Deep learning approaches for segmenting Bruch's membrane opening from OCT volumes

Abstract

Automated segmentation of the eye's morphological features in OCT datasets is fundamental to support rapid clinical decision making and to avoid time-consuming manual segmentation of the images. In recent years, deep learning (DL) techniques have become a commonly employed approach to tackle image analysis problems. This study provides a description of the development of automated DL segmentation methods of the Bruch's membrane opening (BMO) from a series of OCT cross-sectional scans. A range of DL techniques are systematically evaluated, with the secondary goal to understand the effect of the network input size on the model performance. The results indicate that a fully semantic approach, in which the whole B-scan is considered with data augmentation, results in the best performance, achieving high levels of similarity metrics with a dice coefficient of 0.995 and BMO boundary localization with a mean absolute error of 1.15 pixels. The work further highlights the importance of fully semantic methods over patch-based techniques in the classification of OCT regions.