Construction Of Thyroid Ultrasound Image-based Convolutional Neural Network And Its Application In TI-RADS 4 Nodules

Part Ⅰ:Preliminary construction of thyroid ultrasound image-based convolutional neural network.Objective:To construct a classification model of thyroid ultrasound image based on convolutional neural networks(CNN),then optimize and verify its value in differential diagnosis of benign and malignant thyroid nodules.Methods:The original thyroid ultrasound images of patients undergoing thyroidectomy in our hospital from April 2019 to May 2020 were collected as the training data,and all images were divided into training set and test set according to 7:3,and the region of interest(ROI)of each lesion in the ultrasound image was analyzed after image preprocessing.Put it into the optimized CNN then output,and the loss function is minimized through the network optimizer to update the model parameters.Verify the test set.Results:The sensitivity,specificity,accuracy and AUC of the AI model are 88.07%,81.50%,83.26%and 0.9224,respectively.Conclusion:The CNN model has a high accuracy of classification and has a high diagnostic effect on the benign and malignant thyroid nodules.The AI model of thyroid ultrasound based on CNN in this study can be used as a risk prediction model of thyroid cancer,which can objectively and accurately estimate the risk of benign and malignant thyroid nodules.Part Ⅱ:The application of thyroid nodule AI in TI-RADS 4 nodulesObjective:To evaluate the type 4 nodules of Thyroid Imaging Reporting and Data System(TI-RADS)though the thyroid ultrasound AI model based on CNN.To analysis the clinical value of TI-RADS in the diagnosis of the type 4 nodules.Methods:Selected patients with thyroid nodules from June 2020 to December 2020,which diagnosed by thyroid ultrasound in our hospital,with the nodules classified by TI-RADS into 4 categories as the research objects.According to C-TIRADS,divided thyroid nodules into TI-RADS 4a nodules,TI-RADS 4b nodules and TI-RADS 4c nodules.4A nodules was defined as negative group,4b nodules and 4c nodules as positive group.A total of 161 TI-RADS 4 nodules were included in this study,which were divided into malignant group(106 cases,65.8%)and benign group(55 cases,34.2%).The ROC curve of thyroid ultrasound AI model to evaluate of benign and malignant TI-RADS 4 nodules was drawn,and calculated the area under the curve,obtained the sensitivity,specificity and accuracy.Compared and analyzed the diagnostic value of thyroid ultrasound AI model in benign and malignant thyroid nodules.Results:There were 29(18.0%)cases in TI-RADS 4a,35(21.7%)cases in TI-RADS 4b and 97(60.2%)cases in TI-RADS 4c.The diagnostic sensitivity,specificity and accuracy of the TI-RADS were 76.40%,93.81%and 76.40%,respectively.The sensitivity,specificity,accuracy and AUC of AI model were 90.57%,83.64%,88.19%and 0.766,respectively,which were higher than those of the TI-RADS.Conclusion:the thyroid ultrasound AI model based on CNN can detect thyroid cancer more accuracy,then help to choose the best treatment strategy.When the model is combined with TI-RADS classification,it is recommended to follow-up when TI-RADS is classified into 4 categories and AI is diagnosed as low-risk nodules;it is recommended to have a FNA or thyroidectomy or lobectomy when AI is diagnosed as a high-risk nodule.