Advanced Ultrasound in Diagnosis and Therapy ›› 2023, Vol. 7 ›› Issue (1): 16-22.doi: 10.37015/AUDT.2023.220023

• Original Research • Previous Articles     Next Articles

Predicting Malignancy in Sonographic Features of Thyroid Nodules Using Convolutional Neural Networks ResNet50 Model

Jingfang Dong, MDa,1, Jianyun Wang, MDb,1, Xiangzhu Wang, MDa,*()   

  1. a Department of Ultrasound, The Third Affiliated Hospital, Beijing University of Chinese Medicine, Beijing, China
    b Department of Gastroenterology, The Third Affiliated Hospital, Beijing University of Chinese Medicine, Beijing, China
  • Received:2022-06-24 Accepted:2022-08-16 Online:2023-03-30 Published:2023-03-30
  • Contact: Xiangzhu Wang, MD,
  • About author:First author contact:

    1Jingfang Dong and Jianyun Wang contributed equally to this study.


Objective To predict sonographic features of malignancy in thyroid nodules by using convolutional neural networks (CNNs) ResNet50 model.

Methods A cohort of 461 patients having sonographic thyroid nodules with histology diagnosis were randomly split into training set (70%), validation set (15%) and testing set (15%). Labeled sonographic patterns of thyroid nodules were used to train CNNs ResNet50 in training set, where algorithm pipelines were used to explicitly delineate structures of interest using segmentation algorithms to measure predefined characteristics of these structures as to be predictive and to use these features to train models that predict the malignancy in thyroid nodules. The prognostic accuracy of ResNet50 model was evaluated on validation set and compared with the individual sonographic specialists in testing set. Accuracy, sensitivity, specificity, and efficiency of ResNet50 model was measured using receiver operating characteristic (ROC) curve.

Results Measurements showed the evaluation indexes of ResNet50 model were as follows: accuracy: 94.39%, sensitivity: 92.45%, specificity: 96.30%, efficiency: 96.08%, F1 value: 94.23%, and AUC: 93.40%. The prognostic accuracy and other indexes of ResNet50 model was not subordinate compared to sonographic specialists (P < 0.05).

Conclusion These results highlight the emerging role of deep learning techniques including CNNs in precision medicine and suggest an expanding utility for computational analysis of sonographic images in the future practice. This study showed a computational approach can be used for learning sonographic features of thyroid nodules using ResNet50 model to combine the power of adaptive machine learning and algorithms with traditional sonographic assessment.

Key words: Thyroid nodules; Malignancy; Sonographic feature; Deep learning; Convolutional neural networks; ResNet50 model