MAU-Net: A Retinal Vessels Segmentation Method

Abstract

Detailed extraction of retinal vessel morphology is of great significance in many clinical applications. In this paper, we propose a retinal image segmentation method, called MAU-Net, which is based on the U-net structure and takes advantages of both modulated deformable convolution and dual attention modules to realize vessels segmentation. Specifically, based on the classic U-shaped architecture, our network introduces the Modulated Deformable Convolutional (MDC) block as encoding and decoding unit to model vessels with various shapes and deformations. In addition, in order to obtain better feature presentations, we aggregate the outputs of dual attention modules: the position attention module (PAM) and channel attention module (CAM). On three publicly available datasets: DRIVE, STARE and CHASEDB1, we have achieved superior performance to other algorithms. Quantitative and qualitative experimental results show that our MAU-Net can effectively and accurately accomplish the retinal vessels segmentation task.