A Deep Learning Model Based On Ultrasound Images In The Classification Of Thyroid Benign And Malignant Nodules And Prediction Of Cervical Lymph Nodule Metastasis In Thyroid Cancer

Part I: The deep learning model based on ultrasound images in theclassification of thyroid benign and malignant nodules Objective:The purpose of this study was to evaluate the diagnostic performance of deep learning(DL)models based on ultrasound images in the classification of thyroid benign and malignant nodules.Methods:Between January 2019 and December 2022,1330 patients with 3690 thyroid ultrasound images at the Second Affiliated Hospital of Nanchang University and Jingzhou Central Hospital of Yangtze University were retrospectively collected.2041 images was divided into training cohort,while 893 images,416 images and 344 images were split into test cohort,validation cohort A and validation cohort B,respectively.The postoperative pathological results were adopted as the golden standard and ultrasound images of the training cohort were labeled.Four DL models,namely Res Net-50,Vi T,VGG-16 and Goog Le Net models,were proposed to diagnose thyroid nodules by analyzing greyscale ultrasound images respectively.The performance of models was assessed by accuracy,sensitivity,specificity and areas under the receiver operating characteristic curve(AUC).The DL model was re-visualized to help radiologists understand its potential working mechanism.Results:1.The AUC of the Res Net-50 model were 0.900,0.920 and 0.902 in the test cohort,validation cohort A and B,respectively.In the test cohort,the diagnosis performance of the Res Net-50 was superior to that of Vi T,VGG-16 and Goog Le Net,which the accuracy was 90.93%,sensitivity was 89.74% and specificity was 90.16%.The Res Net-50 remained elegantly accuracy in the validation cohort A and B,with88.90% and 89.51%.2.The accuracy of the junior radiologist were 74.71%,81.82% and 67.28% in the test cohort,validation cohort A and B,respectively.For the validation cohort A,the junior radiologist achieved the best diagnostic performance with the sensitivity was 76.06% and specificity was 85.04%,which were inferior to the Res Net-50 model.3.The visualization of the DL model demonstrated that the model accurately distinguish between thyroid nodules and surrounding tissue.Furthermore,the model intelligently assigned larger weights to the solid component in the diagnosis of echogenic thyroid nodules containing solid and cystic components.Conclusions:1.The DL model based on ultrasound images can accurately and objectively differentiate benign and malignant thyroid nodules,which is worthy of application to clinical practice.2.The DL model based on ultrasound images have good diagnostic performance for classify thyroid nodules,which is superior to junior radiologist.Part II: Predictive value of cervical lymph node metastasis in the thyroid cancer by using deep learning model based on ultrasound imagesObjective:To establish a deep learning model based on ultrasound images for the prediction of cervical lymph node metastasis in the thyroid cancer.Methods:Between January 2019 and December 2022,338 patients with 676 thyroid ultrasound images at the Second Affiliated Hospital of Nanchang University and Jingzhou Central Hospital of Yangtze University were retrospectively collected.280 images was divided into training cohort,while 120 images,156 images and 120 images were split into test cohort,validation cohort A and validation cohort B,respectively.The postoperative pathological results were adopted as the golden standard and ultrasound images of the training cohort were labeled.In the test cohort,validation cohort A and B,the image collection and direct prediction were performedby three radiologists who had over 6,3 and 15 years of work experience in thyroid ultrasound,respectively.Res Net-18 model based on greyscale ultrasound images was proposed to predict cervical lymph node metastasis in the thyroid cancer.The performance of model was assessed by accuracy,sensitivity,specificity,positive predictive value and negative predictive value(NPV).Results:1.In the test cohort,validation cohort A and B,the AUC of the Res Net-18 model were 0.714,0.693 and 0.779,while the accuracy were 70.51%,68.33% and71.67%,respectively.The sensitivity and NPV of the model achieved good results.In the test cohort,validation cohort A and B,the sensitivity were 73.68%,76.67% and83.33%,while the NPV were 72.97%,72.00% and 78.26%,respectively.2.The accuracy of three radiologists were 73.33%,66.67% and 86.67% in the test cohort,validation cohort A and B,respectively.The accuracy of the Res Net-18 model was superior to the junior radiologist and inferior to experienced radiologists.Conclusions:1.The DL model based on ultrasound images achieved good performance in predicting cervical lymph node metastasis in thyroid cancer,which may have great potential for clinical practice.2.The diagnostic performance of the DL model in predicting cervical lymph node metastasis in thyroid cancer was superior to the junior radiologist and inferior to experienced radiologists.