ECSD-Net: A joint optic disc and cup segmentation and glaucoma classification network based on unsupervised domain adaptation.

BACKGROUND AND OBJECTIVES

Glaucoma can cause irreversible vision loss and even blindness, and early diagnosis can help prevent vision loss. Analyzing the optic disc and optic cup helps diagnose glaucoma, which motivates many computer-aided diagnosis methods based on deep learning networks. However, the performance of the trained model on new datasets is seriously hindered due to the distribution gap between different datasets. Therefore, we aim to develop an unsupervised learning method to solve this problem and improve the prediction performance of the model on new datasets.

METHODS

In this paper, we propose a novel unsupervised model based on adversarial learning to perform the optic disc and cup segmentation and glaucoma screening tasks in a more generalized and efficient manner. We adopt an efficient segmentation and classification network and employ unsupervised domain adaptation technology on the output space of the segmentation network to solve the domain shift problem. We conduct glaucoma screening task by combining classification and segmentation networks to obtain more stable and efficient glaucoma screening prediction.

RESULTS

We verify the effectiveness and efficiency of our proposed method on three public datasets, the REFUGE, DRISHTI-GS and RIM-ONE-r3 datasets. The experimental results demonstrate that the proposed method can effectively alleviate the deterioration of segmentation performance caused by domain shift and improve the accuracy of glaucoma screening. Furthermore, the proposed method outperforms state-of-the-art unsupervised optic disc and cup segmentation domain adaptation methods.

CONCLUSIONS

The proposed method can assist clinicians in screening and diagnosis of glaucoma and is suitable for real-world applications.