Structural analysis based on machine learning methods for detecting mediastinal lymph node metastases in non-small cell lung cancer cases

Aim: The aim of our study was designed as a retrospective, single-center study aiming at preoperative prediction of medastinal lymph node metastasis (LNM) in non-small cell lung cancer (NSCLC) patients. The primary objective of the study was to develop machine learning models that predict mediastinal LNM with high accuracy by combining radiomic features obtained from contrast-enhanced thorax computed tomography (CT) images with clinical parameters. Materials and Methods: The images of 155 patients who were diagnosed with primary NSCLC, underwent contrast-enhanced thoracic CT scans, and underwent mediastinoscopy and lymph node sampling in our hospital between January 2020 and December 2024 were examined. 156 lymph nodes from 97 patients were manually segmented. Each segmented lymph node was treated as an independent unit. Machine learning-based structural analysis and modeling were created by adding lymph node short axis, maximum standardized uptake value (SUV-max) and clinical information to the segmentation data. The sensitivity, specificity, negative predictive values ​​(NPV) and area under the curve (AUC) results of these models were evaluated. Results: The performance of the models increased when lypmh node short axis, SUV-max and clinical data were added in addition to radiomics features. Stacking ensemble model (AUC: 0.894, Sensitivity: 94.1%, NPV: 96%), a current technique combined with radiomics and 13 clinical data ( lymph node SUV-max, lymph node short axis, age …), has high potential for clinical decision support and demonstrated the highest performance in detecting mediastinal LNM. Conclusion: This study demonstrated that preoperative prediction of mediastinal LNM in NSCLC patients is possible with high accuracy by integrating radiomics and clinical information. Key words: NSLC, mediastinal lymph node metastasis, stacking ensemble learning, machine learning