Deep contour attention learning for scleral deformation from OCT images

Swept-source optical coherence tomography (SS-OCT) is widely used to diagnose high myopia due to its advantage in imaging the ocular anatomic structures. Although the scleral deformation provides information on the risk of the high myopia, further validation of these highly promising findings in clinical studies has been limited by the current semi-automated software, which requires human input, and the automatic analysis of the scleral structure is quite challenging due to the ambiguous boundaries. To address these challenges, we propose a deep contour attention network (DCANet) for automatic segmentation of scleral deformation structure. Specifically, we design a scale-aware attention feature fusion module to achieve cross-scale feature fusion, which can facilitate the network to learn complementary information from multi-scale features. In addition, we develop a pyramid feature enhancement module to allow the network to learn global contextual features through the combination of receptive field and attention mechanism, and we also propose a boundary heatmap label to enrich boundary information. We evaluate the performance of the proposed method on two in-house SS-OCT datasets. In addition to the multiple metrics that are used for evaluating the segmentation performance, including Jaccard similarity coefficient, dice similarity coefficient and boundary distance error, we also propose length similarity coefficient and angle similarity coefficient to evaluate the length estimation and angle estimation, respectively. The experimental results show that our method can effectively improve the segmentation performance, and our DCANet achieves the overall best performance on two datasets compared with other state-of-the-art networks. Our findings motivate the development of clinically applicable deep learning systems for the prediction of high myopia progression on the basis of the scleral phenotypes from SS-OCT images.