Purchase this article with an account.
John William Miller, Jui-Kai Wang, Matthew Thurtell, Randy H Kardon, Mona K Garvin; Differentiation between papilledema and nonarteritic anterior ischemic optic neuropathy using retinal layer shape and regional volume features in spectral-domain optical coherence tomography. Invest. Ophthalmol. Vis. Sci. 2018;59(9):2196.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Deformation of Bruch’s membrane (BM) may reflect severity of optic disc edema due to raised intracranial pressure (ICP) in cases of papilledema. Nonarteritic anterior ischemic optic neuropathy (NAION) also causes optic disc edema but without raised ICP. We trained a random forest classifier (RFC) to distinguish between papilledema and NAION using BM and newly developed inner limiting membrane (ILM) shape models along with regional volume measures.
Volumetric spectral-domain optical coherence tomography (SD-OCT) scans from 20 subjects (10 papilledema, 10 NAION) at the University of Iowa were grouped into 10 volume-matched pairs, ensuring differences in shape between groups were not driven by volume. The principal components (PCs) from five previously-trained shape models (four 2D models of the BM and ILM in superior/inferior and temporal/nasal orientations and one 3D model of the entire BM surface) were used to calculate 35 shape features for each of the 20 scans. The normalized regional volumes of the peripapillary circle contributed four additional features. A RFC was trained on all scans and features and evaluated using 100 repetitions of leave-one-out cross-validation. The RFC was then retrained and evaluated using only the six most important features as determined by the Gini metric (Fig. 1A).
When trained on all 39 features the RFC’s average accuracy was 79.0 ± 3.5% (SD). The four most important features were the first PC of each BM model and the seventh PC of the superior/inferior ILM model (Fig. 1B). The fifth and sixth most important features were the volumes of the temporal and inferior quadrants (Fig. 1B). After retraining on only these six features, the RFC’s accuracy increased to 81.2 ± 3.6%. Figure 2 shows the discriminating ability of these six features for both classes.
Our results confirm the utility of retinal layer shape information for differentiating between causes of optic disc swelling. As we collect more data and incorporate more meaningful features, we will be able to better estimate our classifier accuracy and improve its robustness.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
(A) Sorted feature importance (B) Illustration of each feature. For features 1-4, the model's mean shape and +/- 3 deviations are shown.
Discriminating ability of the six most important features
This PDF is available to Subscribers Only