Multi-scale channel importance sorting and spatial attention mechanism for retinal vessels segmentation

Retinal Vessels segmentation is an important procedure for detecting and diagnosing a variety of pathological diseases. However, the inherent complex properties around the disc make it challenging to improve the segmenting accuracy of capillaries and the retinal blood vessels at the ends. In this paper, we proposed a multi-scale channel importance sorting and important spatial information positioning (MSCS) encoder-decoder for segmentation in Retinal Vessels. Firstly, the fully convolutional encoder-decoder is formed to implement a series of linear and non-linear transformation and achieve end-to-end segmentation tasks. Then, the channel importance sorting module is employed to suppress useless feature responses during the process of encoding and to identify effective channels, whose information is utilized to recognize capillaries and the retinal vessels at the ends. Finally, in the decoding stage, the spatial attention mechanism module is designed to extract the positioning information of multi-scale feature maps. The spatial information of retinal vessels is collected to better locate the position of the vessels. In addition, aiming at taking fully advantage of the network, the multi-scale asymmetric cascade convolution module is proposed to reduce the parameters of the model and increase the operation rate. Experimental results on DRIVE, STARE datasets indicate that the proposed method outperforms other state-of-the-art strategies. This system, as demonstrated, can greatly decrease false positive rate of the blood vessels at the ends and enhance the sharpness of retinal vessels. (C) 2020 Elsevier B.V. All rights reserved.