Deep Generative Modeling Based Retinal Image Analysis

Abstract

In the recent past, deep learning algorithms have been widely used in retinal image analysis (fundus and OCT) to perform tasks like segmentation and classification. But to build robust and highly efficient deep learning models amount of the training images, the quality of the training images is extremely necessary. The quality of an image is also an extremely important factor for the clinical diagnosis of different diseases. The main aim of this thesis is to explore two relatively under-explored area of retinal image analysis, namely, the retinal image quality enhancement and artificial image synthesis. In this thesis, we proposed a series of deep generative modeling based algorithms to perform these above-mentioned tasks. From a mathematical perspective, the generative model is a statistical model of the joint probability distribution between an observable variable and a target variable. The generative adversarial network (GAN), variational auto-encoder(VAE) are some popular generative models. Generative models can be used to generate new samples from a given distribution. The OCT images have inherent speckle noise in it, fundus images do not suffer from noises in general, but the newly developed tele-ophthalmoscope devices produce images with relatively low spatial resolution and blur. Different GAN based algorithms were developed to generate corresponding high-quality images fro its low-quality counterpart. A combination of residual VAE and GAN was implemented to generate artificial retinal fundus images with their corresponding artificial blood vessel segmentation maps. This will not only help to generate new training images as many as needed but also will help to reduce the privacy issue of releasing personal medical data.