Abstract
PURPOSE: The purpose of this study was to develop a deep learning model for automatic binarization of the choroidal tissue, separating choroidal blood vessels from nonvascular stromal tissue, in optical coherence tomography (OCT) images from healthy young subjects. METHODS: OCT images from an observational longitudinal study of 100 children were used for training, validation, and testing of 5 fully semantic networks, which provided a binarized output of the choroid. These outputs were compared with ground truth images, generated from a local binarization technique after manually optimizing the analysis window size for each individual image. The performance was evaluated using accuracy and repeatability metrics. The methods were also compared with a fixed window size local binarization technique, which has been commonly used previously. RESULTS: The tested deep learning methods provided a good performance in terms of accuracy and repeatability. With the U-Net and SegNet networks showing >96% accuracy. All methods displayed a high level of repeatability relative to the ground truth. For analysis of the choroidal vascularity index (a commonly used metric derived from the binarized image), SegNet showed the closest agreement with the ground truth and high repeatability. The fixed window size showed a reduced accuracy compared to other methods. CONCLUSIONS: Fully semantic networks such as U-Net and SegNet displayed excellent performance for the binarization task. These methods provide a useful approach for clinical and research applications of deep learning tools for the binarization of the choroid in OCT images. TRANSLATIONAL RELEVANCE: Deep learning models provide a novel, robust solution to automatically binarize the choroidal tissue in OCT images.