Application Of Deep Learning-based Parathyroid Recognition And Real-time Tracking(PTAIR) In Endoscopic Thyroid Surgery

ObjectiveThere is a high possibility of parathyroid injury or misresection during thyroidectomy.Real-time recognition can protect the parathyroid gland more expediently and timely.In this study,our purpose is to find the best algorithm and develop an artificial intelligence model to recognize parathyroid gland in endoscopic thyroid surgery,PTAIR,analyze its recognition efficiency and compare it with junior and senior surgeons,apply it in actual combat to analyze false positives and false negatives,and put forward professional improvement suggestions for the algorithm.MethodsA total of 166 cases of complete high-definition surgical video of endoscopic thyroid surgery were selected from the surgery done in the same group which was treated by Zhao Wenxin in the Department of Thyroid Surgery of the Union Hospital Affiliated to Fujian Medical University from August 24,2012 to August 24,2019.A total of 249 video clips with possible parathyroid glands were intercut by a junior surgeon.The senior surgeon divided them into parathyroid confirmed group(above 90% assurance),parathyroid unconfirmed group(50% assurance group),and non-parathyroid confirmed group(under10% assurance).parathyroid video clips(above 90% assurance)were converted into images.All images were randomly divided into training set and test set in a ratio of 15:2.YOLO,Faster RCNN and Cascade were used for training and the hyperparameter optimization was carried out.The best algorithm was selected according to the F value.The precision,recall rate,F value,accuracy,sensitivity,specificity,positive predictive value and negative predictive value of image recognition were used to evaluate the ability of the model to identify parathyroid gland under endoscopy.In addition,a retrospective study was conducted on a total of 20 cases of endoscopic thyroid surgery performed in the same group which was treated by Zhao Wenxin in the Department of Thyroid Surgery of the Union Hospital Affiliated to Fujian Medical University from December 20,2019 to March 25,2020.The 20 complete videos act as an independent external validation queue,allowing the AI model to run in the complete video.The recognition rate,initial recognition time and recognition duration of the parathyroid gland in the complete surgical video were compared among the artificial intelligence model,the junior and senior surgeons,the false positives and false negatives of parathyroid gland recognition by artificial intelligence model were analyzed.Results(1)Faster RCNN algorithm has the highest F value in the test data set.As the best choice at present,it is named as PTAIR-Faster RCNN-01 for further research.In the test data set,the precision,recall rate,F value,accuracy,sensitivity,specificity,positive predictive value and negative predictive value of PTAIR-Faster RCNN-01 for endoscopic parathyroid identification were 88.7%,92.3%,90.5%,89.86%,92.3%,87.6%,88.7% and92.5% respectively.(2)In the video where the junior surgeon identified a parathyroid gland and the senior surgeon confirmed that the junior surgeon was wrong(under 10%assurance),PTAIR-Faster RCNN-01 reduced this part of the junior surgeon’s misidentification,especially for tissue types such as thymus,tumor invasion capsule,thyroid nodule,and glomus/vein.(3)In the complete video analysis of the independent external validation cohort,the recognition rate of parathyroid was 96.88% in the AI model group,87.50% in the senior surgeon group,and 71.88% in the junior surgeon group.The recognition rate of artificial intelligence model was higher than that of junior surgeon and comparable to that of senior surgeon.While the artificial intelligence model found parathyroid glands earlier than the senior surgeon,and the sustained recognition time was longer than that of the senior surgeon.(4)The main misidentified tissue types of PTAIRFaster RCNN-01 were adipose tissue and lymph node.The misrecognition areas were mainly concentrated in the central region of lymph node dissection and process during establish the lacunae to the lobectomy.There are bleeding,smoke,occlusion and other situations in the missing identification.ConclusionBased on this Convolutional Neural Network(CNN)model,we successfully developed a computer aided endoscopic parathyroid intelligent recognition model—PTAIR-Faster RCNN-01.The model realizes the real-time recognition and tracking of parathyroid gland in endoscopic thyroid surgery video,and achieves a high precision.It can reduce the misidentification of junior surgeon.The identification rate of AI is higher than that of junior surgeon and comparable to that of senior surgeon.It has the function of early warning and continuous tracking.We analyze misidentified organization types,scenarios and misidentified situations,put forward targeted suggestions to reduce false positives and false negatives.