Changes in Radiomic Features During Radiation Therapy as Predictors for Outcome in Locally Advanced Non-Small Cell Lung Cancer

Abstract

Local recurrence (LR) is common in locally advanced lung cancer, and predictive models may allow risk-based treatment paradigms and adaptations. Radiomic features, which examine quantitative tissue characteristics from radiographic images, have been shown to successfully predict outcomes in various tumors. In lung cancer, pretreatment radiomic features have been studied far more extensively than intra-treatment changes. We hypothesized that intra-treatment scans provide additional information compared to pre-treatment scans alone. Forty-one patients with locally advanced non-small cell lung cancer (Stage III = 38, Stage II = 3) underwent multiple fan beam CT scans during radiotherapy under IRB approved imaging protocols. All were treated with definitive radiotherapy using conventional fractions of 1.8 or 2 Gy to a median dose of 66 Gy. Thirty-six patients also received chemotherapy (29 concurrent, 5 sequential). The primary tumor was contoured on fan beam CT scans acquired pre-treatment and during treatment. Thirty-five repeatable radiomic texture features were extracted from the delineated primary tumor from the following families: histogram (HIST), gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), and neighborhood gray tone different matrix (NGTDM). A hierarchical clustering method was implemented to improve robustness by grouping related features, and the medoid feature for each cluster was selected with those present in more than 50% of time points included with clinical factors for modeling. We used a proportional hazards model for each time point to assess the contribution of each baseline and radiomic feature towards hazard of local recurrence. Twenty (48.8%) patients had LR at the primary tumor site. Median overall survival was 1.74 years, and median follow-up time for living patients was 6.45 years. Three radiomic features (dissimilarity, contrast calculated from the NGTDM, and low gray run emphasis) were statistically significant on the pre-treatment scan. However, none of these were significant during scans acquired at Weeks 3, 4, and 5 during treatment. For the intra-treatment scans, age and patient performance along with two radiomic features (maximum and strength) were shown to be significant across Weeks 3, 4, and 5, and three additional features (variance, mean, and contrast) showed a trend towards significance for all intra-treatment time points. Most predictive models generated by radiomics examine the features of pre-treatment scans. According to our findings, the radiomic features found to be statistically significant on pre-treatment vs intra-treatment scans were mutually exclusive, with the intra-treatment features having a higher predictive value. A possible explanation is that intra-treatment radiomic features better reflect the tumor’s response to treatment and should therefore be incorporated into a predictive model.