Use of Radiomics to Predict Treatment Response in CNS, Head, Neck and Lung Tumors

Abstract

The ability to predict treatment response and local tumor control in cancers treated with radiation might allow for treatment de-escalation in responsive tumors and treatment escalation in non-responsive tumors. The past decade has seen substantial developments in radiomic technology, with some investigators finding radiomic models that predict treatment response and toxicities. While it is standard of care to adapt treatment volumes in the setting of significant tumor shrinkage, it is unclear if deescalating their overall treatment might jeopardize their treatment outcomes. The purpose of this study is to investigate the radiomic features in patients receiving definitive treatment that were re-simulated for significant tumor shrinkage. We looked for radiomic features correlating with tumor response in 12 female and 16 male patients with a median age of 64 and ECOG status between 0-2. These patients received radiation with curative intent in the CNS, head, neck, or lungs. These patients were re-simulated during the course of their treatment for significant tumor shrinkage to optimize their treatment volumes and minimize radiation dose to organs at risk. Of the 28, 1 patient had CNS cancer, 22 had head or neck cancers, and 15 had lung cancer. Pre- and post- radiation CT planning scans were assessed after the initial gross tumor volume (GTV) was shrunk by 3mm to avoid interface effects of proximal structures. Slope for a straight-line fit was computed between each potential radiomic feature and post-treatment tumor reduction (as calculated using GTV - 3mm) to identify radiomic features that correlate with treatment response. We used the criteria slope >0.5, and slope > 0.6 cut offs to determine the highly predictive radiomic features. First order mean, which is a surrogate for CT density or the HU value showed a slope of 0 indicating HU as a poor predictor of treatment response. This radiomic feature was deemed stable between patients and between scans. 17 radiomic features showed a slope >0.5 and 5 features demonstrated a slope > 0.6. The 5 most promising included: original_firstorder_minimum, original_glcm_joint average, original_glcm_SumAverage, original_gldm_LargeDependenceHighGrayLevelEmphasis and original_glxzm_LargeAreaHighGrayLevelEmphasis. Of the patients within this group of predictive radiomic features, none had recurrence at the treated disease site. Our study identified 22 radiomic features that might be used to predict treatment response in malignant tumors of the CNS, head, neck and lung. Current tools we use in the clinic, including density and HU values, showed no predictive power for tumor progression but remained stable between patients and between scans. Further research is necessary to determine if these 22 unique radiomic features can be utilized in predicting patient outcomes.