MDA-Net: Multi-Dimensional Attention Network for Retinal Vessel Segmentation

Fundus images play an important role in the diagnosis and treatment of ophthalmic diseases (such as hypertension, arteriosclerosis and diabetic retinopathy), and the morphological in-for-mation of retinal blood vessels can be used as an important index for the diagnosis of these dis-eases, so it is very important for accurate segmentation of retinal blood vessels. With the con-tinuous development of computer technology, deep learning method provides a new idea for medical image segmentation. Due to the complex structure and different scales of retinal blood vessels, the existing U-Net model still faces significant challenges in dealing with these tasks. To solve these problems, we propose a retinal vascular segmentation network MDA-Net based on multidimensional attention mechanism. Based on the U-Net structure, this method op-timizes the network design by reducing the number of encoder-decoder layers to three layers, and intro-duces the Coordinate Grouped Feature Fusion (CGFF), multi-dimensional feature en-hancement (MDFE) and multi-scale convolution enhancement (MSCE). Firstly, CGFF module in-tegrates mul-ti-scale features by grouping convolution and multi-dimensional pooling, which improves the adaptability of the model to uneven distribution of blood vessels. Secondly, MDFE module com-bines channel shuffling, multi-dimensional attention and pooling operation to en-hance the ex-traction ability of micro-vessel features, especially in low contrast and complex back-ground.Experimental results show that the accuracy (ACC), sensitivity (SE) and specificity (SP) of this method DRIVE 0.9825, 0.9842 and 0.9895, 0.8211, 0.8342 and 0.8452 respectively, and 0.9840, 0.9872 and 0.992 respectively. MDA-Net proposed in this paper provides a new idea and scheme for improving the performance of retinal blood vessel segmentation.