PET-based radiomics as a preoperative predictor of pathologic lymph node metastasis in lung cancer.

266

Background: Lymph node metastasis (LNM) critically influences treatment strategy in lung cancer, as nodal involvement may alter surgical eligibility and the need for multimodality therapy. Preoperative staging relies mainly on CT and PET-CT, which are limited by suboptimal sensitivity and false-positive findings. We therefore developed a prediction model to improve LNM identification and support treatment decision-making. Methods: This study retrospectively enrolled 166 patients with lung cancer who underwent PET-CT for preoperative staging between 2013 and 2020. A total of 175 pathologically confirmed tumors were analyzed at the lesion level, with LNM as the prediction target. To avoid ambiguous lesion-level labeling, patients with multiple synchronous tumors and LNM were excluded. SUV _max -based thresholding was conducted to delineate the metabolic tumor volume of a tumor, from which peritumoral and boundary regions were generated. Radiomic features were extracted from every defined region according to the Image Biomarker Standardization Initiative guidelines, using multi-scale discretization. A stratified ten-fold nested cross-validation framework was employed to estimate model performance. In each outer fold, feature selection using recursive feature elimination and model training with a LightGBM classifier were conducted within the inner loop. Final performance metrics were averaged across the outer folds. For clinical deployment, lesion-level predictions were aggregated at the patient level, with a patient classified as high risk for LNM if any tumor was predicted to be positive. Results: At the lesion level, the model achieved a mean AUC of 81.0% ± 10.5%, with an accuracy of 81.1% ± 6.9%, sensitivity of 86.5% ± 19.5%, specificity of 79.6% ± 10.1%, positive predictive value (PPV) of 58.7% ± 10.4%, and negative predictive value (NPV) of 95.3% ± 6.8% across outer folds. At the patient level, aggregation of lesion predictions yielded an accuracy of 80.7%, sensitivity of 85.4%, specificity of 79.2%, PPV of 57.4%, and NPV of 94.3%. Compared with clinical staging (sensitivity 56.1%, NPV 84.5%), the proposed model demonstrated a substantial absolute increase in sensitivity (29.3%) and NPV (9.8%), while maintaining comparable specificity. Conclusions: The proposed PET-based radiomics model demonstrates promising performance for preoperative prediction of LNM. With substantially improved sensitivity and NPV compared with clinical staging, the model may serve as a noninvasive risk-stratification tool to assist in nodal evaluation and treatment planning.