Retinal vessels segmentation of colour fundus images using two stages cascades convolutional neural networks

Retinal vascular status in fundus images is a reliable biomarker for diabetes, hypertension and many ophthalmic diseases. Therefore, accurate segmentation of retinal vessels is of great significance for the diagnosis of many diseases. However, due to the inherent complexity of the retina itself and the lack of data, it is difficult to obtain the ideal accuracy of the segmentation results of the vascular end. To solve this problem, we propose an innovative two stage cascades convolutional Neural network (TCCNNet) to segment the retinal vessels in fundus images. In stage 1, an entire down sampled fundus images is sent to a probability-anatomical-prior-guided shallow-layer-enhanced location net to perform rough segmentation. In addition, a novel circular inference module and parameter Dice loss are proposed to reduce the uncertain probabilities and false positive points of the boundary. In stage 2, high-resolution 2D image slices are reconstructed to offset the detailed information lost during downsampling, further refining the contours. Moreover, a multi-view 2.5D net composed of three 2D refinement subnetworks is applied to deeply explore the morphological features, compensating for the mistakes and missing spatial information of a single view. The proposed method is evaluated on different evaluation metrics such as sensitivity, specificity and accuracy. The experimental results show that this method shows the most competitive performance compared to other state of the art methods.