In the current image processing field, medical image segmentation needs a lot of groundtruths, and the process of making these groundtruths is time-consuming and laborious. Thus, a novel retinal vessel segmentation algorithm without groundtruth is proposed in this Letter. The hierarchical clustering algorithm is first used to binary classify vessel and non-vessel pixels. Then classification results based on DRIVE databases are used as pseudo groundtruths to train the neural networks and transfer learning is considered for subsequent processing. Next the trained network is used as the feature extraction tool, by calculating and comparing the difference of image features between the target domain data (DRIVE database) and the source domain data (STARE, CHASE DB1, and HRF databases) extracted from the network. The data required for training is expanded based on semi-supervised clustering in this image feature space, finally the deep neural network is further fine-tuned. Experiments on the publicly available fundus image dataset DRIVE demonstrate that the proposed method outperforms many other state-of-the-art weakly supervised and unsupervised methods.

中文翻译：

---

没有groundtruth的弱监督视网膜血管分割算法

在当前的图像处理领域，医学图像分割需要大量的groundtruths，制作这些groundtruths的过程既费时又费力。因此，在这封信中提出了一种没有groundtruth 的新型视网膜血管分割算法。层次聚类算法首先用于对血管和非血管像素进行二元分类。然后将基于 DRIVE 数据库的分类结果用作伪地面实况来训练神经网络，并考虑迁移学习进行后续处理。接下来将训练好的网络用作特征提取工具，通过计算和比较目标域数据（DRIVE 数据库）与从网络中提取的源域数据（STARE、CHASE DB1 和 HRF 数据库）之间的图像特征差异。在该图像特征空间中基于半监督聚类对训练所需的数据进行扩展，最后对深度神经网络进行进一步微调。在公开可用的眼底图像数据集 DRIVE 上的实验表明，所提出的方法优于许多其他最先进的弱监督和无监督方法。