Purchase this article with an account.
Yukun Guo, Honglian Xiong, Tristan Hormel, Jie Wang, Thomas Hwang, Yali Jia; Automated volumetric segmentation of retinal fluid on optical coherence tomography using deep learning. Invest. Ophthalmol. Vis. Sci. 2019;60(11):026.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
In diabetic macular edema (DME), thickening from intraretinal fluid and thinning from neural degeneration can occur concurrently and confound the understanding of disease progression and treatment response. Current quantitative methods only evaluate overall retinal thickness. We propose a deep learning-based method to automatically detect volumetric retinal fluid on optical coherence tomography (OCT).
3 × 3-mm2 OCT scans were acquired on one eye by a 70-kHZ OCT commercial AngioVue system (RTVue-XR; Optove, Inc) from 51 participants in a clinical diabetic retinopathy (DR) study (45 with retinal edema and 6 healthy controls). A deep convolutional network with U-net-like architecture (Fig. 1) was constructed to detect and segment the retinal fluid volume. The network contains two types of input B-scans: structural OCT (Fig. 2A) and the corresponding OCTA (Fig. 2B). OCTA scans are used to enhance differentiation between retinal fluid and normal tissue. To improve the quality of OCT B-scans, two B-scans from adjacent positions were averaged. Experts manually delineated retinal fluid (red in Fig. 2C), non-fluid (black in Fig. 2C), and background area (green in Fig. 2C) on each B-scan to establish the ground truth. Data augmentation methods (horizontal flipping and Gaussian noise addition) enlarged the training data set. Six-fold cross-validation was used to evaluate the algorithm on the entire data set.
Compared to manual delineation, the algorithm detected retinal fluid on B-scans (Fig. 2D) with an accuracy of (Dice coefficient) 89.1 ± 2.4% (mean ± standard deviation) on cross-validation data. Compiling adjacent B-scans generates 3-dimensional volumes of the retinal fluid with a smooth profile (Fig. 2E).
Our deep-learning method accurately segments retinal fluid volumetrically on OCT scans, providing a more complete picture of the anatomic changes in DME.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.
This PDF is available to Subscribers Only