Multimodal Ultrasound-Based Decision Support for Bethesda IV Thyroid Nodules: Integration of Clinical, Radiomics, Deep Learning, and Topological Features.
Authors
Affiliations (4)
Affiliations (4)
- Department of Medical Ultrasound, Affiliated Hospital of Jiangsu University, Zhenjiang, China.
- Cancer Institute of Jiangsu University, Zhenjiang, China.
- Department of Medical Ultrasound, Affiliated Hospital of Jiangsu University, Zhenjiang, China. [email protected].
- Cancer Institute of Jiangsu University, Zhenjiang, China. [email protected].
Abstract
To develop and internally validate a multimodal ultrasound-based decision support framework for benign-malignant risk stratification of Bethesda IV thyroid nodules by integrating clinical information, radiomics, deep learning-derived, and topological features. This retrospective single-center study included 243 patients with Bethesda IV thyroid nodules who underwent preoperative ultrasound between January 2020 and December 2025. The cohort included 138 malignant and 105 benign lesions and was split at the patient level using stratified random allocation at a 7:3 ratio. Four fusion models were developed: M5 (clinical information + radiomics), M6 (clinical information + deep learning), M7 (clinical information + radiomics + deep learning), and M8 (clinical information + radiomics + deep learning + topological data analysis). Model performance was evaluated using receiver operating characteristic analysis, decision curve analysis, calibration curves, DeLong tests, net reclassification improvement, integrated discrimination improvement, and model interpretability analysis. In the validation cohort, the observed AUCs were 0.836 for M8, 0.829 for M7, 0.828 for M5, and 0.805 for M6. However, no pairwise AUC difference among M1 and M5-M8 remained statistically significant after Holm-Bonferroni correction. Compared with the clinical baseline model, M8 showed a significant continuous net reclassification improvement of 0.750 (95% CI, 0.310-1.174; p = 0.002), whereas the integrated discrimination improvement was not significant (0.029; 95% CI, - 0.050 to 0.105; p = 0.478). Interpretability analysis showed that the full-fusion model incorporated contributions from clinical, radiomics, deep learning-derived, and topological features. The multimodal ultrasound-based framework provided a structured approach for individualized risk assessment of Bethesda IV thyroid nodules. Although statistical superiority of the full-fusion model over the clinical baseline or the other fusion models was not established, its potential utility lies in providing an interpretable, probability-based multimodal risk estimate for case-level preoperative assessment rather than a stand-alone treatment decision. Prospective multicenter external validation across different institutions and ultrasound systems is required before clinical implementation.