Related Experiment Video
Updated: Jul 15, 2026
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Enhancing diagnostic accuracy of American College of Radiology TI-RADS 4 nodules: nomogram models based on MRI
Bin Song1,2, Qiaohui Chen2, Hao Wang2
1Shanghai Institute of Medical Imaging, Department of Radiology, Zhongshan Hospital, Fudan University, Shanghai, China.
Background:
Thyroid nodules classified as American College of Radiology Thyroid Imaging Reporting and Data System category 4 (ACR-TR4) present a diagnostic challenge due to their undetermined nature. This study aimed to develop and validate nomogram models using magnetic resonance imaging (MRI) morphological features to enhance the diagnostic accuracy of ACR-TR4 thyroid nodules, thereby reducing unnecessary fine-needle aspiration (FNA) and minimizing missed cancers.
Methods:
We retrospectively analyzed 229 ACR-TR4 nodules from 184 patients who underwent preoperative MRI and surgical thyroidectomy between January 2017 and December 2022 in Minhang Hospital, Fudan University. All nodules were pathologically confirmed and randomly divided into training (n=166) and validation (n=63) cohorts. We recorded MRI morphological features of the nodules, performed logistic regression analysis to identify independent predictors of malignancy, and developed a nomogram and improved models. The performance of the nomogram was assessed for discrimination, calibration, and clinical utility. The diagnostic performance of the improved models was compared with that of the ACR-TR4.
Results:
Among the 229 ACR-TR4 thyroid nodules, there were 140 benign and 89 malignant nodules, with 46 males and 183 females, and a mean age of 51.2±13.5 years. Diffusion restriction and reversed halo sign in the delayed phase were identified as independent predictors of malignancy and included in the nomogram. The nomogram showed robust discrimination and calibration in distinguishing malignant and benign ACR-TR4 nodules in both the training and validation cohorts, with areas under the curve (AUC) of 0.928 [95% confidence interval (CI): 0.887-0.970] and 0.904 (95% CI: 0.825-0.984), respectively. Four improved models were constructed using the two independent predictors either individually or collectively (OR or AND). The unnecessary FNA (21.1%, 11.7%, 5%, and 23.4%, respectively) and missed cancer rates (12.9%, 13.8%, 18.9%, and 5.7, respectively) were significantly lower than those of the ACR-TR4 system (64% and 43%, respectively).
Conclusions:
The nomogram model using MRI features such as restricted diffusion and reversed halo sign in the delayed phase improved the accuracy of diagnosing benign versus malignant ACR-TR4 thyroid nodules, potentially reducing unnecessary FNA and minimizing missed cancers.
