Comparative evaluation of deep learning models for thyroid nodule classification in C-TIRADS-based ultrasound imaging

Abstract

This study aimed to develop an intelligent scoring model for thyroid ultrasound imaging based on the Chinese Thyroid Imaging Reporting and Data System (C-TIRADS) and to evaluate its performance in differentiating benign and malignant nodules. A total of 6,596 ultrasound images were collected, and five deep learning models were tested for feature classification. DenseNet121 achieved the most stable performance and was selected as the final classifier. Comparative analysis showed that the AI system outperformed a junior radiologist, with higher accuracy (84.0% vs. 62.0%), sensitivity (72.5% vs. 41.2%), specificity (95.9% vs. 83.7%), and F1 score (82.2% vs. 52.5%). It also surpassed a senior radiologist in accuracy (84.0% vs. 79.0%), specificity (95.9% vs. 87.8%), and F1 score (82.2% vs. 77.4%), while maintaining comparable sensitivity (72.5% vs. 70.6%). These findings demonstrate that a C-TIRADS-based AI framework can provide accurate and objective risk stratification of thyroid nodules, outperforming less-experienced clinicians and approaching the performance of senior experts. However, the study was limited by data from a single institution, potential dataset imbalance, and variability in image quality and equipment. Future multicenter studies are needed to validate the model’s robustness and generalizability.

Ähnliche Arbeiten