Longitudinal Radiomics-Based Prediction of Local Recurrence after Stereotactic Body Radiation Therapy for Early-Stage Non-Small Cell Lung Cancer

Abstract

Computed tomography (CT)-based radiomic analysis is an emerging tool that can help predict recurrence after radiation in non-small cell lung cancer (NSCLC). Most studies have focused on using radiomic features obtained from pre-treatment CT scans but not on post-treatment CTs. Our goal was to determine if radiomic features derived from post-treatment CTs perform better than those from pre-treatment CTs in predicting local recurrence after stereotactic body radiation therapy (SBRT) in patients with early-stage NSCLC. Patients with T1/T2 N0 early-stage NSCLC were retrospectively reviewed. Criteria for selection included patients with biopsy-proven NSCLC, non-ground glass solid lesions, 4-5 fraction SBRT treatments, and non-contrast thin-slice CTs available at pre-treatment baseline and at 3 and 6 months post-SBRT. Clinical and tumor characteristics were compared using Chi-square and Student's t-tests. CT image pre-processing was performed and 107 radiomic features were extracted using 3D Slicer and open-source software. Univariate analysis for local recurrence was determined by Cox regression with death as a competing risk. The Benjamini-Hochberg Procedure was applied to control the false discovery rate. Pearson correlation analysis was used to exclude redundant features (r > 0.8), and multivariate analysis was conducted with Random Forest on the top-performing features found on univariate logistic regression. Models were trained on a class-balanced loss to account for class imbalance. Twenty iterations of stratified 3 k-fold cross-validation were used to select the best model by the area under the receiver operating characteristic curve (AUC). We identified and analyzed 86 patient tumors, including 76 non-local recurrences and 10 local recurrences (49 males, 37 females; age, median 72, range 52-91 years). No differences in age, histology, standardized uptake values on positron emission tomography scans, gross tumor volume, and radiation dose were found between patients with and without local recurrence (all p >0.05). The median time to local recurrence was 18.9 months (range 5.5-45.6 months). At the baseline, 3-month, and 6-month timepoints, 2, 10, and 10 radiomic features predicted local recurrence on Cox regression univariate analysis, respectively (all q <0.01). On multivariate analysis of top-performing radiomic features, the 3-month timepoint performed the best with a mean (±standard deviation) AUC score of 0.82 (±0.13) compared to a mean AUC of 0.73 (±0.086) at baseline and a mean AUC of 0.74 (±0.10) at 6 months. Post-treatment radiomic features at 3 months outperformed pre-treatment radiomic features in predicting local recurrence after SBRT for NSCLC. These results suggest radiomic data from follow-up CTs may be helpful when developing radiomic models to predict local recurrence in patients with NSCLC.