Nonperfusion area segmentation in three retinal plexuses on wide-field OCT angiography using a deep convolutional neural network

Abstract

Purpose : To train and validate a convolutional neural network (CNN) to segment nonperfusion areas (NPA) in three retinal plexuses on wide-field OCTA.

Methods : We obtained consecutive 6×6-mm OCTA scans at central macular, optic disc, and temporal regions on one eye from 202 participants in a clinical diabetic retinopathy (DR) study with a 70-kHz OCT commercial system (RTVue-XR; Optovue, Inc). Projection-resolved OCTA algorithm was applied to remove projection artifacts in voxel. We designed a deep convolutional neural network [Fig. 1 D] to detect NPA [blue in Fig. 1 E] and distinguish from signal reduction artifacts [yellow in Fig. 1 E] from superficial vascular complexes (SVC), intermediate capillary plexuses (ICP) and deep capillary plexuses (DCP). The input to the network contains the inner retinal thickness map [Fig. 1 A], reflectance mean projection [Fig. 1 B] and en face angiograms of montaged scans at three regions [Fig. 1 C]. In the temporal region where the ICP merges with the DCP, we treated the ICP and the DCP as a single slab for segmentation and NPA measurement. Expert graders manually determined the ground truth for NPA and signal reduction artifacts. Six-fold cross-validation was used to evaluate our algorithm on the entire dataset.

Results : This study had 202 participants, including 39 healthy controls, 25 participants with diabetes without retinopathy, 59 participants with mild to moderate nonproliferative DR (NPDR) and 79 participants with severe NPDR or proliferative DR (PDR). The signal strength index (SSI) ranged from 55 to 87. On the test set, the proposed algorithm had high agreement with ground truth on NPA detection in three retinal plexuses on montaged wide-field OCTA (F-score (mean±standard deviation): SVC 0.83±0.08, ICP 0.81±0.10, and DCP 0.78±0.12). The algorithm showed high performance on both healthy controls and eyes with varying severities of DR [Fig. 2]. Shown by all scans from healthy controls, the proposed method was independent of SSI (Pearson correlation, p-value = 0.146).

Conclusions : A deep learning network can accurately segment NPA in individual retinal capillary plexuses and distinguish from signal reduction artifacts prevalent on wide-field OCTA.

This is a 2021 ARVO Annual Meeting abstract.