Radiomics and Deep Learning Features Based on 18F-FDG PET/CT for Predicting Occult Lymph Node Metastasis in Non-small Cell Lung Cancer

<title>Abstract</title> <p> Background Preoperative identification of occult lymph node metastasis (OLM) in non-small cell lung cancer (NSCLC) remains a diagnostic hurdle, as standard imaging often fails to detect micrometastases, and single-modality radiomic approaches may not fully capture crucial tumor microenvironment information. To overcome this, we sought to construct an artificial intelligence-driven model that synergistically combines tumor radiomic features, deep learning (DL) features derived from ¹⁸ F-FDG PET/CT, and clinical parameters to predict OLM in NSCLC patients. Methods This retrospective analysis included 358 NSCLC patients (clinical stage T1-3N0M0) from Ningbo Mingzhou Hospital, randomly allocated to training and internal validation sets at an 8:2 ratio. An additional 132 patients from Suzhou Hongci Hospital formed an external validation cohort. Radiomic features were derived from manually segmented tumor volumes on PET/CT. Simultaneously, DL features were extracted from the slice with the largest tumor area and its two neighboring slices [(n + 2) and (n-2)]. Feature selection was executed via a multi-step process incorporating Spearman's correlation analysis, the minimum redundancy maximum relevance (mRMR) algorithm, and least absolute shrinkage and selection operator (LASSO) regression. The resultant features were fed into various machine learning classifiers to identify the optimal performer. We established seven distinct models and benchmarked their performance using the area under the receiver operating characteristic curve (AUC) and accuracy. The optimal cutoff was derived from the Youden index of the best model, stratifying patients into high- and low-risk groups for survival comparisons. Results Among the seven models, the logistic regression-based deep learning-radiomics-clinical (DLRC) model demonstrated superior predictive power, achieving AUCs of 0.890 and 0.872 in the internal and external validation cohorts, respectively. It also delivered the highest accuracy (0.833 and 0.826) in these cohorts. Kaplan-Meier analyses revealed that patients in the high-risk group experienced significantly worse progression-free survival (PFS) and overall survival (OS) compared to their low-risk group ( <italic>χ</italic> ²  = 7.339 and 9.464, respectively; all <italic>p</italic>  &lt; 0.05). Conclusion The proposed fusion model, which integrates PET/CT-based radiomics, DL features, and clinical variables, demonstrates robust predictive capability for OLM in NSCLC. This non-invasive tool holds promise for tailoring lymph node dissection strategies and refining patient selection for postoperative adjuvant therapy. </p>