MRI Wavelet-Based Machine Learning Classifier to Automatically Determine OLAR Necessity During MRgART
H Nasief*, E Omari, A Parchur, Y Zhang, X Chen, E Paulson, W Hall, B Erickson, X Li, Medical College of Wisconsin, Milwaukee, WI
Presentations
TH-F-BRC-4 (Thursday, 7/14/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Ballroom C
Purpose: MRI-guided adaptive radiation therapy (MRgART), particularly daily online adaptive replanning (OLAR), can substantially improve RT delivery, however, is generally labor-intensive and time-consuming and is unnecessary for all treatment fractions. Being able to objectively and automatically determine the necessity of OLAR can avoid unnecessary effort. We have previously showed the feasibility of using a machine learning classifier based on wavelet textures to determine OLAR necessity. This study aims to improve and validate the classifier using more sophisticated approaches with larger datasets.
Methods: A total of 109 daily MRI datasets acquired from 22 panceratic cancer patients using a 1.5T MR-Linac were analyzed. For each daily MRI set, two plans were adapted from the reference plan following an OLAR or an adapt-to-position workflows, and 438 multiscale wavelet textures were extracted from the region enclosed by 50-100% reference isodose-surfaces from different decomposition levels. Spearman correlations were utilized to rule-out redundant features. Inter-class correlation (ICC), coefficient of variance (COV), t-test (p<0.05), self organized map (SOM) and maximum stable exteremal region (MSER) algorithm were used to determine significant features. Multivariate Bayesian classifier was used to develop prediction model and its performance was judged using the AUC of the ROC curve.
Results: Spearman correlation showed that 123 features were not redundant (r<0.9). Of these features, 82 showed high ICC for repositioning >0.6, 67 had a COV > 9% for OLAR. In addition, 38 passed the t-test with p value <0.05, of which 25 passed the SOM and 12 passed the MSER. Best performing model can predict OLAR necessity with a training CV-AUC of 0.98 and independent validation AUC of 0.97.
Conclusion: The newly developed machine learning classifier based on multiscale wavelet features from a daily MRI can automatically and objectively determine OLAR necessity with high accuracy, thus, can be implemented to avoid unnecessary effort during MRgART.
Keywords
Taxonomy
IM/TH- MRI in Radiation Therapy: Development (new technology and techniques)
Contact Email