Deep Learning Based Automated Extraction of Intra-Retinal Layers for Analyzing Retinal Anomalies (P-0330) (MFN 6811)

Abstract

Blindness eliminates a person vision and retinal abnormalities are the second biggest reason of blindness across the globe. Each retinal abnormality affects one or more retinal layers or retinal surface. So, if we want to diagnose a particular retinal abnormality we have to extract the retinal layers to analyze the effects of that particular abnormality. Manual extraction of these retinal layers is possible but it is a cumbersome and time consuming task. So, in our final year project we proposed a computer aided self-diagnostic system that could help the ophthalmologists in mass screening of retinal patients. In this project, we have proposed the algorithms which are based on convolutional neural network (CNN). We have used two different approaches to use the CNN along with other techniques. In first method, we have combined CNN with structure tensor and tenned this method as convolutional neural network and structure tensor-based segmentation framework (CNN-STSF). In the first method, AlexNet has been utilized which is trained on more than 1000 patches of retinal layers through transfer learning. Afterwards, it classifies each retinal layer patch which is passed to it. In second approach, we have used Gabor filter along with CNN to find the boundaries of each retinal layer. We have also used the flattening technique before Gabor Filter to increase the horizontal shape of retina by flattening the curvature of retinal layers and then before plotting the lines on retinal layers, we have de-flattened the layers to restore their original shape as well as to represent the boundaries of retinal layers on original OCT scan. Both approaches have been tested and validated on OCT scans from local Armed Forces Institute of Ophthalmology (AFIO) dataset and publicly available OCT dataset from Vision and Image Processing (VIP) lab, Duke University USA.