Automated blood vessel segmentation based on de-noising auto-encoder and neural network

Retinal vessel segmentation has been widely used for screening, diagnosis and treatment of cardiovascular and ophthalmologic diseases. In this paper, we propose an automated approach for vessel segmentation in digital retinal images based on de-noising auto-encoders layer-wise initialized neural networks. The proposed method utilized a deep neural network, which is layer-wise initialized by de-noising auto-encoders and fine-tuned by BP algorithm, to segment vessel structures in retinal images. The proposed method is very competitive with the state-of-the-art methods. It achieves an average accuracy of 0.9612, 0.9614, 0.6761, sensitivity of 0.7814, 0.7234, 0.9702, and specificity of 0.9788, 0.9799, 0.9702 on 3 public databases DRIVE, STARE, and CHASE_DB1 respectively. The proposed method is promising for automated blood vessel segmentation.