One of the consequences of diabetes is glaucoma, which can lead to blindness. It is expensive and time-consuming to conduct tests for glaucoma manually in its earliest phases. A CNN model is used to obtain the high-level attributes. Utilizing the geographic region of Interests (ROIs) of pathologies extracted through the annotation fundus images, the primary layer of an CNN model that has been trained is updated. Subsequently, the model is fine-tuned to ensure the low-level layers detect the lesion’s and healthy regions’ regional architecture. We replace the fully linked layer (FC), and these encodes high-level attributes that are global in scope as well as domain-specific, about an entirely fresh FC layer so as to obtain discriminatory features based on fundus pictures unsupervisedly. By substantially decreasing the model variety in this stage of the process, the problem of overfitting that determines to reveal diabetes has occurred is eliminated. The categorization for every pixel in the image is combined to establish how well it is regular or aberrant, resulting in the final fragmented visualization of the blood vessels within the Fundus image. The precision attained in our work using the suggested methods is superior.