Purchase this article with an account.
Mohammad Shafkat Islam, Jui-Kai Wang, Matthew J. Thurtell, Randy Kardon, Mona K Garvin; Retinal Fold and Peripapillary Wrinkle Segmentation from En-Face OCT Images using Deep Learning. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4032.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The biomechanical stresses associated with optic disc swelling from increased intracranial pressure can give rise to peripapillary wrinkles and/or retinal folds. Previously, Agne et al. (OMIA 2017) developed a traditional machine-learning-based approach to segment the folds and wrinkles using custom-designed features. In this work, we implemented a deep-learning-based automated approach to perform pixel-level segmentation of folds and wrinkles from optic nerve head (ONH) centered en-face OCT images without the need to develop custom-designed features.
For each OCT scan, 2D en-face images corresponding to the internal limiting membrane (ILM) (averaging intensities of 7 pixels above and below the ILM) and retinal pigment epithelium (RPE) were generated based on a custom 3D graph-based segmentation. In order to increase the visibility of folds, histogram equalization and contrast limited adaptive histogram equalization (CLAHE) strategies were used to generate two additional ILM en-face images. Since folds, wrinkles, and vessels are tubular structures, binary vessel maps generated from a previously developed deep-learning-based approach (Islam et al., ARVO 2019) and RPE en-face images were used to maximize the vessel information and reduce the fold/wrinkle segmentation false positive rate. These five images were stacked together as the input to the deep neural network (U-Net) (shown in Fig 1). This network provides an output folds/wrinkles probability map for each pixel. The folds were also manually traced (from each set of five input images) to serve as the reference standard for training and evaluation.
A mean area-under-receiver-operating-characteristic curve (AUC) of 0.86 was achieved from a leave-one-out cross-validation of 29 OCT volumes from 29 subjects with optic disc swelling. Fig 2 shows the qualitative results of segmentation. The proposed approach effectively avoided recognizing vessels as folds and wrinkles.
Our approach successfully segments retinal folds and peripapillary wrinkles at the pixel level. The proposed approach can be used to automatically identify the presence and categorize the folds to help diagnose different causes optic disc swelling and differentiate it from pseudopapilledema.
This is a 2020 ARVO Annual Meeting abstract.
Fig 1: Deep neural network (U-Net) architecture
Fig 2: Results of folds and wrinkles segmentation in three subjects with optic disc swelling.
This PDF is available to Subscribers Only