Recognition Of Brachial Plexus Ultrasound Im Ages Based On Object Detection

Peripheral nerve block（PNB）is widely used in clinical practice,but it is a challenge for doctors to locate the nerve accurately.With the continuous development of artificial intelligence,the application of computer vision in medical field has been paid more and more attention.Deep convolutional neural network helps computer to automatically learn features from images and complete automatic recognition and learning through complex function mapping.It has been widely used in natural images and medical fields,and has its shadow in diagnosis,prediction and treatment stages.In order to apply deep learning technology to peripheral nerve block anesthesia,this thesis studied the target detection model based on brachial plexus ultrasound images.In this thesis,an Iterative Detection Network（IDNet）model is constructed and a one-to-many Loop Feature Pyramid Network（OMLFPN）is designed.In this paper,the two-stage target detector Mask RCNN is taken as the prototype and combined with the characteristics of brachial plexus ultrasound images,a high-precision ultrasonic image multi-target detection network IDNet is built.In this paper,by studying the characteristics of feature pyramid,a lightweight feature pyramid network OMLFPN is designed by using the technique of reducing parameters,and a multi-target detection platform based on brachial plexus neural ultrasound image is realized by using it.The innovation of this thesislies in the design of two new modules and lightweight feature pyramid network.This paper first proposes two new modules.The Fully Connected Feature Pyramid Network（FCFPN）realizes the efficient fusion of high resolution Feature maps with texture information and low resolution Feature maps with semantic information,reduces the loss of key Feature information,and strengthens the divide-and-conquer strategy.In addition,the mean-iterative Region Proposal Network（MIRPN）is designed to improve the response of the Network model to edge features by iteratively predicting the suggestion box.This paper designs a lightweight feature pyramid network by studying the feature pyramid and combining the advantages of the previous network.This paper focuses on the design of network structure.By combining the brachial plexus ultrasound image and the characteristics of the target detector,some modules and networks were improved for the pain points of the previous detector,and a large number of experiments were conducted.In this thesis,a multi-target detection model and a lightweight feature pyramid network based on ultrasonic images are proposed.The module ablation experiment and network comparison experiment prove that the proposed module and network are higher in both accuracy and efficiency.It is more helpful for doctors to locate medical anatomical tissues and complete peripheral nerve block surgery when applied to practical scenarios.