Differentiation of Papilledema from Non-Arteritic Anterior Ischemic Optic Neuropathy (NAION) using 2D Image-Based Features and 3D Deep-Learning-Based Shape Features in Color Fundus Photographs

Abstract

Purpose : Previous studies have shown that optical coherence tomography (OCT) can help differentiate papilledema from NAION based on peripapillary 2D/3D structural and morphological features (Miller et al., ARVO 2018, Wang et al., ARVO 2020). However, OCT is not always available (e.g., in the emergency room and primary care settings). To overcome this limitation, in this study, we focus on the automated differentiation in more available 2D color fundus images and include novel use of deep-learning-based 3D morphological features.

Methods : Using 102 OCT and fundus image pairs (51 papilledema and 51 NAION subjects; severity matched by the optic-nerve-head [ONH] volume measured in OCT) from the University of Iowa, a feature pyramid neural network was trained to estimate the OCT-like thickness maps from the color fundus photographs (Fig. 1A). Based on these 102 estimated thickness maps, seven statistical shape models were built using principal component analysis (PCA) covering 90% of the shape variance (Fig. 1B). Then, for each input fundus image, the corresponding seven 3D shape measures were computed. Meanwhile, 12 features based on the image intensity, 12 features based on a segmented vessel map, 24 features based on the image texture analysis, and six regional peripapillary volumetric measures were also computed (Fig. 2) for a total of 61 features. A random forest classifier was used to classify papilledema/NAION cases using leave-one-subject-out validation.

Results : When the random forest classifier considered the entire feature set (feature importance of 61 features are shown in Fig. 2), the classification accuracy was 84% (44/51: papilledema; 42/51: NAION). Without considering the 3D features (i.e., the 48 pure 2D image-based features), the accuracy dropped to 80% (42/51: papilledema; 40/51: NAION).

Conclusions : Deep-learning-based OCT-like thickness maps make retinal 3D structural and morphological analysis possible in 2D color fundus photographs, which is beneficial for differentiating papilledema from NAION when OCT is not available.

This is a 2021 ARVO Annual Meeting abstract.

 

Fig 1. (A) Example of a NAION input fundus image, output deep-learning-based thickness map, and true OCT reference. (B) Example of two shape models.

 

Fig. 2. Feature examples and feature importance from the random forest classifier.