Answers are grounded in this paper and cite their sources.
Background Cervical lymph node metastasis (CLNM) is the most common route of spread in papillary thyroid carcinoma (PTC) and a key determinant of surgical extent, radioactive iodine therapy and recurrence risk. Pre-operative ultrasound has limited sensitivity for nodal disease, and objective, reproducible risk-stratification tools are needed. We developed and validated a multimodal model fusing deep ultrasound image features with clinical risk factors, compared it with clinical and image-only models, and evaluated its transportability in an independent external cohort. Methods In this multicentre retrospective study, 607 consecutive PTC patients (2022–2026; N1, 48.4%) with pre-operative grey-scale ultrasound and definitive histopathology were included, plus an external validation set of 252 consecutive patients (N1, 83.3%). Lesions were segmented by a YOLOv8-seg model trained on 390 manually labelled frames (mask mAP50 = 0.985), retaining 1,322 network-generated masks and making delineation operator-independent. The development cohort was split 7:3 into training (n = 424) and internal validation (n = 183) sets; univariable and multivariable logistic regression identified independent clinical risk factors. Ultrasound deep features (896-d pooled activations) were extracted by a same-domain transfer-pretrained backbone, reduced to 30 principal components and fused with clinical factors in a logistic regression head. Models were evaluated by AUC with bootstrap 95% CIs, DeLong tests, calibration curves, DCA and SHAP. Results Age (OR 0.959, 95% CI 0.936–0.982), female sex (OR 0.316, 0.178–0.564), tumour diameter (OR 5.321, 2.438–11.617), minimal extrathyroidal extension (OR 2.001, 1.071–3.741) and multifocality (OR 6.106, 2.178–17.118) were independent predictors. On internal validation, the clinical, ultrasound deep-learning and fusion models achieved AUCs of 0.709 (0.641–0.777), 0.642 (0.561–0.723) and 0.737 (0.665–0.808); the fusion model significantly outperformed the imaging model (dAUC = 0.094, P = 0.0015) and showed the highest net benefit on DCA. The clinical model showed limited transportability in the external validation set (n = 252; AUC = 0.528), consistent with case-mix-dependent shift. Conclusions Fusing same-domain transfer-learning ultrasound deep features with clinical risk factors achieved the highest internal-validation AUC and clinical net benefit, significantly outperforming image-only modelling; external validation of the fusion model remains to be performed. The same-domain transfer strategy offers a reproducible paradigm for small-sample medical imaging tasks.
No takes yet. Share an insight, caveat, or question.
Zheng et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: