Deep learning for retinal image understanding

Abstract

This project aims to aid in the improvement of automated diagnosis of retinopathy via improving structural studies of retinal vessel tree structure. To begin with, three main processes of automated diagnosis of retinopathy have been identified. The first process is the segmentation of retinal vessel tree network from retinal background of the raw retinal image input. Immediately after the segmentation, boosting algorithm is applied to the segmentation results. The second process involves taking the boosted segmentation results as input and carry out tracing of individual retinal vessel tree structure. These two processes are done with the help of cutting edge algorithm as provided by Dr Li Cheng from A*STAR BII. Following which, in the final stage of the project, a Deep Convolutional Network (DCN) has been designed with three different learning algorithms. The three algorithms are: Mini-batch Gradient Descent, Mini-batch Gradient Descent with Momentum and Mini-batch Gradient Descent with RMSProp Algorithm. This DCN enables the classification of resultant traced retinal images into two main categories: with or without retinopathy. The efficiency of each learning algorithm will be further discussed in the later part of the report.