Dual-modal radiomics ultrasound model to diagnose cervical lymph node metastases of differentiated thyroid carcinoma: a two-center study

Objectives: To establish and validate a dual-modal radiomics nomogram from grayscale ultrasound and color doppler flow imaging (CDFI) of cervical lymph nodes (LNs), aiming to improve the diagnostic accuracy of metastatic LNs in differentiated thyroid carcinoma (DTC).

Methods: DTC patients with suspected cervical LNs in two medical centers were retrospectively enrolled. Pathological results were set as gold standard. We extracted radiomic characteristics from grayscale ultrasound and CDFI images, then applied lasso (least absolute shrinkage and selection operator) regression analysis to analyze radiomics features and calculate the rad-score. A nomogram based on rad-score, clinical data, and ultrasound signs was developed. The performance of the model was evaluated using AUC and calibration curve. We also assessed the model's diagnostic ability in European Thyroid Association (ETA) indeterminate LNs.

Results: 377 DTC patients and 726 LNs were enrolled. 37 radiomics features were determined and calculated as rad-score. The dual-modal radiomics model showed good calibration capabilities. The radiomics model displayed higher diagnostic ability than the traditional ultrasound model in the training set [0.871 (95% CI: 0.839-0.904) vs. 0.848 (95% CI: 0.812-0.884), p<0.01], internal test set [0.804 (95% CI: 0.741-0.867) vs. 0.803 (95% CI: 0.74-0.866), p = 0.696], and external validation cohort [0.939 (95% CI: 0.893-0.984) vs. 0.921 (95% CI: 0.857-0.985), p = 0.026]. The radiomics model could also significantly improve the detection rate of metastatic LNs in the ETA indeterminate LN category.

Conclusions: The dual-modal radiomics nomogram can improve the diagnostic accuracy of metastatic LNs of DTC, especially for LNs in ETA indeterminate classification.