Ultrasound Image Segmentation Method Of Cervical Lymph Nodes Based On Deep Learning

Cervical lymph nodes as an important immune organ,lymph node disease is a prone area.Therefore,the examination of cervical lymph nodes is of great significance in clinical diagnosis.Ultrasound imaging technology has become the first choice for cervical lymph node examination due to its advantages of simple,real-time and economical imaging process.However,due to the defects caused by reverberation artifacts and ultrasonic speckle noise in ultrasonic imaging technology,medical ultrasonic images are blurred and subjective in the manual detection of diseases.Therefore,it has become an urgent problem to effectively process the ultrasound images of cervical lymph nodes,obtain the regions of interest through automatic segmentation,and conduct relevant analysis,so as to provide objective reference indicators for doctors’ diagnosis.For the segmentation of cervical lymph node ultrasound images,most of the current algorithms will cause certain false edges.To solve this problem,this thesis proposes a method of using Masked Area Convolutional Neural Network(Mask R-CNN)structure to achieve segmentation of cervical lymph node ultrasound images.In order to further improve the accuracy of lymph node segmentation in ultrasound images,the Mask RCNN algorithm has been improved accordingly.First,in order to solve the poor training effect of small sample data sets,this paper introduces the method of transfer learning to train the network.Second,in order to extract more accurate target information and make better segmentation of target information,a convolution block attention mechanism module is introduced.The convolution block attention mechanism module performs attention supervision on the deepest features.And the importance of learning different pixels on the and channels,thereby improving the accuracy of the segmentation effect of lymph node ultrasound images.The thesis used 209 cervical lymph node ultrasound images to conduct experiments.Among them,167 images were selected as the training set data in the experiment,and the remaining 42 cervical lymph node images were used as the test set.The experimental results show that the qualitative and quantitative results of the Mask R-CNN method with attention mechanism are superior to the U-Net network.Compared with the original Mask R-CNN,the improved Mask R-CNN method improves the segmentation effect.The Mask R-CNN algorithm with the attention mechanism module added.The Dice coefficient reached 0.95,the volume overlap error was-0.02461,and the relative volume difference,accuracy,and sensitivity were 2.431%,0.9562,and 0.933,which demonstrated the effectiveness of the proposed method in this thesis.