Tiwari, Anil Kumar2026-01-192026-01-192025-04-21https://ir.iitj.ac.in/handle/123456789/240Retinal diseases such as Diabetic Retinopathy (DR) and Hypertensive Retinopathy (HR) are serious health concerns. HR is a retinal disease caused by elevated blood pressure for a prolonged period. Diabetic Retinopathy(DR)is a progressive retinal disease caused by long-term diabetes. Non-proliferative diabetic retinopathy(NPDR) is an early stage of DR, damages blood vessels of the retina. Moreover, HR serves as a biomarker for several illnesses, including retinal diseases, atherosclerosis, strokes, kidney disease, and cardiovascular risks. Early identification of these diseases helps in the timely and proper treatment can prevent blindness. Manual diagnosis of these retinal diseases is time-consuming, resource demanding, inconvenient, costly, and demands specialized skills and experience. On the other hand, Computer-aided diagnosis (CAD) and Artificial intelligence (AI) based systems are expected to solve the above-mentioned challenges effectively. This thesis proposes an efficient algorithm for HR detection. Further, for effective management of these diseases, we develop an effective algorithm for severity grading of HR and NPDR.To achieve this, a careful study of existing literature in the field of automatic diagnosis of HR and DR is conducted to identify important research gaps in this area. It has been observed that the availability of public data sets for HR diagnosis is limited. Moreover, there is no such public dataset available for HR grading. Severity grading helps in the timely and effective management of these diseases and reduces future risks. However, very limited research has been conducted for the severity grading of HR and NPDR. Additionally, previous studies mainly relied on Arteriovenous Ratio (AVR) and the manual selection of regions of interest (ROI) around the optic disc (OD) for HR detection and its grading. Due to ROI selection primarily around OD, these method had risk of missing several important clinical features present elsewhere, resulting in less accurate diagnostic outcomes. Furthermore, one of the key challenges observed in DR and HR severity grading is the high-class imbalance high-class imbalance. Such class imbalance makes the training of learning based recognition models very challenging, as the majority of class samples predominate in the training process of the learning model. To address these limitations, a novel HR detection approach has been proposed, based on few-shot learning, using a pretrained initial baseline model to obtain transferable knowledge for feature embedding on few-shot prediction. This approach aims to reduce overfitting and improve generalization, which is especially advantageous for smaller datasets. Unlike previous methods, the proposed approach uses complete images to capture all clinical features. Experimental results demonstrate the effectiveness of the proposed HR detection method. In this thesis work, we develop “HRSG: Expert-Annotated Hypertensive Retinopathy Severity Grading Dataset” dataset, encompassing fundus images, and categorizing the severity into four classes: normal, mild, moderate, and severe. The grading process is conducted by three experienced ophthalmologists affiliated with prestigious medical institutions in India, including the All India Institute of Medical Sciences (AIIMS) in Jodhpur, the Aravind Eye Hospital in Madurai, and the Sri Aurobindo Institute of Medical Sciences Hospital in Indore. For effective severity grading of HR, a hybrid deep learning (DL) architecture is proposed that leverages the combined strengths of pretrained ResNet-50 and a modified Vision Transformer (ViT) enhanced with both global and locality self-attention mechanisms, enabling accurate grading into four classes: normal, mild, moderate, and severe. The proposed method effectively captures both local and global contextual information across the input image, leading to a robust and resilient classification model. Further, to address the class imbalance issue, we introduce a novel decouple representation and classifier (DRC) based training method. The proposed DRC method effectively addresses the class imbalance by improving the module’s capacity to identify effective feature learning while maintaining the original dataset’s distributional properties, leading to improved diagnostic accuracy. The extensive experimental results demonstrate the effectiveness of the proposed method in accurately grading HR severity. This thesis also presents a reliable method for NPDR severity grading into normal, mild, moderate, and severe classes. This method includes an initial image enhancement to improve quality for subsequent processing. A feature set is then developed using various descriptors, capturing rich information to identify distinct and unique properties of NPDR lesions. To address the class imbalance, we employ the Synthetic Minority Oversampling Technique (SMOTE) and an ensemble-learning-based Random Forest (RF) classifier to improve the model’s performance on the imbalance classes. Moreover, the comparison results show that the proposed method performs better than existing methods, making it suitable for the early diagnosis and effective management of NPDR. The contributions of this thesis aim to assist healthcare professionals in early HR detection, regular screening, risk stratification, and patient categorization based on HR and NPDR severity grading. These developments aim to optimize clinical decision-making, improve resource allocation, and enhance overall disease management. The proposed systems can assist clinicians in referral decisions and facilitate mass screening.XVIII;117enThesisTP00198