Modeling Study Of Multimodal Ultrasound In Preoperative Diagnosis Of ACR TI-RADS 4、5 Thyroid Nodules

BackgroundUltrasonography has become the first choice for thyroid nodules because of its convenience and real-time performance.On this basis,the Thyroid Imaging Reporting and Data System(TI-RADS)launched by the American College of Radiology(ACR)provides standardized management for the diagnosis of thyroid nodules.Among them,thyroid nodules classified as TI-RADS 4 and TI-RADS 5 have complex biological behaviors and diverse ultrasonic image features,and the diagnosis often depends on the empirical judgment of sonographers.convolutional neural network(CNN)has great application prospects in the differential diagnosis of tumors.At present,there are few reports on the use of ultrasound multimodal image information to assist the diagnosis of TI-RADS class 4 and class 5 thyroid nodules.ObjectivesTo construct a multimodal DL fusion model of thyroid based on conventional ultrasound images combined with real-time elastography,and to explore its value in the differential diagnosis of benign and malignant thyroid nodules of TI-RADS 4 and 5 that are difficult to differentiate before operation.The diagnostic ability of the model was compared with that of sonographers with different experience,and whether the model could provide good auxiliary value for sonographers.MethodsThe clinical data of patients who underwent thyroid nodule resection in our hospital from January 2019 to May 2021 were retrospectively analyzed.The nodules classified as TI-RADS 4 and 5 by preoperative ultrasound were screened out.The two-dimensional ultrasound images and real-time elastography were collected and used as training data,and all images were divided into 8:2 is divided into training set and test set.Each ultrasound image was labeled as benign or malignant,pre-processed and input into the designed thyroid multimodal CNN model to predict the output.The diagnostic efficacy of the trained model was compared with that of different seniority sonographers in the differentiation of benign and malignant thyroid nodules to explore whether the model could improve the diagnostic accuracy of benign and malignant thyroid nodules by different seniority sonographers.The accuracy(ACC),sensitivity(SEN),specificity(SPE),and F1 score of the model were obtained by evaluating the performance of the model with surgical pathological results as the gold standard.The receiver operating characteristic(ROC)curve of the multimodal thyroid CNN model for evaluating the benign and malignant thyroid nodules in TI-RADS 4 and 5 was drawn.the area under the curve(AUC)was calculated to analyze the value of the model in differentiating benign and malignant thyroid nodules,and the diagnostic ability of the model was compared with that of sonographers with different seniority.Results1、Diagnostic efficacy of thyroid multimodal CNN fusion modelThe AUC of thyroid multimodal ultrasound CNN fusion model training set for benign and malignant diagnosis of TI-RADS 4 and 5 nodules was 0.940(95%CI:0.926,0.968),and the model had high classification accuracy.2、The efficacy of thyroid multimodal CNN fusion model in clinical diagnosisThe AUC,SEN and SPE were 0.867,0.840 and 0.890 in the senior sonographer and 0.715,0.721 and 0.708 in the junior sonographer,respectively.The AUC of thyroid multimodal ultrasound CNN fusion model validation set was 0.978(95%CI:0.942,0.999),ACC was 0.918,SEN was 0.980,SPE was 0.750,F1-Score was 0.980.The diagnostic performance and sensitivity of the model were significantly different from those of senior and junior sonographers(P<0.05).ConclusionsThe multi-modal CNN fusion model of thyroid constructed in this study has high classification accuracy for the diagnosis of TI-RADS 4 and TI-RADS 5 nodules that are difficult to distinguish clinically,which has great clinical application value.The model is expected to help sonographers with different seniority,especially junior sonographers to improve the diagnostic accuracy.