Self-Calibrated Convolution U-Net For Retinal Vessel Segmentation

Retinal vessels in fundus images are the only deep microvessels that can be directly observed in the human body,it is an important method for doctors to diagnose many serious diseases by analyzing the length,diameter,branching mode,curvature and angle information of vessels,for example,in patients with cardiovascular disease,diabetes,and eye disease,the structure of the retinal vessels is often changed and there are disaffected areas around the vessels.Fundus images are easy to obtain and accurate segmentation of retinal vessels is the basic step for diagnosis and screening of these diseases.Artificial segmentation of fundus vessels is time-consuming and labors,and the segmentation effect depends on the subjective perception of the specialist.Therefore,more and more automatic segmentation technologies have been proposed to solve this problem,however,many methods are not enough in segmentation accuracy.Therefore,it is necessary to design a more accurate and efficient algorithm to realize the automatic segmentation of retinal vessels.In this paper,I propose a self-calibrated convolutional U-Net(SCCU-Net)to achieve accurate segmentation of retinal vessels.Convolution in SCCU-Net,using the calibration module instead of the traditional convolution operation,unlike traditional convolution,which uniformly performs all convolution on the input features in the original space,self-calibration convolution first transforms the input features into low-dimensional features and convolves them through subsampling,and then calibrates the convolution transform in the other part through the low-dimensional features after convolution transformation.Thanks to this heterogeneous convolution,the receptive field of each spatial location can be effectively expanded and richer output information can be obtained.In addition,an improved spatial attention module is proposed,which can obtain more complete spatial information by obtaining spatial attention weights before convolution operation.The proposed algorithm was evaluated on three publicly available datasets,DRIVE,STARE and CHASE DB1.The comparative experimental results show that SCCU-Net has a great improvement in segmentation results compared with U-NET.Compared with the current state of the art methods,the proposed method can also achieve better or comparable results in the segmentation of retinal vessels.