Purchase this article with an account.
JIE WANG, Tristan Hormel, Qisheng You, Yukun Guo, Xiaogang Wang, Liu Chen, David Huang, Thomas Hwang, Yali Jia; Robust Nonperfusion Area Detection in Three Retinal Plexuses using Convolutional Neural Network in OCT Angiography. Invest. Ophthalmol. Vis. Sci. 2019;60(11):PB086. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop a robust nonperfusion area (NPA) detection method in three retinal plexuses using convolutional neural network (CNN) in optical coherence tomography angiography (OCTA).
3 × 3-mm2 OCTA scans were acquired on one eye by a 70-kHz OCT commercial AngioVue system (RTVue-XR; Optove, Inc) from 428 participants in clinical diabetic retinopathy (DR) studies. No scans were excluded due to image quality. OCTA scans from 10 healthy volunteers were degraded with manufactured artifacts from neutral density filters and defocus. En face superficial vascular complex (SVC), intermediate capillary plexus (ICP) and deep capillary plexus (DCP) angiograms were generated on projection-resolved OCTA. The angiogram and the corresponding en face reflectance map of each plexus, as well as the inner retinal reflectance comprised the input to the designed CNN.A trained grader determined the ground truth NPA by evaluating both the angiogram and the reflectance maps. We then randomly selected 59 diabetic and 5 healthy cases with manufactured signal reductions for testing. The remaining cases were used for training the CNN. To preserve the feature resolution and detect the features in multi-scales, the dilated convolution kernels replaced the pooling layers and parallelized the output layer by extending the field of view. The softmax activation function generated the probability maps of NPA output. Pixels with probability > 0.5 were classified as NPA.
On the healthy eye scans with manufactured signal reduction, the CNN consistently detected foveal avascular zones (FAZ) as NPA, regardless of the signal strength (Fig. 1). On the scans from clinical studies, NPAs were effectively detected in DR subjects of various disease stages with minimal interference by low signal strength and bulk motions. On all testing scans with the wide range of signal strength index (SSI: 44-88), the mean intersection-over-union (mIOU) of NPA detection was 0.81±0.11 in all layers (Fig. 2), indicating the high accuracy of the algorithm. The coefficient of variation of NPA detection on healthy eyes between scans within same visit was 0.07±0.01, suggesting the high repeatability.
The proposed algorithm detects NPA with a high level of agreement with manual grading in all retinal capillary plexuses, regardless of manufactured signal strength attenuation or low signal quality in clinical studies.
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