Multicenter CT-Based Deep Learning for Preoperative Nodal Risk Stratification in Papillary Thyroid Carcinoma.
Authors
Affiliations (9)
Affiliations (9)
- Department of Thyroid and Breast Surgery, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- Division of Hepato-Pancreato-Biliary Surgery, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- Department of Thyroid and Breast Surgery, The Third Affiliated Hospital of Sun Yat-Sen University, Guangzhou, China.
- Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- Radiology Department, Nanchuan Hospital of Chongqing Medical University, Chongqing, China.
- Radiology Department, Red Cross Hospital of Yulin City, Yulin, China.
- Department of Physician Services, Xinjiang International Medical Center (Xinjiang Medical Hospital), Urumqi, China.
- Department of Head and Neck Surgery, Hubei Cancer Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
- Department of Nuclear Medicine, Tianjin Medical University General Hospital, Tianjin, China.
Abstract
Accurate preoperative risk stratification for lymph node metastasis (LNM) in papillary thyroid carcinoma (PTC) remains difficult. We developed and validated ThyLNT, a Transformer-based 2.5D deep learning model using preoperative computed tomography (CT), and explored related biological features in multi-omics analyses. A total of 1560 patients with PTC from six hospitals were retrospectively included. The Tongji Hospital cohort was divided into training and internal validation sets, and five external cohorts were used for independent testing. Seven 2.5D tumor slices were extracted per patient, encoded with a DenseNet201 backbone, and fused by a Transformer. Model performance was compared with radiomics, ensemble learning, and multi-instance learning using AUC, calibration, decision curve, and precision-recall analyses. Bulk RNA-seq, single-cell RNA-seq, spatial transcriptomics, and spatial metabolomics were used to explore related biological features. ThyLNT achieved an AUC of 0.882 (CI, 0.858-0.905) in the training cohort, 0.787 (CI, 0.738-0.835) in the internal validation cohort, and 0.772 (CI, 0.662-0.878) to 0.827 (CI, 0.763-0.885) across the five external cohorts. It showed multicenter discrimination, whereas calibration and threshold-dependent operating characteristics varied across cohorts. Multi-omics analyses provided exploratory biological context, including candidate associations with vascular endothelial growth factor A-related signaling and lipid metabolism. ThyLNT demonstrated multicenter performance for preoperative LNM prediction in PTC and may support risk stratification, although calibration, threshold use, and clinical utility require prospective validation.