Deep Learning-Driven Multimodal Integration of Ultrasound and CT for Identifying Central Lymph Node Metastasis in Papillary Thyroid Carcinoma.

OBJECTIVE: To evaluate the utility of multimodal imaging biomarkers in papillary thyroid carcinoma (PTC), this study developed a deep learning model using ultrasound (US) and computed tomography (CT) imaging to predict central lymph node metastasis (CLNM) in PTC by integrating radiomic features of perivascular adipose tissue (PVAT) at the cervical vertebra 6 (C6) level. METHODS: We retrospectively analyzed 477 PTC patients from two centers. Using ResNet101 with Hierarchical Attention Integration (ResHAI-101), we extracted deep learning features of the tumor area from US images. Separately, we extracted radiomic features of PVAT from CT images. Univariate and multivariate analyses were conducted based on clinical data, followed by feature selection and integration. RESULTS: Multimodal prediction model performance was quantified by the area under the receiver operating characteristic curve (AUC). Predictive performance was validated using an external test cohort (ETC). The Multimodal Prediction Model 3, integrating deep learning features, PVAT radiomic features and clinical features, showed superior diagnostic performance for predicting CLNM compared to the Unimodal Model, achieving AUCs of 0.915 in the internal test cohort and 0.890 in the ETC. Binary logistic regression identified gender, age, tumor size and extrathyroidal extension as independent risk factors for CLNM. CONCLUSION: This study establishes and validates the significant impact of PVAT on CLNM in PTC, providing a non-invasive tool for pre-operative risk assessment to aid in surgical decision-making.