MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Machine Learning with Radiomic Features to Detect the Types of Errors in IMRT Patient-Specific QA

M Sakai*1, H Koarai1, M Ueda1, S Shigeta2, H Nakano2, T Takizawa23, S Tanabe2, R Sasamoto1, H Aoyama4, S Utsunomiya1, (1) Niigata University Graduate School of Health Sciences, Niigata, Japan , (2) Niigata University Medical and Dental Hospital, Niigata, Japan , (3) Niigata Neurosurgical Hospital, Niigata, Japan, (4) Niigata University Graduate School of Medical and Dental Sciences, Niigata, Japan

Presentations

(Sunday, 7/14/2019) 5:00 PM - 6:00 PM

Room: 301

Purpose: To develop machine learning models to detect the types of error in patient-specific quality assurance (QA) for intensity modulated radiation therapy (IMRT), and evaluate their accuracy.

Methods: We created the IMRT treatment plans with an intentional error for 10 prostate and 10 head-and-neck cancer patients previously treated with IMRT. We created the four types of errors: single MLC misalignment in 2, 3 mm, positional error in measurement in 2 mm, changing MLC transmission factor (TF) and dosimetric leaf gap (DLG) of the treatment planning system Eclipse (Varian) in 5%, 10%, 15%, and 20%, respectively. We subtracted the error-free 2-D dose distributions from ones with errors and calculated the 8 histogram-based radiomic features with the subtracted dose distributions for each type of error. The Wilcoxon rank-sum test was used to find the radiomic features useful for distinguishing a type of error from the others. The machine learning models using a support vector machine (SVM) and a logistic regression were created with MATLAB (MathWorks) for each type of error and the accuracy of the models were evaluated by the area under the ROC curve (AUC) in 10-fold cross validation.

Results: The results of the Wilcoxon rank-sum test showed that almost all radiomic features were useful for distinguishing a type of error from the others, and the effect size of the best feature for DLG were less than those for the other types of error. The machine learning models showed high accuracy in detecting single MLC misalignment, positional error in IMRT QA measurement, and error in TF. However, the models showed moderate accuracy in detecting error in DLG.

Conclusion: The machine learning models using radiomic features showed high or moderate accuracy in detecting the types of error in patient-specific QA for IMRT.

Keywords

Intensity Modulation, Quality Assurance, Commissioning

Taxonomy

TH- External beam- photons: Quality Assurance - IMRT

Contact Email