Retinal Blood Vessels Semantic Segmentation Method Based on Modified U-Net
【摘要】：Automatic segmentation of retinal blood vessels from fundus images plays a key role in the computer aided diagnostic system, which is helpful for the early treatment of many fundus diseases including diabetic retinopathy, glaucoma and hypertension. In this paper, a modified U-Net is proposed to train semantic segmentation models for retinal blood vessels. In addition, we use Condition Random Field to integrate the global information. The comparison between our method and other typical methods is given to evaluate the proposed method. Our experiments shows that our scheme leads to improved accuracy on DRIVE datasets and having obtained an average accuracy of 86.5% for retinal blood vessels segmentation task.