Related Experiment Video
Updated: Jun 3, 2025
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Neural Network and Logistic Regression Models Based on Ultrasound Radiomics and Clinical-Pathological Features to
Jia-Wei Feng1, Shui-Qing Liu2, Yu-Xin Yang3
1Department of General Surgery, Sir Run Run Shaw Hospital, School of Medicine, Zhejiang University, Hangzhou, Zhejiang, China (J-W.F., H.L.); Department of thyroid surgery, The Third Affiliated Hospital of Soochow University, Changzhou First People's Hospital, Changzhou, Jiangsu, China (J-W.F., Y-X.Y., J.Y., Y.J.).
Rationale And Objectives:
Papillary thyroid carcinoma (PTC) often metastasizes to lateral cervical lymph nodes, especially in level II. This study aims to develop predictive models to identify level II lymph node metastasis (LNM), guiding selective neck dissection (SND) to minimize unnecessary surgery and morbidity in low-risk patients.
Methods:
A retrospective cohort of 313 PTC patients who underwent modified radical neck dissection (MRND) between October 2020 and January 2023 was analyzed. The patients were randomly assigned to a training cohort (70%) and a validation cohort (30%). Five predictive models were developed using neural networks (NNET) and logistic regression (LR) based on ultrasound radiomic features, clinical-pathological data, or a combination of both. Each model's performance was evaluated based on accuracy, area under the receiver operating characteristic curve (AUC), sensitivity, and specificity in predicting occult level II LNM. SHapley Additive exPlanations and nomogram were used to interpret the most important features in the models.
Results:
The occurrence rate of level II LNM was 28% in the cohort. Among the five predictive models developed, the LR-radiomics signature model demonstrated the highest performance, achieving an accuracy of 96.8% and an AUC of 0.989 in the validation set. In comparison, the NNET-radiomic + clinical feature model achieved an AUC of 0.935, while other models exhibited moderate to low accuracy and AUCs ranging from 0.699 to 0.785. The decision curve analysis demonstrated that the LR-radiomics signature model provided the greatest clinical utility, offering the highest net benefit across a range of decision thresholds for identifying occult level II LNM.
Conclusion:
Our study developed predictive models using ultrasound-derived radiomic features and clinical-pathological data to assess the risk of occult level II LNM in PTC. The LR-radiomics signature model demonstrated high accuracy, making it a valuable tool for guiding personalized treatment decisions, by informing MRND for high-risk patients and supporting SND for low-risk patients to minimize unnecessary surgical interventions and optimize clinical outcomes.
