Abstract
BACKGROUND: This study aims to develop a deep learning-based algorithm dedicated to the automated classification of choroidal layers in en face swept-source optical coherence tomography (SS-OCT) images of the eye. METHODS: A cohort of 117 healthy subjects (117 eyes), underwent SS-OCT volume scans covering a 12 × 9 mm range. En face SS-OCT images of the choroid were acquired at 2.6 μm intervals from Bruch's membrane to the chorioscleral border. The images were classified into 5 layers using landmarks, such as the start of the choriocapillaris, the start of Sattler's layer, and the start and end of Haller's layer. The dataset consisted of 16,025 en face SS-OCT images, and model performance was evaluated using 5-fold cross validation. RESULTS: We developed the deep learning system based on ResNet, incorporating boundary-enhancing undersampling and subclass ensemble techniques. The developed algorithm successfully classified choroidal layers with a balanced accuracy of 84.30% and 92.61% within error ranges of 0 μm and 5.2 μm, respectively. CONCLUSIONS: Automated stratification of choroidal layers from en face SS-OCT images can be achieved accurately through deep learning.