Deep learning combined habitat radiomics analysis of central lymph node metastasis in papillary thyroid carcinoma

Accurate preoperative assessment of central lymph node metastasis (CLNM) in papillary thyroid carcinoma (PTC) remains challenging. This retrospective study was registered at ClinicalTrials.gov (NCT06725628) on December 9, 2024. A total of 934 patients with confirmed PTC who underwent cervical central lymph node dissection between January 2023 and December 2024 were included. Radiomics features were extracted and selected. Deep learning models were trained and used for automatic segmentation and classification. We developed and validated a Thy-CLNM Fusion model, which achieved an AUC of 0.862, with a sensitivity of 78.9% and a specificity of 79.3% in the validation cohort. We compared the proposed model with other models that also predict CLNM, and the results demonstrated competitive performance compared with previously reported models. Histopathological slides from 231 patients were retrospectively reviewed to provide pathological interpretability. In the CLNM low-risk group, model activation predominantly localized to hyperechoic Habitat I regions, corresponding pathologically to fibrous hyperplasia and active lymphocyte–tumor cell interactions. In the CLNM high-risk group, model attention showed dominant localization in hypoechoic Habitat II regions, which were associated with nuclear atypia, poor differentiation, reduced intercellular adhesion, and vascular invasion. The proposed model offers an additional preoperative decision-support option for CLNM risk assessment in PTC patients who are planning to undergo surgery and are being considered for central lymph node dissection.