Program Information
A Novel Error Detection Methodology for IMRT Quality Assurance Using Radiomic Features of Gamma Distributions
L Wootton1*, A Chaovalitwongse2 , M Nyflot1 , E Ford1 , (1) University of Washington Department of Radiation Oncology, Seattle, WA, (2) University of Arkansas Department of Industrial Engineering, Fayetteville, Arkansas, (3) University of Washington Department of Radiology, Seattle, WA
Presentations
MO-DE-FS1-1 (Monday, July 31, 2017) 1:45 PM - 3:45 PM Room: Four Seasons 1
Purpose: Gamma analysis has well-known shortcomings for quality assurance of intensity-modulated radiation therapy (IMRT). The use of threshold-based passing criteria discards information like the spatial and intensity distributions of dose deviations. This work outlines a novel method of analyzing QA results, treating gamma distributions as images and extracting radiomic features to detect treatment delivery errors.
Methods: 23 patient IMRT treatments were measured in phantom with electronic portal imaging device dosimetry. Gamma distributions were generated for each beam by comparing the measured dose to three treatment planning system calculated plans with: 1) no errors, 2) random multi-leaf collimator (MLC) mispositioning errors, and 3) systematic MLC errors. These 558 gamma distributions were randomly divided into equal-sized training and validation datasets for model development. Logistic-regression models were trained using the radiomics features calculated for each distribution and tested on the validation set. Models were compared to traditional gamma analysis with area under the receiver operator characteristic curve (AUC).
Results: The AUC of the random MLC error model on the validation set was 0.76, compared to 0.51 for traditional gamma analysis. The AUC for the systematic MLC error model was 0.71 vs 0.66 for gamma analysis. The size zone radiomic features were most predictive of errors, but models combining multiple radiomic features outperformed any individual feature. Models for the two error types were demonstrated to be error specific, exhibiting AUC of approximately 0.5 (equivalent to random guessing) when applied to the error they were not intended to detect.
Conclusion: A method of analyzing gamma distributions using radiomic features has been developed that outperforms traditional gamma analysis. Furthermore, each model has been shown to detect only the specific error it was designed to detect. This methodology is a step towards improved quality assurance that provides a probable cause for QA failures when they do occur.
Contact Email: