Hybrid Deep Learning Framework for the Analysis of Ganglion Cell Complex and Optic Nerve Head from SD-OCT Scans

Abstract

Abstract Glaucoma is an eye condition that occurs due to increase in intra-ocular pressure which damages the optic nerve. It is a progressive degenerative optic neuropathy and optic nerve damage is irreversible. Glaucoma is world’s leading cause of irreversible blindness, but can be prevented with proper and timely treatment. It’s prevention and treatment has been a major focus of international directives and different latest imaging tool and techniques have been developed for early detection and monitoring of glaucoma. Optical Coherence Tomography (OCT) is one of the most advanced imaging technique for detection and analysis of various retinal diseases. OCT is now widely being used for detection, analysis and monitoring of glaucoma but in most of the scenarios a detailed manual effort is required. Different researchers and leading research groups across the globe are working in this area to assist ophthalmologists for reliable and early detection of this disease. However, detailed automated analysis of OCT scans to provide more useful and in depth clinical insights is still an area to explore. In this research, we present a robust framework for detailed analysis of OCT scans which provides insights related to major retinal layers directly associated with glaucoma to help ophthalmologists in glaucoma screening, grading and progress monitoring. The propose framework is divided into two main modules where the first module focuses on extraction of cup to disc ratio (CDR) as a clinical indicator and second module deals with glaucoma monitoring through analysis of ganglion cell complex (GCC) and lamina cribrosa (LC) regions. The first module classifies the input OCT scan into health and glaucomatous image based on CDR estimation. The CDR value is calculated using the inner limiting membrane (ILM) and retinal pigmented epithelium (RPE) layers of retina. The proposed framework uses structure tensors to extract candidate layer pixels, and a patch across each candidate layer pixel is extracted, which are further classified as ILM and RPE using convolutional neural network (CNN). These layers are further refined using graph search by addressing missing and noisy points. The clinically defined geometry of ILM and RPE layers is used to compute CDR value and diagnose glaucoma. The second module performs classification into healthy and glaucoma image, and grading into early and advance cases based on GCC analysis. The proposed framework encompasses a hybrid convolutional network that extracts the retinal nerve fiber layer (RNFL), retinal ganglion cell (RGC) with the inner plexiform layer (IPL) and GCC regions. The thickness profiles of these extracted regions are computed and their mean values are passed as a feature vector to the supervised support vector machines for grading the screened glaucomatous scan as either early suspect or a severe case. Both modules of proposed framework are validated using a local OCT dataset from 196 patients, a subpart of which has been made publicly available to other researchers. The first module is able to extract ILM and RPE with absolute mean error of 6.01 and 5.44 pixels, respectively, and it finds CDR value within average range of ± 0.09 as compared with glaucoma expert. The second module of proposed framework achieves the F1 score of 0.9577 for diagnosing glaucoma, a mean dice coefficient score of 0.8697 for extracting the RGC regions and an accuracy of 0.9117 for grading glaucomatous progression. Furthermore, the performance of the proposed framework is clinically verified with the markings of four expert ophi thalmologists, achieving a statistically significant Pearson correlation coefficient of 0.9236. The proposed framework contributes by providing quantitative assessment of structural abnormalities like CDR, GCC and LC to detect, analyze and grade glaucoma using OCT scans