Exhibit Hall | Forum 8
Purpose: Online adaptive replanning (OLAR), although can improve radiation therapy (RT), is generally labor-intensive and time-consuming and is not necessary for every fraction. We have previously reported that the need of MRI-guided OLAR can be indicated by measuring structure similarity index map (SSIM) on daily MRI. This study aims to improve the SSIM-based indication by training and testing a machine learning(ML) algorithm with updated datasets.
Methods: A total of 109 daily MRI sets acquired on a 1.5T MR-Linac during 5-fraction RT of 22 pancreatic cancer patients were analyzed. For each daily MRI set, a reposition plan (Adapt-to-Position, ATP) and an OLAR plan (Adapt-to-Shape, ATS) were developed. All fractions (daily MRIs) were divided into ATS and ATP groups. A fraction belonged to the ATS group if its ATS plan was superior to the ATP plan. SSIM maps were extracted from each daily MRI set, with respect to its reference MRI in the region enclosed by 50–100% isodose surfaces. Selected features SSIM were extracted and engineered. A 5-fold cross-validated ensemble bagged tree classifier was developed based on significant features determined using Pearson's correlation and t-test. Performance of the obtained classifier was judged using the AUC of the ROC curve.
Results: A single engineered SSIM feature, MMF ratio₌ln(Mean/(Moment×FWHM), combined with entropy can significantly discriminate between ATP and ATS groups. The obtained ML classifier using the two features can predict OLAR necessity with a cross-validated AUC of 0.93, compared to 0.90 if using MMF alone. Misclassification occurred primarily for the fractions that had minimal quality difference between their ATP and ATS plans.
Conclusion: The ML classifier based on MMF ratio and entropy from SSIM map was improved in determining when OLAR is necessary for a daily MRI set to a practical acceptable prediction accuracy. This classifier can be implemented in a MR-guide OLAR process.
MRI, Structure Analysis, Treatment Planning