Automatic venous segmentation in venipuncture robot using deep learning

Vein identification plays a pivotal role in realizing automatic venipuncture, and it has become a difficulty to segment the veins efficiently as well as accurately in the research of full-automatic venipuncture robots. Most studies in the field of vein segmentation have only focused on traditional image processing methods, the segmentation accuracy and generalization performance of which are poor. Therefore, we propose an automatic image segmentation algorithm using the U-Net model with the attention mechanism (Attention-UNet) which can suppress unnecessary features. Besides, the encoder-decoder and the skip-connection structure are applied for multi-scale feature recognition so that the segmentation accuracy can be improved. Meanwhile, on digital arm images for the vein segmentation data set (DAIVS data set), the newly-built human forearm veins data set, the effectiveness of the proposed method in vein segmentation is verified. Finally, we conduct experiments to acquire and process venous images with the Attention-UNet in real-time on the venipuncture robot. These results indicate that machine vision has better performance in complex visual tasks and can be translated into clinical application.