Parathyroid Gland Recognition Based On Convolution Neural Network

In recent years,with the rapid development of medical information,a variety of medical image data are becoming more and more abundant,and the medical image object detection is developing continuously.For the scene of endoscopic thyroidectomy,parathyroid injury can lead to hypocalcemia for life after thyroid surgery.Doctors need to protect parathyroid gland from injury during endoscopic surgery.However,parathyroid glands are very small glands,and congestion and obstruction may occur during endoscopic thyroidectomy,which makes it difficult for experienced surgeons to identify them.Therefore,the object detection methods in computer vision can help surgeons to successfully complete endoscopic thyroidectomy without damaging parathyroid gland.The paper is aim to identify and locate the parathyroid gland through the object detection methods based on convolutional neural network,so that surgeons can protect the parathyroid gland from damage during the operation.The main work of the paper is to identify and locate parathyroid gland and system implementation from the practical application effect.(1)The parathyroid video data provided by the cooperative hospital were processed.Firstly,an algorithm based on difference method is proposed to extract the key frames of video effectively to form parathyroid image data set;then,the images in the data set are labeled and sorted to form a standard format that can meet the model input;finally,because the image data collected during endoscopic thyroidectomy has the characteristics of unstable signal,low color discrimination,and is also affected by many environmental factors such as fog,angle of view,distance,occlusion,illumination,etc.,it is necessary to enhance the image data.(2)Considering the physical characteristics of parathyroid gland,this paper replaced the traditional rectangle bounding box with the circular bounding box,and proposed a full convolution one-stage pixel-by-pixel regression model for circular objects(FCOSC),which is suitable for the scene of endoscopic thyroidectomy.In order to calculate the parameters of the circular label box,this paper improve the IOU loss function and propose the center-ness-d branch.The average accuracy of the model in parathyroid data set experiment was 37.9%.In addition,this method performs better than the common models which can achieve an average accuracy of 70.4%on the open BCCD cell data set.(3)In order to show the experimental results in detail,the paper have made a specific system implementation and developed a parathyroid data analysis and research platform.This paper transform the python model into ONNX model and develop a website through onnxjs.In this site,the main function is to display the results of object detection model.Online resources or local upload resources can be selected for reasoning detection,and then the corresponding images and visual detection results will be displayed.