Radiomics Predict Pathological Response in Non-Small Cell Lung Cancer

Abstract

Radiomics is an emerging field of quantitative imaging that aims to describe tumors using a large set of advanced imaging features. Pathological response is a direct measure of tumor response to neoadjuvant chemoradiation assessed at time of surgery. Predicting pathological response at an early time point would allow for potential modification of the treatment regimen. In this study we assessed if pre-treatment radiomics data are able to predict pathological response after neoadjuvant chemoradiation in patients with locally advanced non-small cell lung cancer (NSCLC). 127 NSCLC patients undergoing with trimodality treatment at a single institution were included in this retrospective study. The median radiation dose was 54 Gy with 107 (84%) and 20 (16%) patients that respectively had concurrent or induction chemoradiation. The median time to surgery after completion of chemoradiation was 1.4 months (0.3 to 5 months). Pathological response was evaluated at the time of surgery and scored as gross residual disease (GRD) or pathological complete response (pCR). Radiomic features (over 1600) were extracted from CT imaging obtained at the time of radiation treatment planning and reduced using unsupervised selection (principal component analysis). Association of selected features with pathological response was evaluated using area under the curve (AUC) analysis. Cross-validation (CV) was computed to evaluate model performance (100 splits, 80% training and 20% validation). Permutation test (1000 CV) was used to control model performance from random. At the time of surgery, 27 (21.3%) patients had a complete pathological response and 67 (52.7%) had a gross residual disease. PCA analysis selected 15 features for further analysis. Of these, seven features were predictive associated with GRD (AUC range from 0.61 to 0.66, P<0.05), and one with pCR (AUC=0.63, P=0.01). No conventional imaging features were predictive associated with pathological response (range AUC=0.51 to 0.59, P >0.05). Tumors that did not respond well to neoadjuvant chemoradiation were more likely to present with heterogeneous texture (GLCM Entropy, AUC = 0.61, P = 0.029) and rounder shape (Spherical disproportionality, AUC = 0.63, P <0.01). On cross validation, a combined model of radiomic and conventional imaging features performed significantly better than conventional measurements for prediction of pCR (0.68 vs 0.57, P<0.05) and GRD (0.65 vs 0.60, P<0.05) We identified CT radiomic features associated with pathological response in NSCLC patients undergoing trimodality therapy. Clinically used conventional imaging measurements, such as tumor diameter, were not associated with pathological outcome. This study demonstrates that radiomics can provide valuable complementary clinical information, and perform better than conventional imaging features for clinical outcomes.