voxels; 6×6×2 mm³) of 22 papilledema subjects with available expert-defined Frisén scale grades were retrospectively obtained. After applying our automated 3-D graph-theoretic approach to segment the inner limiting membrane and lower boundary of the retinal pigment epithelium (Figure 1), the following eight features were computed on each volumetric scan. Feature 1 was the total retinal (TR) volume. Features 2 and 3 were the mean circular RNFL and TR thickness, respectively. Feature 4, 5, 6, and 7 were the mean volume of nasal, superior, temporal, and inferior region, respectively (Figure 1). Feature 8 was the ratio of temporal to nasal regional volume. A random forest classifier was used to predict Frisén scale grades, with a sequential forward selection approach being used to select the optimal set of features.

 

Results

 

The mean values of the regional features at each grade are reported in Figure 2. When using TR volume alone (but not any other features), the accuracy rate (AR) of Frisén scale prediction was 38.57% and the mean Frisén grade difference (MGD) was 0.657; however, the AR and MGD improved to 62.86% and 0.414, respectively, when a combination of features was used (either feature set 4, 5, 8 or feature set 2, 3, 4, 6).

 

Conclusions

 

A machine-learned combination of region-based volumetric features from SD-OCT is better than volume alone for predicting Frisén scale grades in papilledema.

 

 

Figure 1. A) Segmentation of ILM and RPE on central B-scan from ONH SD-OCT volume. B) 3-D visualization of swollen optic nerve head with C) indicated boundary of swollen region (in red), and D) four sectors used to obtain regional volumes.

 

 

Figure 2. Mean value of region-based features at each scale grade.

 

Keywords: 613 neuro-ophthalmology: optic nerve • 612 neuro-ophthalmology: diagnosis • 549 image processing