Involved-Field Irradiation vs Elective Nodal Irradiation for Locally Advanced Thoracic Esophageal Squamous Cell Carcinoma: A Comparative Interim Analysis of Clinical Outcomes and Toxicities (NCT01551589, CSWOG 003)

Abstract

Purpose/Objective(s): Cancer treatment outcome predictors are being examined in recent studies. Quantitative image features have shown potential for predicting treatment outcomes in several body sites. This study focused on quantitative analysis of cervical cancer patient’s PET/CT scans. Materials/Methods: We investigated the treatment outcome predictive power of SUV, textural and shape features from pre-treatment F18-FDG PET/CT scans. A cohort of 80 patients diagnosed with cervical cancer, FIGO stage IB-IVA, age range 25e86 years, (median age at diagnosis: 50 years) all treated between 2009 and 2014 with external beam radiation therapy to a dose range between: 45e50.4 Gy (median Z 45 Gy), concurrent cisplatin chemotherapy and MRI-based brachytherapy to a dose of 20e30 Gy (medianZ 28 Gy). Metabolic Tumor Volume (MTV) in patient’s primary site was delineated on pretreatment PET/CT by two board certified Radiation Oncologists. More than 80 quantitative image features were computed for each volume. Distant Metastases (DM) and Locoregional Recurrence (LRR) were analyzed based on each patient’s clinical record. Textural features were extracted using Co-occurrence Matrix (COM), Run Length Matrix (RLM), Gray Level Size Zone Matrix (GLSZM), and Intensity Based Matrix (IBM). SUVmax and SUVpeak were all measured from MTV. The correlation between the extracted features, SUV, and treatment outcomes were measured by three tests; Spearman’s rho, Pearson and Receiver Operator Curve (ROC). Results: Pearson’s and Spearman’s tests correlation scale for this study was: No correlation 0e0.19; Weak 0.2e0.249; Moderate 0.25e0.4; relatively strong 0.41e0.5; and strong 0.5e1. A total of 23 features showed correlation with both treatment outcomes. With DM: 10 features showed moderate correlation and 1 relatively strong correlation. With LRR: 21 features showed moderate correlation. SUVpeak correlated moderately with DM only on both tests. SUVmax showed weak correlations with all outcomes. ROC results were concordant with other tests. However, SUVpeak and SUVmax correlated with both DM and LRR. The correlation coefficients are shown in Table 1. Conclusion: Quantitative image features, based on four different calculation algorithms, predicted treatment outcomes better than SUV on three statistical tests. The results suggest that Radiomics may produce more robust tools to predict for treatment outcomes. Author Disclosure: B.A. Altazi: None. G.G. Zhang: None. E.G. Moros: None. M.C. Biagioli: None. D.C. Fernandez: None.