Research On Extrathyroidal Extension Of Thyroid Cancer By Deep Convolution Neural Network Model Based On Ultrasound Images

Objective To construct and verify a deep convolution neural network model based on ultrasound images to evaluate extrathyroidal extension in patients with thyroid cancer.Methods A retrospective study was conducted on 483 patients with thyroid cancer who underwent preoperative ultrasound examination in Zhejiang Cancer Hospital from January 2019 to July 2019.All patients were randomly divided into three groups:training queue (n=338),verification queue (n=72) and test queue (n=73).Three neural network model algorithms,ResNet-50,Inception-V3 and DenseNet-121,were used to evaluate the diagnostic performance of extra-glandular invasion according to the area under the working characteristic curve(Area Under Curve,AUC),accuracy,sensitivity,specificity,positive predictive value and negative predictive value.Finally,we compared the optimal DCNN model algorithm with the performance of three experienced ultrasound doctors in the verification cohort.Results In the test queue,the accuracy,sensitivity,specificity,positive predictive value,negative predictive value and the AUC value of the best deep convolution neural network model ResNet-50 in the diagnosis of extraglandular invasion of thyroid cancer are 82.1%,80.2%,83.6%,79.3%,84.4% and 0.870,respectively.Besides,the accuracy,sensitivity,specificity,positive predictive value,negative predictive value and the AUC value of ultrasound doctors are 74.0%,71.4%,76.3%,73.5%,74.4% and 0.739,respectively.Conclusion The depth convolution neural network model based on ultrasound images is a non-invasive,reliable and accurate tool for predicting ETE in patients with thyroid cancer.