Laboratory for Medical Imaging Informatics, Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai, 200083, China.
Laboratory for Medical Imaging Informatics, University of Chinese Academy of Sciences, Beijing, 100049, China.
J Med Syst. 2020 Mar 20;44(5):96. doi: 10.1007/s10916-020-01561-2.
Optic disc (OD) and optic cup (OC) segmentation are important steps for automatic screening and diagnosing of optic nerve head abnormalities such as glaucoma. Many recent works formulated the OD and OC segmentation as a pixel classification task. However, it is hard for these methods to explicitly model the spatial relations between the labels in the output mask. Furthermore, the proportion of the background, OD and OC are unbalanced which also may result in a biased model as well as introduce more noise. To address these problems, we developed an approach that follows a coarse-to-fine segmentation process. We start with a U-Net to obtain a rough segmenting boundary and then crop the area around the boundary to form a boundary contour centered image. Second, inspired by sequence labeling tasks in natural language processing, we regard the OD and OC segmentation as a sequence labeling task and propose a novel fully convolutional network called SU-Net and combine it with the Viterbi algorithm to jointly decode the segmentation boundary. We also introduced a geometric parameter-based data augmentation method to generate more training samples in order to minimize the differences between training and test sets and reduce overfitting. Experimental results show that our method achieved state-of-the-art results on 2 datasets for both OD and OC segmentation and our method outperforms most of the ophthalmologists in terms of achieving agreement out of 6 ophthalmologists on the MESSIDOR dataset for both OD and OC segmentation. In terms of glaucoma screening, we achieved the best cup-to-disc ratio (CDR) error and area under the ROC curve (AUC) for glaucoma classification on the Drishti-GS dataset.
视盘 (OD) 和视杯 (OC) 分割是自动筛查和诊断视神经头异常（如青光眼）的重要步骤。许多最近的工作将 OD 和 OC 分割表述为像素分类任务。然而，这些方法很难明确地对输出掩模中的标签之间的空间关系进行建模。此外，背景、OD 和 OC 的比例不平衡，这也可能导致模型出现偏差，并引入更多噪声。为了解决这些问题，我们开发了一种遵循粗到细分割过程的方法。我们首先使用 U-Net 获得一个粗略的分割边界，然后裁剪边界周围的区域，形成一个以边界为中心的轮廓图像。其次，受自然语言处理中序列标记任务的启发，我们将 OD 和 OC 分割视为序列标记任务，并提出了一种名为 SU-Net 的新型全卷积网络，并结合维特比算法共同解码分割边界。我们还引入了基于几何参数的数据增强方法，以生成更多的训练样本，从而最小化训练集和测试集之间的差异，减少过拟合。实验结果表明，我们的方法在 2 个数据集上的 OD 和 OC 分割方面都取得了最先进的结果，在 MESSIDOR 数据集上，我们的方法在 6 位眼科医生中达成了一致，在 OD 和 OC 分割方面都优于大多数眼科医生。在青光眼筛查方面，我们在 Drishti-GS 数据集上实现了最佳的杯盘比（CDR）误差和 ROC 曲线下面积（AUC），用于青光眼分类。
