Related Experiment Video
Updated: Sep 13, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.0K
Optimizing Thyroid Nodule Management With Artificial Intelligence: Multicenter Retrospective Study on Reducing
Jia-Hui Ni1, Yun-Yun Liu1, Chao Chen2
1Department of Medical Ultrasound, Shanghai Tenth People's Hospital, Tongji University School of Medicine, YanChang Middle Street 301, Shanghai, China, 86 21-66307539.
JMIR Medical Informatics
|July 30, 2025
Summary
Artificial intelligence (AI) effectively identifies benign thyroid nodules, significantly reducing unnecessary fine needle aspirations (FNA). While AI shows promise as a diagnostic tool, careful surveillance remains crucial for potentially missed malignant nodules.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Current artificial intelligence (AI) models for thyroid nodules primarily focus on malignancy screening but lack widespread clinical implementation.
- There is a need to evaluate AI's efficacy in real-world clinical settings for managing thyroid nodules.
Purpose of the Study:
- To assess the performance of an AI system in identifying potentially benign thyroid nodules that radiologists initially flagged as suspicious for malignancy.
- The goal is to reduce unnecessary fine needle aspiration (FNA) procedures and optimize thyroid nodule management.
Main Methods:
- A retrospective validation cohort of 4572 thyroid nodules with prior FNA was analyzed using a deep learning-based AI system.
- The AI's diagnostic performance was evaluated for correctly identifying benign nodules and misclassifying malignant ones.
- A comparison cohort was used to directly compare the AI's accuracy in identifying benign nodules against that of junior and senior radiologists.
Main Results:
- The AI system correctly identified 86.8% of benign nodules and reduced unnecessary FNAs from 68.5% to 9.1%.
- However, 8.6% of malignant nodules were misclassified as benign, primarily those with low or intermediate suspicion.
- In a comparison cohort, the AI achieved an 81.4% accuracy for benign nodules, outperforming junior (40%) and senior (55%) radiologists, with a superior area under the curve (AUC) of 0.88.
Conclusions:
- AI demonstrates potential as a "goalkeeper" tool, significantly decreasing unnecessary FNAs by accurately identifying benign thyroid nodules initially suspected of malignancy.
- Despite AI's high performance, continued vigilance and active surveillance are essential due to the small possibility of misclassifying low-aggressiveness malignant nodules.
