Artificial intelligence-based multi-modal multi-tasks analysis of thyroid ultrasound image features predicts thyroid cancer: a multicenter study.
Authors
Affiliations (9)
Affiliations (9)
- Department of Breast and Thyroid Surgery, Southwest Hospital, Army Medical University, Chongqing, 400038, China.
- College of Mathematics and Statistics, Chongqing University, Chongqing, 400038, China.
- Department of Ultrasound, Southwest Hospital, Army Medical University, Chongqing, 400038, China.
- Department of General Surgery, The People's Hospital of Dazu, Chongqing, 402360, China.
- College of Mechanical and Transportation Engineering, Chongqing University, Chongqing, 400038, China.
- Department of Oncology Surgery, Shaanxi Provincial People's Hospital, Xian, 710086, China.
- Department of Oncology Plastic Surgery, Hunan Province Cancer Hospital and The Affiliated Cancer Hospital of Xiangya School of Medicine, Central South University, Changsha, Hunan, China.
- Breast Surgery Department, Fuyang Cancer Hospital, Fuyang, 236000, China.
- Centre for Medical Big Data and Artificial Intelligence, Southwest Hospital, Army Medical University, Chongqing, 400038, China.
Abstract
Thyroid nodule ultrasound (US) images and their features are of great importance in thyroid nodule diagnosis, and can be helpful for radiologists' clinical decision-making. To evaluate whether an AI-assisted system can accurately characterize thyroid nodule ultrasound features and assist radiologists in diagnosing thyroid cancer. The AI-assisted system (MDT-TC) was trained and internally validated on B-mode US images from 7204 lesions in 6884 patients in Southwest Hospital (SW). The model performance was validated using three independent external validation cohorts. Echogenicity (ECH) and shape (SHA) are features of high importance for model recognition, and these features lead to excellent model performance. The model achieved up to 87.56% accuracy in determining ECH attributes and 69.21% in identifying shape categories. The AUC of the internal validation cohort and three independent external validation cohorts for MDT-TC were 0.951, 0.837, 0.816, and 0.871, respectively. The sensitivity values were 98.7%, 91.2%, 90.3%, and 85.6%, respectively. The AUC for the accurate diagnosis of radiologists with MDT-TC assistance was significantly higher than that of radiologists without MDT-TC assistance (p < 0.001). In addition, the AUC for the accurate diagnosis of junior doctors with MDT-TC assistance was significantly higher than that for those who did not (p < 0.01). MDT-TC incorporates radiomic features extracted from thyroid lesion US images, and can significantly improve the diagnostic performance of radiologists. This result was particularly strong for junior doctors. Therefore, our data support the idea that MDT-TC can help to identify patients with thyroid cancer and could greatly benefit clinical practice.