Research On Vein Detection Algorithm In Ultrasound Image Based On Deep Learning

The first step in medical clinical intervention is usually to establish a safe,effective and convenient access to the vein.The success of traditional artificial blind puncture and ultrasound-guided artificial puncture depends heavily on the medical staff's puncture experience and the patient's vascular condition,so an intelligent venipuncture solution is eagerly needed.The subject of this project is the ultrasound guidance image of intelligent puncture system.By intelligently detecting,locating and tracking the puncture target vessel in the image,the system can better guide puncture.Firstly,by comparing the detection effect of the vein detector composed of different artificial features and machine learning algorithms,cascaded algorithm with Haar feature was used to detect blood vessels.The recall of this model was 0.90.Then the author tried to use the detection algorithm based on deep learning for blood vessel detection.The acquisition of medical images is difficult,and the amount of training image usually is not enough for fully training a deep learning model.So the author firstly validated that finetuning the pre-trained ImageNet model is good for the vessels' recognition in the ultrasound image,and then Faster R-CNN(Regions with Convolutional Neural Network features)object detection algorithm was used to detect blood vessel.The alternating training of Faster R-CNN achieves transfer learning from the natural dataset to the venous ultrasound image dataset.Therefore,the vein detection training in the case of limited annotation samples can be completed,and the final AP value of the detector is 0.896.The proposed layer-by-layer filter strategy not only accurately screens the puncture target vessel in the ultrasound image,but also facilitates human-machine friendly interaction during puncture.In order to accurately locate the target blood vessel,two venous radius measurement methods are proposed.One method is to segment blood vessels based on adaptively selected threshold,and then scan the upper and lower edges of the blood vessel up and down.Another method is to segment the blood vessel based on the Otsu method,and use the morphological opening operation to preprocess the binary image,and then calculate the radius in a similar manner to the method 1.At last to correct the vein depth with the radius.Experiments were performed on the test image set,and the accuracy rates of methods 1 and 2 were 93.0% and 99.4%,respectively.The measurement based on Otsu and morphological is better.For the real-time tracking of puncture target vessel during puncture,kernelized correlation filter is used to track the position change of the puncture target in real time,so that the puncture robot can adjust the angle and distance of the needle in the puncture process.And it is the guarantee of successful puncture.