Retinal vessel segmentation in fundus images using deep learning / Chen Chun Hui

Chen , Chun Hui (2021) Retinal vessel segmentation in fundus images using deep learning / Chen Chun Hui. Masters thesis, Universiti Malaya.

	PDF (The Candidate’s Agreement) Restricted to Repository staff only Download (637Kb)
Preview	PDF (Thesis M.A) Download (1644Kb) | Preview

Abstract

Retinal vessel is the only microvascular system that can be viewed from digital fundus cameras directly and non-invasively, they are closely relative with human blood circulation, so the appearance and change of retinal blood vessels can reflect cardiovascular and cerebrovascular diseases, such as diabetic and hypertensions. In practice, observing retinal blood vessels has become a crucial step for ophthalmologist to make diagnosis and conduct timely treatment. However, retinal vessels in fundus images are difficult, tedious and time-consuming to recognize for ophthalmologist. Hence, computer-aided diagnosis was introduced to make automatic retinal vessels segmentation. Deep learning is used as it is a promising technique since its high efficiency and accuracy. In this project, we proposed a new method to conduct automatic retinal vessel segmentation. Firstly, we enhanced the quality of raw fundus images by using image processing technique, which the pre-processed images present a better quality than before. Secondly, we proposed a deep-learning based model to make predictions. Inspired by U-net and ensemble learning, our model is comprised of two cascaded U-shaped networks, and each of the sub-network is composed of CBR (Conv., BatchNormalization, ReLU) blocks. The second sub-network aims to fine-tune coarse vessel maps produced by the first sub-network, since it learns features from combination of coarse vessel map and raw input. To enlarge the receptive field, dilation convolution was adopted with whose dilated rates arranged deliberately to make dense sampling. In addition, residual learning was adopted to ease the optimization and sufficient skip connections were added between the two sub-networks to make full use of feature maps. Finally, three public databases were chosen to verify the proposed model and compared its performance with other recent publications. The model produced an accuracy of 0.9552/0.9699/0.9642, an AU_ROC of 0.9787/9852/9846, a sensitivity of 0.8211/0.8466/0.8395 on DRIVE, STARE and CHASE_DB1 databases, respectively. Cross-validation was also conducted to evaluate the generation capacity of the proposed model. In these intensive experiments, the proposed model can produce a good performance after training, so it can provide a good reference for ophthalmologist to perform diagnosis

Actions (For repository staff only : Login required)