Multimodal multi-task deep learning for preoperative prediction of central and lateral lymph node metastasis in papillary thyroid carcinoma.

Abstract

To develop a multi-task deep learning model using preoperative tumor ultrasound (US) and C6-level adipose tissue CT images to predict cervical lymph node metastases (LNM) at different anatomical sites in papillary thyroid carcinoma (PTC) patients. This retrospective study included 620 pathologically confirmed PTC patients from two medical centers. Independent predictors of central compartment (CLNM) and lateral compartment (LLNM) metastasis were identified through logistic regression, and key clinical variables were extracted. A multi-task feature interaction network (MT-FINet) with ResNet50 backbone was constructed to jointly incorporate tumor US, adipose CT, and clinical features. Model performance was evaluated in internal (I-T) and external (E-T) test sets. Models with adipose features outperformed single tumor-modality models. The multi-task strategy mitigated feature competition between CLNM and LLNM tasks and achieved superior results over single-task models. In the I-T, the AUCs for CLNM and LLNM prediction were 0.906 (95 % CI: [0.845, 0.948]) and 0.901 (95 % CI: [0.820, 0.956]); in the E-T, the corresponding AUCs were 0.860 (95 % CI: [0.792, 0.925]) and 0.868 (95 % CI: [0.795, 0.924]). These performances were significantly superior to independent diagnoses made by three experienced radiologists (average AUC = 0.687). Multivariate logistic regression further revealed that tumor size, gender, age, and the number of tumor foci were independent predictors of LNM in PTC patients. MT-FINet integrating tumor and adipose features significantly improved preoperative prediction of CLNM and LLNM. This model may support individualized surgical planning and provide reliable decision support for lymph node dissection in PTC patients.