A Deep Learning Model for Preoperative Prediction of Lymph Node Metastasis in cT1-Stage Lung Adenocarcinoma: A Multicenter External Validation Study

Abstract

Purpose

To develop and validate a deep learning (DL) model for preoperative prediction of lymph node metastasis (LNM) in clinical T1-stage lung adenocarcinoma (LUAD), and to compare its performance with conventional semantic and radiomics signatures.

Methods

This multicenter retrospective study enrolled 2503 patients with 2568 pathologically confirmed cT1-stage LUAD nodules from eight institutions. Data from six centers (1994 patients/2059 nodules) were randomly divided into training (1600 patients/1664 nodules) and internal test (394 patients/395 nodules) cohorts. Two independent external validation cohorts (Set-1: 283 patients/nodules; Set-2: 226 patients/nodules) were included. Three predictive models were developed as follows: 1) a semantic model incorporating spiculation, pleural traction, air bronchogram, and vacuole signs; 2) a radiomics model; and 3) ResLNM—a residual network-based DL model. Diagnostic performance was evaluated using the area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA).

Results

Both ResLNM and radiomics models significantly outperformed the semantic model in predicting LNM (AUC range: 0.71–0.85 and 0.72–0.84 vs. 0.58–0.74, respectively; P<0.05). While ResLNM demonstrated comparable performance to the radiomics model in the internal test set (AUC: 0.85, 0.81–0.89 vs. 0.84, 0.80–0.88; P = 0.624) and external validation set-2 (0.71, 0.63–0.79 vs. 0.72, 0.64–0.80; P = 0.472), it achieved superior accuracy in external validation set-1 (0.82, 0.76–0.88 vs. 0.77, 0.71–0.83; P = 0.039). DCA confirmed the clinical superiority of ResLNM. Notably, integrating ResLNM with either semantic or radiomics signatures provided no incremental value (P>0.05), whereas incorporating radiologically mediastinal enlarged lymph node status (short-axis diameter ≥10 mm) significantly enhanced predictive performance, achieving AUCs of 0.88 (95%CI:0.85–0.92), 0.89 (95%CI:0.84–0.94), and 0.76 (95%CI:0.68–0.83) in the test set and two validation cohorts, respectively.

Conclusion

The ResLNM model provides a clinically feasible tool for preoperative LNM prediction in cT1-stage LUAD, outperforming conventional semantic and radiomics approaches. Its performance can be further optimized by integrating routinely available lymph node size criteria, offering potential to refine surgical decision-making and reduce overtreatment.