Validation of a novel deep learning algorithm to segment geographic atrophy in spectral domain and swept source OCT en face images

Abstract

Purpose : To validate a novel deep learning algorithm (DLA) to segment and measure the size of geographic atrophy (GA) in eyes with non-exudative age-related macular degeneration (neAMD) based on the presence of choroidal hyper-transmission defects (HTDs) applicable to both spectral domain (SD-OCT) and swept source (SS-OCT) en face images.

Methods : The DLA for GA segmentation is a deep convolutional neural network modified from U-net where average pooling, rather than max pooling, is used in the encoder stage to tolerate noise and a two-step non-linear transformation employing convolution and residual connections are applied in the skip connections to propagate the encoder features to the corresponding decoder features. Subjects with normal eyes without any evidence of retinal disorders and patients diagnosed with GA secondary to neAMD were enrolled in a prospective OCT study and underwent both SD-OCT and SS-OCT imaging using 6x6 mm scan patterns. The DLA was trained with the OCT images with and without GA accurately annotated by professional graders. For validation, manual gradings were compared to automatic algorithm segmentations to assess the algorithm’s classification and segmentation performance.

Results : A total of 160 OCT en face images, including 48/32 SS-OCT images with/without GA, 48/32 SD-OCT images with/without GA, were used to train the DLA algorithm. Another independent series of 80 SD-OCT and SS-OCT en face images made of two balanced sets of 40 eyes (20 with GA, 20 without GA) were used for validation. The DLA achieved a 100% sensitivity and 100% specificity for SD-OCT and a 100% sensitivity and 95% specificity for SS-OCT, respectively. The average Jaccard Index was 0.83 for SD-OCT and 0.80 for SS-OCT. A strong positive correlation was established for the manual and automatic measurements of the GA square root areas for SD-OCT (R = 0.996, P < 0.001) and SS-OCT (R = 0.985, P < 0.001). The Bland Altman plots had a bias of -0.01 mm for SD-OCT and 0.05 mm for SS-OCT with no obvious trends.

Conclusions : The DLA had an excellent sensitivity and specificity for identifying and measuring GA in both SD-OCT and SS-OCT images. This algorithm should be useful clinically in providing quantitative measurements of GA and tracking the appearance and enlargement of GA in patients diagnosed with neAMD.

This is a 2021 ARVO Annual Meeting abstract.