Human eye retinal image classification for glaucoma screening using machine and deep learning approaches

Abstract

In modern times, non-invasive human eye examinations using retinal images play a newlinevital role in the prevention of eye disorders. Glaucoma is a leading disorder that is a set of eye impairment conditions that can lead to loss of vision due to optic nerve damage.However, manual screening is always tedious and fallible due to various factors such as mis-interpretation of disease existence, disagreements in expert opinion, and unaware of retinal images complex internal patterns. Automation of the medical field through Computer-Aided Diagnosis (CAD) can handle these limitations. Through the in-depth survey of the glaucoma state-of-the-art approaches, a few serious limitations have been identified, such as longer time vs. complexity, non-involvement of detailed inter pixel intensity relations in pattern extraction, and utilising existing or fixed architectures for the model construction. These limitations are resolved in the best possible manner in this research through the proposed approaches. newlineOur contribution extracted the significant retinal image patterns using Principal newlineComponent Analysis (PCA) and its variants using orthogonal and bi-orthogonal wavelets. newlinePrior to patterns extraction, images are enhanced using new 3D-block alpha rooting ap- newlineproach. In this approach, a trade-off between accuracy and runtime has been achieved. newlineLater, neighbour pixels intensity relations have been thoroughly analyzed using Local newlineGraph Structures (LGS) and Graph Shortest Path (GSP) to extract the image features. newlineThe input images quality has been improved using morphological procedures. The ex- newlinetracted features are classified using Wavelet Neural Networks (WNN) to test the role of newlinepixel intensity co-relationships in glaucoma identification. Subsequently, features are newlineextracted from undirected complex graphs formed by performing Deterministic Tree newlineWalk (DTW) on visually improved retinal images in wavelet domain. These features newlineare classified using Machine Learning (ML) classifiers for glaucoma screening. In this newlineapproach, an in-dept