The Value Of Deep Learning Based Convolutional Neural Network CT Images In The Differential Diagnosis Of Benign And Malignant Thyroid Nodules

Objective:To investigate the value of deep learning-based convolutional neural network CT images in the differential diagnosis of benign and malignant thyroid nodules.Five convolutional neural network（CNN）models were selected,and one ensemble model was generated to discriminate benign and malignant thyroid nodules,and the diagnostic performance of all models was compared with that of radiologists.Methods:We retrospectively included CT images of 880 patients with surgically pathologically confirmed thyroid nodules between July 2016 and December 2019,and collected a total of 986 thyroid nodules.Firstly,CT images of all patients were exported in DICOM format to the Darwin Intelligent Research Platform,where two radiologists manually map regions of interest（ROI）on thyroid nodules layer by layer without knowledge of the pathology results.These nodules were then randomly divided into a training-validation set and a test set.For the ROI of the training set,random horizontal flipping and random rotation were performed as image augmentation.Adjust all ROI images to a window width of 350 and a window level of 40,and scale to a size of224×224×3（Xception network of 299×299×3）,normalized to pixels between 0-1.5CNNs（Res Net50,Dense Net121,Dense Net169,SE-Res Ne Xt50,and Xception）were trained-validated and tested using 788 and 198 thyroid nodule CT images,respectively.All networks adopted the pre-trained models on Image Net.All models performed 5-fold cross-validation on the training-validation set.The maximum number of iterations in training was 50,the batch size was 4,the optimizer was Adam,the initial learning rate was 5e^-5,and the learning rate decayed to the 9th power of the number of iterations.Three models with better diagnostic performance on the test set were selected for model collection.Two radiologists retrospectively diagnosed benign and malignant thyroid nodules on CT images in a test set.The diagnostic performance of the five CNN models,the ensemble model and two radiologists for benign and malignant thyroid nodules was measured by the area under the receiver operating characteristic curve（AUC）,sensitivity,specificity,accuracy,positive predictive value（PPV）and negative predictive value（NPV）to measure the diagnostic performance of five CNN models,the ensemble model and two radiologists for benign and malignant thyroid nodules.Results:A total of 986 thyroid nodules were obtained,541 malignant and 445 benign,which were randomly divided into training-validation set（359 benign and 429malignant nodules）and test set（86 benign and 112 malignant nodules）.The mean size of the nodules,female-to-male ratio and age of patients were not significantly different between the benign and malignant thyroid nodules（P>0.05）,with the Res Net50,Dense Net121,and Dense Net169 models having higher AUC values and better diagnostic performance,the three models were integrated,and the ensemble model was obtained with an AUC of 0.901（95%confidence interval[CI]:0.906-0.974）,sensitivity of 0.872,specificity of 0.768,accuracy of 0.808,PPV of 0.740,and NPV of 0.878.The AUCs of the two radiologists for the diagnosis of benign and malignant thyroid nodules were 0.587（95%confidence interval[CI]:0.515-0.656）and 0.754（95%confidence interval[CI]:0.688-0.812）,with sensitivities of 0.593 and 0.802,specificities of 0.580and 0.705,accuracies of 0.586 and 0.748,PPV of 0.520 and 0.677,and NPV of 0.644and 0.705.the AUCs of the two radiologist and CNN models were significantly different（p<0.05）.The ensemble model had the highest AUC values.Conclusion:5 CNN models and ensemble model performed better than radiologists in distinguishing malignant thyroid nodules from benign nodules on CT.Compared with the single model,the diagnostic performance of the ensemble model was improved and showed good potential.