Retinal vessel segmentation with deep learning techniques

Abstract

The retinal vasculature is the only vascular system that can be viewed in a non-invasive manner and the characteristics of the retinal vascular offer diagnostic information. Therefore, it plays an important role in ophthalmology examination and is commonly studied in the research community. Developing automatic retinal vessel segmentation methods from retinal images is in urgent need, however, this task is difficult. The main obstacles include poor image quality, tiny structures of vessels, and various abnormalities, which increase difficulties in the retinal vessels segmentation. In recent years, deep learning techniques have emerged and led to promising progress for retinal vessel segmentation. However, this task is not fully solved. The unsolved issues include: current methods usually fail to make precise predictions for micro-vascular; predictions from CNNs based methods have a unsatisfactory graphical structure compared with the ground truth; learning-based methods highly rely on perfect annotations and flawed annotations can lead to bad performance.

In this study, we further explore deep learning techniques for retinal vessel segmentation and focus on the issues mentioned above. First, we propose a U-Net based method with deep supervision to preserve spatial information in deep layers. We specifically design a vessel-dependent loss to deal with different errors in cases of thick vessels and thin vessels. The proposed method provides better performance on vessel segmentation. Second, we design a novel method for exploring graphical structures of vessels by graph convolutional networks, which further improved the performance of vessel segmentation. The proposed method produces vessels with improved graphical structures. Third, we present a robust learning scheme under the condition of wrong annotations. A self-attention module with a label adaptation scheme increases the robustness toward label noise and enables learning with imperfect annotations. Fourth, we proposed a Transformer-U-Net hybrid model which improves segmentation results by a global receptive field. We also mitigate an imbalance issue between thin vessels and thick vessels by a dual-path decoder, in which the model learns the skeleton of vessels as an auxiliary task. The proposed model enables a effort-efficient labeling process in which annotators only need to roughly delineate vessel skeleton. The proposed method has been evaluated in different imaging modalities and achieve promising performances. In conclusion, this study will not only provide cutting-edge techniques for retinal vessel segmentation but also offer insights into the segmentation of tiny tubular structures.âǍŃ

Date of Award	2022
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Chi Keung TANG (Supervisor) & Albert Chi Shing CHUNG (Supervisor)