Predicting Circulating Tumor Cells from Intratumoral Radiomic Heterogeneity of 18F-FDG-PET/CT in Early Stage Non-Small Cell Lung Cancer Treated with Stereotactic Body Radiation Therapy

Abstract

Patients with early stage non-small cell lung cancer (ES-NSCLC) treated with stereotactic body radiation therapy (SBRT) have excellent local control rates although they have a relatively high rate of regional and distant recurrences. Recently, patients with elevated pre-treatment (pre-RT) circulating tumor cells (CTCs) or persistently detectable post-treatment (post-RT) CTCs were found to have significantly increased risk of regional and distant recurrence outside the treatment volume. Knowledge about radiomics biomarkers associated with CTCs may have additional value in predicting these recurrences. This study aims to evaluate the predictability of radiomic phenotypes of intratumoral heterogeneity for pre- and post-RT CTCs in ES-NSCLC patients treated with SBRT. A total of 56 patients with stage I NSCLC treated with SBRT underwent 18F-FDG-PET/CT imaging pre-SBRT and post-SBRT (median, 5 months; range, 3 to 10 months). For each patient, CTCs were assessed via a telomerase-based assay before (“pre”) and within 3 months after SBRT (“post”), and dichotomized at 5 and 1.3 CTCs/mL. Pre-RT, post-RT and delta radiomics features were extracted from the gross tumor volume of the original, wavelet-filtered, Laplacian of Gaussian-filtered PET/CT images, each of which included 1,562 CT and 1,548 PET radiomics features that consisted of shape, first-order, second-order, higher-order local gray-level statistics and global texture features. CT and PET radiomics features were considered different omics data types. Two (CT and PET)-block radiomics-based data integration analysis for biomarker discovery using latent components (DIABLO) was performed to seek for common information between CT and PET radiomics data types through feature selection, while discriminating CTC levels. ROC AUC value and 20-repeated 5-fold cross-validation (CV) were used to evaluate the prediction performance of the DIABLO models. CV prediction scores between one class vs the other were compared using the Wilcoxon rank sum test. For predicting pre-SBRT CTCs, a pre-SBRT radiomics-derived DIABLO model showed the CV AUC of 0.841 (p = 0.018) with two global texture features from wavelet-filtered CT and PET. For predicting post-RT CTCs, post-RT and delta radiomics-derived DIABLO models showed the CV AUCs of 0.794 (p = 0.018) with four global texture features from wavelet-filtered CT and original PET, and 0.799 (p = 0.013) with three first-order features from wavelet-filtered PET and a single second-order texture feature from wavelet-filtered CT. Radiomics signatures that reflect intratumoral anatomic and metabolic heterogeneities derived by maximizing common information between CT and PET may reveal surrogate biomarkers for pre-RT and post-RT CTCs.