HRNet:A hierarchical recurrent convolution neural network for retinal vessel segmentation

The extraction of retinal vessel is of great importance in the diagnosis of fundus disease. Many approaches have been proposed for vessel segmentation. However, these models have some drawbacks. First, the encoder-decoder structures, U-Net i.e., will generate redundant information during successive convolution and sampling operations. Second, most methods only have feedforward process, and the feedback is also crucial for contextual feature representations from high to low layers. In this article, we overcome these limitations by proposing a hierarchical recurrent convolution neural network (HRNet). The proposed HRNet first integrates the advantage of the ResNet and Squeeze and Excitation (SE) to build SE-residual block in multi-scale layers, which capture the important channel-wise information and remove the redundant feature in deep network. Further, we design a hierarchical recurrent(feedback) mechanism to explore features from different upper to the lower layer by adding the output of each layer to its corresponding encoding layer iteratively. The feedback path encourages the feature reuse to improve the power of weak retinal vessel detection. Comprehensive experiments on three public retinal datasets (DRIVE, STARE and CHASE) demonstrate that the proposed HRNet is superior or equivalent to the state-of-art methods in terms of most of the indicators, including accuracy, F1-Score, sensitivity.

Automated summary

In this article, a hierarchical recurrent convolution neural network (HRNet) is proposed to improve the detection of weak retinal vessels. The HRNet integrates the advantages of ResNet and Squeeze and Excitation (SE), and employs a hierarchical recurrent mechanism to explore features from different layers.