Enhancing Glaucoma Detection from Retinal Fundus Images Innovative Methods for Specularity Mitigation and Image segmentation

Abstract

newline strategy integrates mathematical models with AI techniques to address specularity as a newlinepreprocessing step and to perform segmentation and classification tasks. The research newlineexplores four approaches, which are outlined as follows: newline1. An efficient high-emphasis filter-based model for specularity mitigation involves newlinea Dichromatic Reflection Model (DRM) approach. This method focuses on the newlineseparation of specular and diffused components from fundus images, achieved newlinethrough the application of the similarity function in the DRM method. newline2. A Low-rank-based Robust Principal Component Analysis (RPCA) model is utilized newlineto attenuate the specularity problem in fundus images, to advance the accuracy newlineof glaucoma detection. newline3. Nonconvex Rank Approximation (NRA) approach is applied for segmentation of newlineOD and OC, and an ensemble model contributes to the classification of glaucoma. newline4. An ensemble model called Cycle-GAN-AE, is employed to generate high-quality newlinefundus images for automatic glaucoma detection. This involves utilizing the newlineCycle-consistency GAN (Cycle-GAN) and Auto-encoder (AE) collaboratively. newlineThe proposed models undergo testing using publicly available fundus image datasets. newlineThe outcomes of the four models are then analogized to those of state-of-the-art models, newlineutilizing performance metrics assessed for each method. The introduced specularity newlineremoval models contribute to an improvement in the quality of retinal images, consequently newlineleading to increased realism in glaucoma detection. The experimental findings newlinereveal maximum accuracy, precision, recall, and F-measure values of 0.968, 0.821, newline0.974, and 0.891, respectively using the Cycle-GAN-AE model for glaucoma detection. newlineKeywords: Fundus image, Glaucoma detection, Specularity removal, Cup to Disc Ratio, newlineOptic Disc, Feature extraction, Autoencoder, Generative Adverserial Network. newlineii newlineACKNOWLEDGEMENT newlineWith immense pleasure and a deep sense of gratitude, I wish to express my newlinesincere thanks to my supervisor Dr. M. Z. Lazarus, Assistant Professor, Department newlineof Electrical Engineering, C. V. Raman Global University, without his motivation and newlinecontinuous encouragement, this research would not have been successfully completed. newlineI am grateful to the Dean Academics, the Head of the Department, and the faculty newlinemembers for motivating me to carry out research in the University and also for newlineproviding me with infrastructural facilities and many other resources needed for my newlineresearch. newlineI wish to extend my profound sense of gratitude to my respected and lovable newlineparents, Mr. Ramakanta Lenka and Mrs. Basumati Lenka and my family members for newlineall the sacrifices they made during my research and also for providing me with moral newlinesupport and encouragement whenever required. newlineLast but not least, I would like to thank my friends for their constant encouragement newlineand moral support along with patience and understanding. newlinePlace: Bhubaneswar newlineDate: SATYABRATA LENKA newlineiii