MDANet: Multi-Direction Attention Network for Curvilinear Structure Segmentation of Biomedical Images

Abstract

Automated segmentation of curvilinear structures is a crucial step for the computer-aided diagnosis of many diseases. Recently, deep learning-based segmentation models have achieved state-of-the-art performance for curvilinear structures. However, most deep learning-based models ignore the complex curvilinear structure, leading to foreground objects being misclassified as background noises. To this end, we propose a Multi-Direction Attention (MDA) mechanism to learn adequate multi-direction spatial attention and channel attention information. To be specific, to segment the curve structure more effectively, we use four strip convolutions in different directions (horizontal, vertical, left diagonal, and right diagonal) rather than the commonly used regular square convolution. Furthermore, we propose a supervised feature aggregation module (SFAM), which can effectively consider the multi-scale feature information to boost the segmentation performance. Integrating MDA and SFAM, we design a novel deep model, MDANet, for curvilinear structure segmentation. Extensive experiments are conducted on three public datasets, including DRIVE, CHASE-DBI and EM datasets. Experimental results demonstrate that MDA outperforms popular attention modules, and MDANet achieves comparable or superior segmentation performance.