Premium
SU‐E‐T‐01: Tumor Control Probability Models Derived Using Machine Learning for Head and Neck and Lung Tumors including a New Metric, Total Clonogen Survival (TCS)
Author(s) -
Oh J,
Zuniga A,
Thorstad W,
Bradley J,
Deasy J
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 - 10.1118/1.3611951
Subject(s) - medicine , rank correlation , head and neck cancer , lung cancer , nuclear medicine , spearman's rank correlation coefficient , radiation therapy , head and neck squamous cell carcinoma , cancer , correlation , machine learning , mathematics , computer science , geometry
Purpose: Empirically validated and highly predictive models of tumor control probability could help guide daily clinical practice, but are currently unavailable. Using a newly‐identified dosimetric predictor (Total Clonogen Survival (described elsewhere at this meeting)), and a newly acquired dataset on head and neck tumors, we explore the predictive power available in a TCP model using machine learning methods.Methods: We modeled two different datasets collected at Washington University School of Medicine: (A) lung cancer data consisting of 56 non‐small cell carcinoma patients (a local failure group (22) and a control group (34)) who received 3D conformal radiation therapy with a median prescription dose of 70 Gy (60–84 Gy) and had a median follow‐up of 32 months; (B) head and neck cancer data consisting of 80 squamous cell carcinoma patients (a local failure group (23) and a control group (57)) who received IMRT as definitive treatment with a median prescription dose of 70 Gy (66–72 Gy) and had a median follow‐up of 19 months. Using dose‐volume parameters extracted from these datasets, we found significant parameters for predicting local failure and evaluated the performance using support vector machines (SVM) with leave‐one‐out cross‐validation (LOO‐CV) in conjunction with several feature selection strategies. Results: For lung cancer data, V75 and TCS were chosen as significant parameters with a Matthewˈs correlation coefficient (rCV) of 0.462 and a Spearmanˈs rank correlation coefficient (RsCV) of 0.642. For head and neck cancer data, with Min dose, V70, and TCS, we achieved the best performance: rCV=0.367 and RsCV =0.554. Conclusions: Combining the newly identified metric TCS with other relevant metrics within a machine learning framework allowed us to produce a predictive TCP model that should be further tested for potential clinical use.