|
内容記述 |
<p>The choroid is a dense vascular layer that lies between the retina and the sclera and contributes to the blood supply of the outer retina. In recent years, optical coherence tomography (OCT), which enables non-destructive acquisition of cross-sectional images of the choroid, has revealed the relationship between morphological changes in the choroid and eye diseases. In this context, automatic and accurate segmentation of OCT images is critical, but many existing methods face challenges, as they 1) rely on convolutional neural network (CNN)-based architectures, which struggle to capture long-range dependencies, and 2) primarily focus on two-dimensional OCT images and thus have difficulty identifying the complex three-dimensional (3D) structure of the choroid. In this study, we propose an automatic choroid segmentation method, 3DChoroidSwin, which incorporates 3D CNN and 3D Swin Transformer frameworks, achieving both short- and long-distance learning. Furthermore, our method uses a combined loss function that includes the boundary loss, which leverages morphological information, achieving shape-aware training and decreasing unnatural false positives. Experimental results using clinical data demonstrate that the proposed method outperforms comparison methods, delivering performance comparable to ground truth; moreover, it achieves smooth and continuous 3D segmentation with reduced segmentation errors at the choroid margins.</p> |
