The segmentation effect is not very good due to some problems such as inconsistent thickness of blood vessels and poor contrast of blood vessel boundaries. In order to solve the problem of small blood vessels breaking during network segmentation caused by the above reasons, this paper takes U-Net as the network basis, and firstly adds CBAM module in the process of network coding. Secondly, PPM module is introduced into the model to fuse the vascular features of different scales, so as to obtain more rich contextual information for the original. In addition, CDCM module is added at the bottom of the model to enhance the receptive field of the model, so that the network has more spatial scale information, and strengthen the awareness of context features, so that the segmentation task has better performance. The experiments on DRIVE, CHASE_DB1 and self-made dataset100 show that compared with the original U-Net, SE increases by 9.22%, SP by 0.83%, Acc by 1.65%, AUC by 1.56% and F1 by 5.93%. Compared with other methods, the sensitivity, accuracy and other indicators are improved, indicating that the vascular segmentation method in this paper has the ability to capture complex features and has higher advantages.
|