Deep Convolutional Neural Networks for Thyroid Tumor Grading using Ultrasound B-mode Images

Abstract

The performances of deep convolutional neural network (DCNN) modeling and transfer learning (TF) for thyroid tumor grading using ultrasound imaging were evaluated. This retrospective study included input patient data (ultrasound B-mode image sets) assigned to the training group (115 participants) or testing group (28 participants). DCNN (ResNet50) and TF (ResNet50, ResNet101, ResNet152, VGG16, Inception V3, and DenseNet201), which trains a convolutional neural network that has been pre-trained on ImageNet, were used for image classification based on thyroid tumor grade. Supervised training was performed by using the DCNN or TF model to minimize the difference between the output data and clinical grading. The performances of the DCNN and TF models were assessed in the testing dataset with receiver operating characteristic analyses. Results showed that TF based on Resnet50 and VGG16 had better performance than DCNN (ResNet50) in differentiating thyroid tumor with areas under the receiver operating characteristic (AUCs) curve more than 0.8. However, TF based on ResNet101, ResNet152, InceptionV3, and Densenet201 had equal or worse performances than DCNN (ResNet50) in grading thyroid tumor with AUCs less than 0.5. TF based on ResNet50 and VGG16 had a superior performance compared to DCNN (ResNet50) model for grading thyroid tumors based on ultrasound images.