Abstract

In this project, we first study the classification performance of four different neural network models - a custom CNN5 (5 layers convolutional layers CNN), a VGG16, a MobileNet and a custom VGG on the original OCT-retinal images. Both custom VGG and MobileNet achieve an accuracy higher than 95% on test dataset. We then pollute the retinal images with varying degrees of speckle noise which is one of the main physical parameters that may reduce the imaging quality of OCT devices. On the basis of this, we discuss the impact of different levels of noise on the performance of our classifiers. We find that different models have very different sensitivity towards speckle noise. Additionally, we try to use median filter, Lee filter and autoencoder to eliminate the noise, and research how different denoisers may influence the detection of retina damage.