An Intelligent Approach for Retinal Vessels Extraction Based on Transfer Learning

An essential step in diagnosing ocular illnesses is automatically segmenting retinal blood vessels. Although current deep-learning techniques have attained excellent accuracy in artery segmentation, maintaining vascular structural connections is still tricky. The most frequent consequence of diabetes, and the main factor in adult blindness, is diabetic retinopathy. Currently, diagnosing fundus pictures is the primary method clinicians use to pinpoint the origin of diabetic retinopathy. Conducting manual screening on a large scale for retinal health is challenging. Therefore, the objective of my article is to state the segmentation of retinal blood vessels for early retinal disease detection and diagnosis. The proposed research method includes a U-net network with VGG19 and ImageNet is the backbone for extracting retinal vessels, and the extraction performance is found to be better with respect to another existing approach in the literature. The proposed methodology includes pre-processing, augmentation, patching, and model setup and implementation. First, pre-processing the available data to enhance the vessels of retinal image. The skip connections, the semantic gap left by a direct simple connection is filled using the nested connection method to combine the feature maps obtained from the intermediate decoder with the original features from the encoder. Data augmentation is also done on the original image to increase resilience and avoid the overfitting issue brought on by insufficient data. A mixed loss function is suggested to address the class imbalance in vascular pictures. After exhaustive experimental analysis observed, the Accuracy, Area Under Curve (AUC), and Jaccard index values are 0.9728, 0.9961, and 0.9175 for the DRIVE Date-set and) 0.9668, 0.9851, and 0.8975 for STARE Date-set respectively. Compared to the DRIVE data set, the proposed strategy performed better at the beginning but worse towards the end for the STARE data set. For the DRIVE and STARE data sets, the final loss values were 0.1029 and 0.1154, respectively.