Color-Fundus-Feature-Based Prediction of Regional SD-OCT-Based ONH-Volume in Optic Nerve Edema

Abstract

 

Purpose

 

Optic nerve edema is commonly assessed via direct funduscopic observation or digital fundus photographs using a six-stage Frisén grading scale. While 3D image-analysis of spectral-domain optical coherence tomography (SD-OCT) can provide volumetric measures of optic nerve head (ONH) swelling (Wang et al., IOVS 2012), developing automated objective quantitative measures of the degree of swelling is important in situations where SD-OCT is not commonly available, such as in emergency rooms. The purpose of this work is to develop and evaluate a machine-learning approach for the prediction of regional SD-OCT volumetric measures from color-fundus-photographic features.

 

Methods

 

Our previously developed 3D image-analysis approach was used to measure regional (nasal, superior, temporal, and inferior regions within 1.73 mm of the center of the ONH) from 48 ONH-centered SD-OCT volumes (Zeiss Cirrus) of 48 patients with varying stages of optic nerve edema (Figure 1a,b). Features concerning the boundary of the ONH swelling (such as the area of the bounded region and the smoothness of the boundary), global and local texture features (such as joint entropy of various image transformations and local entropy of the area bounded by the ONH swelling), and vessel features (such as edge derivative values, and a measure of vessel discontinuity) were extracted from color fundus photographs taken on the same day (Figure 1c) using our automated approach, and used in a leave-one-patient-out random forest regression to predict the volumetric information for each region of the SD-OCT image.

 

Results

 

Pearson correlation coefficients of R_nasal = 0.74, R_superior = 0.66, R_temporal = 0.68, and R_inferior = 0.67 were obtained between our fundus-based predicted measures and the actual SD-OCT-based regional measures. Scatter plots are shown in Figure 2.

 

Conclusions

 

Regional 3D volumetric information can be predicted from 2D fundus information and demonstrates the feasibility of objectively and quantitatively measuring optic nerve edema from fundus photographs alone.  

 

From top to bottom: (a) Example 3D rendering of top surface from SD-OCT image with the primary region indicated. (b) Computed volumetric measures of ONH displayed on a thickness map. (c) Corresponding fundus image.

 

 

The volume for each sub-region as predicted from this method plotted against the volume for each sub-region as computed with a segmentation of SD-OCT images