Design And Implementation Of Ultrasound Blood Vessel Image Segmentation And Recognition Software Based On Deep Learning

Vascular segmentation and recognition technique is of great clinical significance,which can assist medical personnel to quickly determine parameters such as blood vessel types,locations,shapes,so as to more accurately diagnose and treat diseases.The basis for realizing vascular segmentation and recognition is that the acquisition of blood vessel images,which depends on medical imaging equipment.Compared with current imaging equipment,ultrasound imaging has the advantages of high security,low price and easy operation,so this paper chooses ultrasound blood vessel images as the experimental object.Based on the developing trend of vascular segmentation and recognition technique,combined with the advantages of ultrasound imaging technique,this paper designs a software for ultrasound blood vessel image segmentation and recognition based on the combination methods of SSD(Single Shot MultiBox Detector)and Seeded Region Growing(SRG)and Fully Convolutional Networks(FCN),and research on the basic functional modules of the software and the algorithm models involved,which lays the foundation for the acquisition of vascular related parameters.Firstly,according to the MicrUs EXT-1H ultrasound module provided by a third party,combined with the requirements analysis,this paper proposes the overall design scheme of ultrasound blood vessel image segmentation and recognition software based on combination of SSD and SRG and FCN,and describes the basic functional modules and algorithm model design of the software in detail.Secondly,as the premise of blood vessel image processing is the acquisition of images,based on this requirement,MFC is used to create a dialog box program,and the basic functional modules such as display,state adjustment and parameter adjustment of the ultrasound image are realized by coding,so as to achieve the goal of blood vessel image acquisition.Thirdly,based on caffe,transfer learning was carried out on the officially provided SSD_Mobile Net model and FCN-8s model,and SRG algorithm was used to further segment the results processed by SSD_Mobile Net model.Then,the detection results of the SSD_Mobile Net was evaluated by the calculated values of mAP(mean Average Precision)and FPS(Frames Per Second)of 100 ultrasound blood vessel images,and the segmentation results of FCN-8s and SRG algorithm for 50 ultrasound blood vessel images with different features randomly selected was compared and analyzed by using the differential experiment method.Finally,in the OpenCV development environment,the SSD_Mobile Net model and the FCN-8s model are transplanted from Ubuntu system to Windows system across platforms,so as to complete the fusion of algorithm model and software basic function module and realize the design of blood vessel segmentation and recognition software.And according to the principle of software testing,writing test cases is to complete the test of software function and performance.The test results show that the segmentation and recognition algorithm proposed in this paper has high accuracy,and the designed software function behaves normally and practicably.Among them,the detection and recognition accuracy of the SSD_Mobile Net model is as high as 83.7957 and it has a detection speed of 25.2262 FPS,which means that the real-time goal can be achieved.In addition,the FCN-8s model has higher segmentation accuracy than SRG algorithm.So the SSD_Mobile Net model is chosen as a vascular detection and recognition algorithm,and the FCN-8s model is used as a vascular segmentation and recognition algorithm.The combination of them in software can better assist the doctors in selecting different algorithm functions according to actual needs.