Related Experiment Video
Updated: May 1, 2026
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Identification of Novel Biomarkers for Malignant Thyroid Nodules: A Preliminary Study Based on Ultrasound Omics
Zufei Li1, Kaifeng Wang2, Junpu Qu3
1Department of Otorhinolaryngology, Head and Neck Surgery, Beijing Chaoyang Hospital, Capital Medical University, Beijing, 100020, China.
Background And Objective:
The identification of thyroid nodules primarily relies on the ultrasound physician's assessment of nodule morphology and other visually identifiable features. Ultrasound omics technology can reveal additional features that are not visible to the naked eye, which may assist in the evaluation of malignant thyroid nodules. This study aims to explore novel markers for malignant thyroid nodules using ultrasound omics and machine learning (ML) techniques.
Methods:
A total of 1056 thyroid nodules with confirmed pathology were included, comprising 469 malignant and 587 benign cases. Traditional ultrasound features, such as "aspect ratio," "shape," "margins," "blood flow signal," and "calcification pattern," were recorded. Regions of interest (ROIs) were drawn for each ultrasound image, and features were extracted using the Python-based pyRadiomics tool. The Least Absolute Shrinkage and Selection Operator (Lasso) algorithm and correlation analysis were applied to select relevant features. Data were divided into training and testing sets at an 80:20 ratio. Various ML algorithms were employed to construct the models, and SHapley Additive exPlanations (SHAP) was used to assess feature importance.
Results:
A total of 104 ultrasonic omics features were extracted from each image, and seven ultrasonic omics markers for thyroid malignant nodules were identified. The model developed using the random forest (RF) algorithm performed best on the test set, achieving accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUC) values of 89.6%, 90.2%, 89.2%, and 89.7%, respectively. However, when the seven ultrasonic omics markers were excluded from the ML features, the model performance decreased to 83.5%, 80.4%, 85.8%, and 83.1%. SHAP analysis indicated that all seven markers were significant features.
Conclusion:
These novel ultrasonic omics markers may improve the accuracy of thyroid nodule diagnosis, and further research is needed to confirm their clinical utility.
