Development and Validation of a Picture Archiving and Communication System-Integrated Artificial Intelligence for Predicting Cervical Lymph Node Metastasis in Papillary Thyroid Carcinoma.
Authors
Affiliations (11)
Affiliations (11)
- The Fourth School of Clinical Medicine, Zhejiang Chinese Medical University, Hangzhou , Zhejiang, 310053, China.
- Department of Radiology, Affiliated Hangzhou First People's Hospital, Westlake University School of Medicine, Hangzhou , Zhejiang, 310006, China.
- The Second School of Clinical Medicine, Zhejiang Chinese Medical University, Hangzhou , Zhejiang, 310053, China.
- School of Artificial Intelligence, Hangzhou Dianzi University, Hangzhou , Zhejiang, 310018, China.
- Department of Radiology, Hangzhou First People's Hospital Tonglu Campus, Hangzhou , Zhejiang, 311500, China.
- Department of Radiology, Hangzhou Ninth People's Hospital, Hangzhou , Zhejiang, 310012, China.
- Department of Radiology, Affiliated Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou , Zhejiang, 310016, China.
- Department of Radiology, The Third People's Hospital of Yuhang District, Hangzhou , Zhejiang, 311100, China.
- Department of Radiology, Affiliated Hangzhou First People's Hospital, Westlake University School of Medicine, Hangzhou , Zhejiang, 310006, China. [email protected].
- School of Artificial Intelligence, Hangzhou Dianzi University, Hangzhou , Zhejiang, 310018, China. [email protected].
- Department of Radiology, Affiliated Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou , Zhejiang, 310016, China. [email protected].
Abstract
The objective of this study is to develop and validate a Picture Archiving and Communication System-integrated artificial intelligence (PACS-AI) tool for automated neck-level localization, three-dimensional (3D) segmentation, and metastasis risk prediction of cervical lymph nodes (LNs) in patients with papillary thyroid carcinoma (PTC). CT images from 710 patients with PTC, including 4942 LNs, were retrospectively collected from two medical centers and divided into training, internal test, and external test cohorts. The AI model consisted of a 3D TransUNet partition network, a Swin UNETR segmentation network, and a 3D U-Net classification network, all of which were integrated into the PACS. Additional data from 202 patients with PTC (including 202 LNs) were collected from the aforementioned two centers to evaluate the improvement in radiologists' diagnostic performance with PACS-AI assistance. In the internal and external test sets, the partition model achieved precision values of 0.919-0.990 and 0.848-1.000 across levels I-VI, respectively. The segmentation model showed mean Dice similarity coefficients (DSCs) of 0.674 and 0.633, with corresponding precision values of 0.832 and 0.722. The classification model achieved AUCs of 0.948 (internal) and 0.943 (external) for metastasis prediction. In the reader study, the classification model outperformed junior radiologists and significantly improved their diagnostic accuracy (all p < 0.05). The PACS-AI provides an integrated, visual, and clinically practical tool for the preoperative assessment of cervical LNs in patients with PTC, demonstrating the potential to enhance diagnostic accuracy among junior radiologists and to support broader clinical adoption.