Opt Express. 2022 Apr 25;30(9):14723-14736. doi: 10.1364/OE.452208.
The foveal avascular zone (FAZ) is sensitive to retinal pathological process in the macular fovea area. For the purpose of efficient FAZ 3D quantification, we firstly propose a priors-guided convolutional neural network (CNN) to provide a tailor-made solution for 3D FAZ segmentation for optical coherence tomography angiography (OCTA) images. Location and topology priors are taken into account. The random central crop module is utilized to restrict the region to be processed, while the non-local attention gates are contained in the network to capture long-range dependency. The topological consistency constraint is calculated on maximum and mean projection maps through persistent homology to keep topological correctness of the model's prediction. Our method was evaluated on two OCTA datasets with 478 eyes and the experimental results demonstrate that our method can not only alleviate the over-segmentation prominently but also fit better on the contour of FAZ region.
黄斑中心凹的无血管区（FAZ）对黄斑中心凹的视网膜病理过程很敏感。为了有效地进行 FAZ 的 3D 量化，我们首先提出了一种基于先验的卷积神经网络（CNN），为光学相干断层扫描血管造影（OCTA）图像的 3D FAZ 分割提供了一个定制的解决方案。该方法考虑了位置和拓扑先验。随机中心裁剪模块用于限制待处理的区域，而网络中包含的非局部注意门用于捕获长程依赖关系。拓扑一致性约束是通过持久同调来计算最大和平均投影图上的，以保持模型预测的拓扑正确性。我们的方法在两个包含 478 只眼睛的 OCTA 数据集上进行了评估，实验结果表明，我们的方法不仅可以显著减轻过分割的问题，而且还可以更好地适应 FAZ 区域的轮廓。
