Abstract: Objective To evaluate the diagnostic accuracy and clinical utility of a Vision-LSTM-based artificial intelligence(AI)model in classifying Thyroid Imaging Reporting and Data System Category 4b(TI-RADS 4b)thyroid nodules on ultrasound. Methods This study utilized ultrasound imaging data from 401 TI-RADS 4b thyroid nodules. A Vision-LSTM model was developed and validated. The AIs diagnostic performance was compared against that of junior and senior physicians using key metrics, including the area under the curve(AUC)and the precision-recall curve(PRC). Results On an independent validation set, the Vision-LSTM model achieved an AUC of 0.88 and an accuracy of 89.4%, significantly outperforming junior physicians(AUC: 0.624)and performing on par with senior physicians(AUC: 0.787). The model demonstrated a high capability for identifying complex sonographic features and delivering consistent diagnostic outcomes.The AI model was able to accurately identify the complex features in ultrasound images, and consistently produced consistent diagnostic results, demonstrating a high degree of accuracy and reliability. Conclusion The Vision-LSTM-based AI model significantly improves the diagnostic efficiency and accuracy for TI-RADS 4b thyroid nodules, showing great potential as an effective tool to support clinical decision-making and reduce physician workload.

Key words: Thyroid Imaging Reporting and Data System, Thyroid nodules, Vision-LSTM model, Diagnostic accuracy, Artificial intelligence