Program Information
Robust Statistical Estimators for IMRT Plan Analysis
O Nohadani1*, I Das2 , (1) Northwestern University, Evanston, IL, (2) NYU Langone Medical Center, New York, NY
Presentations
TU-C1-GePD-JT-2 (Tuesday, August 1, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Theater
Purpose: To investigate the efficacy of maximum likelihood estimators (MLE) in IMRT plan analysis. Robust estimators are compared to standard MLE for a large cohort of cancer patients treated with IMRT.
Methods: Standard MLEs are known to be sensitive to sample choice and sample size. This renders the derived conclusions unreliable, despite their clinical popularity. Using robust optimization principles, we develop robust MLEs that are immune against uncertainties and remain reliable even in worst-case deviations. This method is well suited for large-scale and high-dimensional data. We compare them to standard MLEs on 491 delivered treatment plans based on their DVH points Dₓ. All treatments followed the same protocols and were planned with the same TPS. We simulate typical institutional analysis for means and covariances of Dₓ. An estimator is considered reliable, when its value remains stable for differing sample sets, i.e., when estimators exhibit a narrow spread over sampled data sets. We conduct two experiments: a) constant sample size but varying samples and b) for varying sample sizes.
Results: The spread of robust means are stable and the same as the standard sample means for input data deviations. With regard to sample-size dependence, sample means exhibit a sizable sensitivity, whereas the robust means remain unchanged. These results apply to all relevant Dₓ, making robust estimators superior over the clinically relevant range. These advantages are magnified for higher estimators, such as covariances and correlation coefficients which are key metrics for analyzing protocol adherence.
Conclusion: Robust estimators are immune against variations in sample and sample size, hence are suited to provide more reliable dosimetric plan analysis. Their computation is efficient for big data and independent of the error sources.
Funding Support, Disclosures, and Conflict of Interest: O. N. acknowledges the support of National Science Foundation.
Contact Email: