Abstract
Purpose :
Variation in peripapillary retinal deformation may help differentiate causes of optic disc swelling. Previously, we presented a machine-learning method for differentiating papilledema from NAION using a mixture of 2D/3D optical coherence tomography (OCT) features (Miller et al., ARVO 2018). In this study, we improved the previous algorithm performance by utilizing a pure set of 3D retinal shape features in a larger dataset.
Methods :
For each optic-nerve-head (ONH) OCT volumetric scan, the internal limiting membrane (ILM) and Bruch's membrane (BM) were automatically segmented; then the volume between the ILM and BM (i.e., ONHV) was computed. Next, 23 3D retinal shape models of the ILM, BM, and ILM+BM were generated using principal component analyses (PCA) based on four manual landmarks on the BM opening contour and 481 automated equidistant landmarks on the ILM and BM surfaces. An example is shown in Fig. 1. Random forest classifiers were then applied to interpret these shape features and to classify the papilledema/NAION cases. Leave-one-subject-out cross-validation was used for evaluation.
Results :
The dataset included 57 papilledema and 57 NAION subjects who were matched for ONHV from the University of Iowa Hospitals and Clinics. When the random forest classifiers considered all the 23 features from the ILM, BM, and ILM+BM 3D shape models, the overall classification accuracy rate achieved was 86% [i.e., total: 98/114; 89% (82%) for the papilledema (NAION) group]. Fig. 2 shows the feature importance and the shape variations of the top two models. Compared to the classifiers only using the traditional 2D and 3D BM shape features, the accuracy rates dropped to 73% and 77%, respectively.
Conclusions :
Our proposed method improves the differentiation of papilledema from NAION eyes using 3D retinal morphological information in regular ONH OCT volumetric scans. This study also sets up a foundation for future efforts to investigate the association between retinal shape change and intracranial biomechanical stress/strain.
This is a 2020 ARVO Annual Meeting abstract.