Deep learning has an enormous impact on medical image analysis. Many computer-aided diagnostic systems equipped with deep networks are rapidly reducing human intervention in healthcare. Among several applications, medical image semantic segmentation is one of the core areas of active research to delineate the anatomical structures and other regions of interest. It has a significant contribution to healthcare and provides guided interventions, radiotherapy, and improved radiological diagnostics. Experts believe that intelligent applications designed for medicine will soon take over the role of a radiologist. The goal of bringing semantic segmentation, specifically in healthcare, is to boost efficiency in diagnostics by labeling every pixel with its corresponding class. The core concept revolved around taking random input size and produce output with similar size and sufficient inference. Over time, researchers have proposed distinct architectures for end-to-end and pixel-to-pixel self-sufficient training. In retinal image analysis, the development of semantic segmentation techniques has opened doors for researchers to precisely extract regions of interest and automatically detect the symptoms of various retinal diseases such as diabetic and hypertensive retinopathy. These diseases are common in subjects with diabetes and hypertension. It may cause vascular occlusion and produce fragile micro-vessels in an advanced stage of neovascularization. An excessive amount of sugar in the blood and extended force on vessels rupture the newly developed micro-vessels, providing blood and other fluids leak into the retina. It may cause blindness in severe cases if not diagnosed and treated at an early stage. It is critical to find the initial symptoms, including abnormal vascular growth, hard exudates, arterial and venous occlusion, and the appearance of bulges on the outer layer of vessels. The primary step is to automate the analysis of variations, appear in vessels for detecting retinopathy. In our dissertation, we have performed a context-sensitive semantic segmentation to capture long-range dependencies and restore lossy pixels of manually annotated groundtruths. We also applied morphological image processing techniques to create masks for unmasked datasets. The method adopted literature to apply leave-one-out and k-fold strategies for unordered data distributions. The use of context information in predicting target pixels bring added precision to the vision-critical system. We also designed a fully automated screening system based on a unified modeling approach of diagnosis. The system can extract multiple ocular features with a novel semantic segmentation network to early detect the symptoms of retinal disease. We proposed a novel technique of dynamic inductive learning with single-point decision criteria, striving to optimize the image segmentation model using multi-criteria decision support feedback. It is found that dynamic inductive transfer learning reduces the subjectivity of hyperparameter selection in a model validation process. We further designed a feature-oriented ensemble network for extracting multiple retinal features. It includes a set of models that reflect feature-based needs to prevent intensity loss, micro-vessels overlap, and data redundancy. The proposed learning protocol with a minimalist approach can compete with state-of-the-art work without a performance compromise. The pitfall of previously proposed work is also addressed through the self-defined assessment criteria. Our research also proposed an architecture inspired by the generative adversarial network. The network enhanced the base model with residual, dense, and attention mechanisms. The residual mechanism helped extend the network depth without falling into the gradient problem and improved the model response by transmitting useful feature representations. The dense block helped increase the information flow for reusing feature representations that reduced the number of trainable parameters. However, the attention mechanism performed the domain-centric synthesis and helped conserve local context by highlighting fine details. Our model shown a promising response in extracting both macro and micro-vessels and reported high true positive rate and structural similarity index scores.