Fundus Image Based Retinal Vessel Segmentation Utilizing a Fast and Accurate Fully Convolutional Network

Monitoring the condition of retinal vascular network based on a fundus image plays a vital role in the diagnosis of certain ophthalmologic and cardiovascular diseases, for which a prerequisite is to segment out the retinal vessels. The relatively low contrast of retinal vessels and the presence of various types of lesions such as hemorrhages and exudate nevertheless make this task challenging. In this paper, we proposed and validated a novel retinal vessel segmentation method utilizing Separable Spatial and Channel Flow and Densely Adjacent Vessel Prediction to capture maximum spatial correlations between vessels. Image pre-processing was conducted to enhance the retinal vessel contrast. Geometric transformations and overlapped patches were used at both training and prediction stages to effectively utilize the information learned at the training stage and refine the segmentation. Publicly available datasets including DRIVE and CHASE DB1 were used to evaluate the proposed approach both quantitatively and qualitatively. The proposed method was found to exhibit superior performance, with the average areas under the ROC curve being 0.9826 and 0.9865 and the average accuracies being 0.9579 and 0.9664 for the aforementioned two datasets, which outperforms existing state-of-the-art results.