A Transformer-Based 2.5D Deep Learning Model for Preoperative Prediction of Lymph Node Metastasis in Papillary Thyroid Carcinoma

Authors

Abstract

BackgroundAccurate preoperative prediction of lymph node metastasis (LNM) in papillary thyroid carcinoma (PTC) remains challenging, particularly in clinically node-negative (cN0) patients, leading to potential overtreatment. We aimed to develop and validate a Transformer-based 2.5D deep learning model (ThyLNT) using preoperative computed tomography (CT) images for robust prediction of LNM and to explore its underlying biological basis through multi-omics analyses. MethodsA total of 1,560 PTC patients from six hospitals were retrospectively included. The Tongji Hospital cohort (n=1,010) was divided into training (70%) and internal validation (30%) sets, while five independent institutions served as external test cohorts. For each lesion, seven 2.5D slices were extracted and modeled using a DenseNet201 backbone. Slice-level features were integrated using a Transformer-based feature-level fusion strategy and compared with ensemble learning, multi-instance learning (MIL), and traditional radiomics approaches. Model performance was assessed using area under the receiver operating characteristic curve (AUC), calibration analysis, decision curve analysis (DCA), and precision-recall curves. Multi-omics analyses, including bulk RNA-seq, single-cell RNA-seq, spatial transcriptomics, and spatial metabolomics, were performed to investigate biological correlates. ResultsThe Transformer-based model consistently outperformed comparator models across cohorts. In the training and validation cohorts, ThyLNT achieved AUCs of 0.882 and 0.787, respectively, with external AUCs ranging from 0.772 to 0.827. Compared with ultrasound (US) and CT, ThyLNT showed superior predictive performance (all P < 0.001 in the validation cohort). Simulation analysis in cN0 patients suggested that ThyLNT could reduce unnecessary lymph node dissection (LND) from 52.16% to 4.88%. Transcriptomic analysis combined with WGCNA and correlation analysis identified VEGFA as the gene most strongly associated with ThyLNT prediction scores. Single-cell and spatial transcriptomic analyses suggested metastasis-related tumor microenvironment remodeling, while enrichment analysis of genes affected by virtual knockout of VEGFA indicated involvement of angiogenesis- and epithelial-mesenchymal transition (EMT)-related pathways. Spatial metabolomics further revealed coordinated lipid metabolic reprogramming in metastatic tissues. These findings suggest that ThyLNT provides robust predictive performance while capturing biologically relevant features associated with metastatic progression.

Tags

Topics