Thyroid Nodule and Lymph Node Metastasis Assessment from Ultrasound Images Using Deep Learning

Xiaohui Zhao¹Gnag Zhang²Xueqin Shen³Jin Dianshengin¹Wei Yanrong⁴Yu Zhang¹Xin Liu¹Yang Liu¹Dongfang Yang¹Huiying Xiao^1*Xianquan Shi^5*Xiaoguang Yang^1*

• ¹Hohhot First Hospital, Hohhot, China
• ²Beijing Institute of Technology, Beijing, China
• ³People's Liberation Army Strategic Support Force Information Engineering University - Luoyang Campus, Luoyang, China
• ⁴Guilin University of Electronic Technology, Guilin, China
• ⁵Capital Medical University Affiliated Beijing Friendship Hospital, Beijing, China

Objectives: The preoperative differentiation of thyroid nodules into benign thyroid nodules (BTN), non-metastatic malignant thyroid nodules (NMTN), and metastatic malignant thyroid nodules (MMTN) is critical for guiding clinical management strategies. Ultrasound (US) examinations frequently exhibit diagnostic inconsistencies due to operator-dependent variability. Computer-Assisted Diagnosis (CAD), an artificial intelligence (AI) model based on convolutional neural network (CNN), can help overcome inconsistencies in US examination outcomes by leveraging large-scale ultrasound imaging datasets to improve classification accuracy. Our study seeks to establish and validate this AI-powered ultrasound diagnostic model for precise preoperative discrimination among BTN, NMTN, and MMTN. Methods: A total of 209 patients (BTN=66; NMTN=15 and MMTN=128) were consecutively identified and enrolled from multi-center database. A subset of 195 patients (BTN=60; NMTN=15 and MMTN=120) was selected for final analysis. These patients were divided into two groups: a training set (BTN=50; NMTN=11 and MMTN=100) and a testing set (BTN=10; NMTN=4 and MMTN=20). A total of 3537 ultrasound images from the 195 patients were preprocessed by normalizing grayscale values and reducing noise. The processed images were then input into the AI model, which was trained to classify thyroid nodules. The model' s performance was evaluated using the testing set and assessed through receiver operating characteristic (ROC) curve analysis and confusion matrix. Finally, the diagnostic accuracy of the AI model was compared with that of radiologists to determine its clinical utility in ultrasound-based diagnosis. Results: Compared with junior and senior radiologists, the AI model achieved near-perfect AUC values of 0.97 (BTN), 0.99 (NMTN), and 0.96 (MMTN), significantly outperforming the senior radiologist's AUCs (0.88 for NMTN) and the junior radiologist's weaker discrimination. In addition, the accuracy of this model was higher compared with all ultrasound radiologists (95% vs. 73%, and 84% for the junior radiologist and senior radiologist respectively). Conclusions: The AI-based ultrasound imaging diagnostic model showed excellent performance in differentiating BTN, NMTN, and MMTN, supporting its value as a diagnostic tool for the clinical decision-making process.