Premium
SU‐E‐T‐38: Validating the Use of a New Tumor Irradiation Quality Metric for Lung and Head and Neck Tumors: Total Clonogen Survival
Author(s) -
Zuniga AA,
Thorstad WL,
Oh JH,
Apte AP,
Bradley JD,
Deasy JO
Publication year - 2011
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.3611989
Subject(s) - medicine , univariate , multivariate statistics , logistic regression , mathematics , statistics , nuclear medicine , lung cancer , head and neck cancer , spearman's rank correlation coefficient , multivariate analysis , metric (unit) , receiver operating characteristic , univariate analysis , oncology , radiation therapy , operations management , economics
PURPOSE: Ideally, a single factor would allow objective ranking of dose distributions with respect to the probability of local control. The purpose of this work is to study the applicability of a new metric as a predictor of local control (LC) indicator in head and neck tumors and lung tumors. MATERIALS: Two datasets were used: (A) 80 head and neck (H&N) patients using IMRT, and (B) 56 non‐small cell lung cancer who received 3D conformal RT both treated at Washington University in Saint Louis at standard fractionation. 23 H&N and 22 lung patients failed locally. A simple model was used to derive the expectation value of the number of surviving clonogens per initial clonogen density (not initial number). Other dose‐volume parameters were compared to TCS. Correlation with LC for univariate and multivariate logistic models of the available parameters was assessed using the area under receiver operating characteristic curve (AUC) and Spearmanˈs rank correlation coefficient (Rs). We also tested bootstrap cross‐validation of the models. Results: Total Clonogen Survival was the most predictive variable on univariate analysis in both datasets using an optimal value of SF2=0.8 ((A) Rs=0.444 p‐value<0.0001 (B) Rs=0.441 p‐value<0.0001). In logistic regression for both datasets, all statistically significant multivariate models included TCS parameter. Multivariate logistic regression for (A) shows that a two parameter (TCS and stage_group) model was the most significant when cross‐validated (RsCV=0.4644, pCV‐value=0.02). Higher order models did not appear more predictive after cross‐validation. However, we report in an abstract at this meeting that machine learning using TCS can improve predictive power. Conclusions: TCS was shown to be a simple and useful parameter when predicting LC in these two datasets. It differs from cell‐kill based EUD by mechanistically accounting for tumor volume variation. Partially supported by NIH R01 grant CA85181.