Purchase this article with an account.
Bryan Dev Gopal, Tingwei Zhang, Anthony Norcia, Jeffrey L Goldberg, Alfredo Dubra, Hao Zhang, Brian Soetikno; Automatic retinal layer segmentation of visible-light optical coherence tomography images using deep learning. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2069 – F0058.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Visible-light optical coherence tomography (vis-OCT) is an emerging imaging modality that can provide retinal imaging with an axial resolution of ~1.2 microns. Accurate retinal layer segmentation has yet to be established for vis-OCT. In this study, we built an end-to-end, deep learning-based segmentation algorithm to obtain boundaries for 12 retinal layers in human vis-OCT images.
The dataset consisted of 96 Vis-OCT B-scan images of the macula from human participants collected using the Aurora X2 vis-OCT system (Opticent Health, Evanston IL). Fourteen areas of the retina (12 layers + 2 backgrounds) were manually annotated. The images were resized to a height of 512 pixels and divided into strips of 32 pixels in width for training. The images were augmented with linear contrast, scaling, rotation, and translation. The final augmented dataset was split into training, validation, and test sets at a ratio of 60:20:20, respectively. A four-level U-net was then constructed with 5x3 pixel kernel sizes. The deep learning model was trained for 60 epochs with an Adam optimizer.
Fig. 1A shows an example vis-OCT image from the testing dataset. The image was used as an input to the deep learning model, which produced as a segmented image with 14 labels as output. The boundaries of each layer segmentation were determined and plotted in Fig. 1B. The combined average dice coefficient for the segmented layers was 0.90 +/- 0.06 (max: 0.98, min: 0.76).
We successfully created an automatic retinal layer segmentation algorithm that utilized an end-to-end deep learning approach for vis-OCT images. We characterized the accuracy of the modelmodel's accuracy on our testing dataset and demonstrated its utility in investigating the sub-laminae of the inner plexiform layer. This study serves as the foundation for future work on automatic segmentation of vis-OCT images.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
Fig. 1A. Vis-OCT B-scan of the human macula. Fig. 1B. Thirteen-layer segmentation boundaries in rainbow colors. Retinal nerve fiber layer (RNFL), ganglion cell layer (GCL), inner nuclear layer (INL), outer plexiform layer (OPL), outer nuclear layer (ONL), external limiting membrane (ELM), photoreceptor inner segment / outer segment junction (IS / OS), cone outer segment tip (COST), rod outer segment tip (ROST), retinal pigment epithelium (RPE), and Bruch's membrane (BM).
This PDF is available to Subscribers Only