Prediction of non-small cell lung cancer N2 metastasis using HIF-1α expression–related multicenter 18 F-FDG PET/CT radiomics

The prognosis for stage III N2 metastatic non-small cell lung cancer (NSCLC) is poor. We aimed to develop an N2 metastasis prediction machine learning model using multicenter 18 F- fluorodeoxyglucose (FDG) positron emission tomography (PET) radiomics correlating with hypoxia-inducible factor (HIF)-1α expression levels. Internal and external cohorts consisted of data of 66 patients from public databases and 102 patients from external institutions. Hub genes associated with metastasis were identified via functional enrichment analysis using stage III N2 metastasis and stage Ⅰ non-metastasis patients. The data of patients with pathological stage T2a or higher (tumor size > 3 cm) were extracted. Features were calculated from 18 F-FDG PET/computed tomography (CT) images; internal and external cohorts were harmonized using the ComBat algorithm. Image features of the prediction model were selected based on the area under the receiver operating characteristic curve (AUC). Using functional gene analysis, HIF-1α was confirmed to be associated with N2 metastasis. Either radiomics correlated with hub genes or HIF-1α gene expression levels were used to construct RF models. The prediction performance of the model was the highest when using image features correlated with HIF-1α expression (accuracy = 0.83, AUC = 0.819). The RF model utilizing harmonized image features showed high performance for the cohort with high tumor size. Harmonization of radiomics is required when developing a machine learning model for predicting N2 metastasis using multicenter data. As lymph node involvement is the main prognostic factor, the prediction of N2 metastasis could facilitate personalized therapeutic strategies for NSCLC.