Application of convolutional neural networks for the retinal vessel segmentation

etinal vessels segmentation is critical for diagnosing systemic diseases like diabetic retinopathy and hypertension. Automated segmentation using deep learning offers an efficient alternative to manual annotation, overcoming challenges such as variability in vessel patterns and image noise. This research evaluates four deep learning architectures (U-Net, W-Net, MU-Net, and DeepLabV3+) and ensemble of U-Net based models for retinal vessel segmentation using the DRIVE, STARE, and CHASE-DB1 datasets. Data preprocessing included green channel extraction to enhance vessel visibility, as well as the CLAHE method and gamma correction for enhanced contrast. TensorFlow and OpenCV libraries were used for models training and image augmentation. The results indicate that U-shaped networks performed better at segmenting small vessels in good-quality images. In contrast, DeepLabV3+ performed better on noisy images compared to U-shaped networks.