Blood Vessel Segmentation Using Fully Convolutional Networks and ResNet50

Abstract

Retinal vessel segmentation represents a critical component in automated ophthalmological diagnosis and screening systems. This research presents a deep learning-based approach utilizing Fully Convolutional Networks integrated with ResNet50 architecture for precise blood vessel extraction from retinal fundus images. The proposed system addresses limitations of traditional image processing methods by implementing DeepLabV3+ with atrous spatial pyramid pooling to capture multi-scale vessel features. The model was trained and validated on standard benchmark datasets including DRIVE and STARE, achieving superior performance metrics with Dice coefficients exceeding 0.95 and sensitivity rates above 0.92. The deployment architecture incorporates a Flask-based web application enabling real-time segmentation accessible to clinicians without specialized technical expertise. Experimental results demonstrate robust generalization across diverse imaging conditions, establishing the feasibility of automated vessel analysis for early detection of diabetic retinopathy and cardiovascular disorders.