Objective To construct a multimodal fusion model integrating ultrasound habitat radiomics features, clinical parameters, and immunological markers for preoperative assessment of central lymph node metastasis (CLNM) risk in clinically node-negative (cN0) papillary thyroid carcinoma (PTC) patients.

Methods A retrospective analysis was conducted on 748 PTC patients treated at Changzhou First People’s Hospital and Xuzhou Central Hospital from January 2022 to June 2024. Patients were randomly allocated to the training group (n = 404) and internal validation group (n = 179) with a ratio of approximately 7:3, with an additional external validation group (n = 165). Superpixel over-segmentation and K-means clustering algorithms were used to identify 3 different habitats in ultrasound images. Radiomics features were extracted from each habitat and screened using a three-step method. Multivariate Logistic regression analysis was used to analyze clinical and immunological parameters. Three fusion strategies of early fusion, late fusion, and ensemble fusion were used to integrate data from different modalities to build machine learning models.

Results Multivariate Logistic analysis showed that Hashimoto’s thyroiditis was a protective factor of CLNM (OR = 0.357, 95% CI: 0.146-0.873, P = 0.024), while multifocality (OR = 2.627, 95% CI: 1.142-6.039, P = 0.023) and elevated systemic immune-inflammation index (OR = 1.002, 95% CI: 1.001-1.003, P ＜ 0.001) were independent risk factors. In ultrasound habitat analysis, habitat 3 (hypoechoic heterogeneous region) showed the strongest association with CLNM. The ensemble fusion voting classifier performed best in predicting CLNM, with the areas under curve (AUC) of receiver operating characteristic (ROC) were respectively 0.98 (95% CI: 0.96-0.99), 0.98 (95% CI: 0.96-0.99) and 0.97 (95% CI: 0.95-0.99) in the training, internal validation, and external validation sets, the accuracies were respectively 0.93, 0.95 and 0.86, and sensitivity and specificity were both over 0.85.

Conclusions The multimodal habitat imaging fusion model provides an accurate and interpretable tool for preoperative assessment of CLNM risk in cN0 PTC patients, potentially improving clinical decision-making and optimizing individualized treatment strategies.