A lightweight, computationally efficient deep learning system for preoperative prediction of occult lymph node metastasis in NSCLC: a multicenter study

Accurate preoperative identification of occult lymph node metastasis in non-small cell lung cancer is crucial for staging and surgical planning, yet conventional computed tomography (CT) assessment alone remains inadequate. In this multicenter study, we developed and validated four deep learning models using preoperative CT images from 2464 patients, with a focus on evaluating the performance and computational efficiency trade-offs between two-dimensional (2D) and three-dimensional (3D) convolutional architectures. The primary aim was to determine whether a computationally efficient 2D architecture could achieve diagnostic performance comparable to resource-intensive 3D models, thereby assessing its suitability for real-time clinical use. The lightweight Res2Net_2D architecture achieved an optimal balance between high predictive accuracy (AUC = 0.871) and superior computational efficiency (12 million parameters; 0.5 s per case) in external validation, demonstrating strong potential for real-time clinical integration. Explainable artificial intelligence (AI) analysis identified spiculation sign (31.2%) and solid proportion (22.1%) as the most influential predictors, providing interpretable radiographic biomarkers for preoperative risk assessment.