Purchase this article with an account.
Jui-Kai Wang, Mohammad Shafkat Islam, Yashila Marie Permeswaran, Samuel Johnson, Randy H Kardon, Mona Garvin; Automated Classification of Retinal Folds and Wrinkles in En-Face Optical Coherence Tomography Images with Optic Disc Swelling. Invest. Ophthalmol. Vis. Sci. 2019;60(9):5231.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Retinal folds and wrinkles can result from biomechanical stress/strain in cases of optic disc swelling. We have published a method to segment the folds and wrinkles around the optic-nerve-head (ONH) in the en-face optical coherence tomography (OCT) image from a thin image slab centered at the internal limiting membrane (ILM) (Wang et al., ARVO, 2018). In this study, we further classify the ILM en-face images with retinal concentric wrinkles, radial folds, or neither. This study may help distinguish the causes of optic disc swelling in ophthalmic images.
For each ONH OCT scan, the ILM and retinal pigment epithelium (RPE) complex were first automatically segmented, and the corresponding en-face images were created (Fig. 1a, b). The Bruch’s membrane opening (BMO) center was estimated; the total retinal volumes of the four peripapillary quadrants were computed (Fig. 1c). Next, all the vessels were detected by a deep-learning method in the RPE en-face image for removal (Fig. 1d). Gabor filter banks were designed to respond to the concentric wrinkles and radial folds in the ILM en-face image (Fig. 1e, f). Next, these two Gabor response maps were unwrapped using the BMO center (Fig. 1g, h) and then represented by the histograms of oriented gradients (HOG) feature descriptors (Fig. 1i, j). Finally, a random forest classifier was utilized to classify the fold types with the features of the regional retinal volumes and HOG descriptors. Fig. 1 shows an example of obtaining features from a case with the concentric wrinkles.
Of a dataset with 107 optic-disc-swelling subjects from the University of Iowa Neuro-Ophthalmology Clinic, a computer program randomly selected one OCT scan (without artifacts) per subject with either clear concentric wrinkles (14 scans), radial folds (16 scans), or neither (16 scans). With the selected most 25% important features from the classifier, the leave-one-subject-out cross-validation showed the accuracy rates: 13/16 for the no-folds group, 11/14 for the concentric wrinkle group, and 14/16 for the radial fold group (Fig. 2).
Our proposed method can classify the concentric wrinkles or radial folds from the no-folds group in the OCT ILM en-face images. This study sets up a foundation for future efforts about automated quantifications of association between the retinal folds and stress/strain.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only