Purchase this article with an account.
Yukun Guo, Tristan Hormel, Acner Camino, Jie Wang, David Huang, Thomas S Hwang, Yali Jia; Development and validation of a deep learning algorithm for distinguishing capillary dropout from signal reduction artifacts on OCT angiography. Invest. Ophthalmol. Vis. Sci. 2019;60(9):2208. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Avascular areas shown on retinal optical coherence tomography angiography (OCTA) can result from either capillary dropout or signal reduction effects (shadowing or defocusing). A deep learning algorithm is developed to distinguish real capillary dropout area from artifacts.
6 × 6-mm2 OCTA scans were acquired on one eye by a 70-kHz OCT commercial AngioVue system (RTVue-XR; Optove, Inc) from 180 participants in a clinical diabetic retinopathy (DR) study (76 healthy controls, 34 participants with diabetes without retinopathy, 31 participants with mild or moderate nonproliferative (NPDR) and 39 participants with severe DR). No scans were excluded due to image quality. OCTA scans were also acquired from 12 healthy volunteers, with signal strength reduction manufactured by defocus and shadow. A U-net-like convolutional neural network was built to detect the avascular area due to perfusion loss (blue in Fig. 1F) or signal reduction artifact (yellow in Fig. 1F). The input to the network contains the en fac angiogram of the superficial vascular complex (SVC), the mean projection of inner retinal reflectance (Gaussian filtered) and the thickness map of the inner retina. The ground truth of the two types of areas were manually graded by retinal experts. Six-fold cross-validation was used to evaluate our algorithm on the entire dataset.
On the tested scans from normal controls, we found that the algorithm can detect the actual foveal avascular area, as well as the areas affected by manufactured signal reduction from either defocusing with 1-3 diopters (Fig. 2A-2B), vignetting (Fig. 2C), artificial shadows (Fig. 2D), and the areas affected by natural media opacity (cataract and vitreous floaters, etc. Fig. 2E). On the tested scans from clinical DR study, we found that the algorithm can classify avascular areas related to perfusion defect or signal reduction (Fig. 2F). By averaging the results from six cross-validations, the detection accuracy of perfusion related avascular area reached 92.58±2.30% (mean ± standard deviation).
Our method can distinguish perfusion defects from the signal reduction on retinal OCTA, which will allow more scans included in the analysis and make the quantitative measurements on retinal ischemia more reliable.
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