Click here to

Session: MRI for Adaptive Treatment Planning and Delivery [Return to Session]

Auto-Detection of Structural Similarity Index Map From Daily MRI as An Indicator for MR-Guided Online Adaptation Replanning

Abdul K Parchur, Sara Lim, Haidy Nasief, Eenas Omari, Ying Zhang, X. Allen Li*, Medeical College of Wisconsin, Milwaukee, WI

Presentations

WE-C-TRACK 6-7 (Wednesday, 7/28/2021) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]

Purpose: MRI-guided adaptive radiation therapy(MRgART), particularly daily online adaptive replanning (OLAR), can substantially improve radiation therapy delivery, however, can be labor-intensive and time-consuming. Currently, whether to perform OLAR for a treatment fraction is determined subjectively. In this work, we develop a machine learning classifier based on structural similarity index map (SSIM) to quickly, automatically, and objectively determine whether OLAR is necessary for a daily MRI set.

Methods: The classifier was developed based on a total of 70 daily MRI sets acquired on a 1.5T MR-Linac during MRgART of 14 pancreatic cancer patients each treated with 5 fractions. For each daily MRI set, OLAR and reposition plans were created and the superior plan for the daily fraction was determined per clinical dose-volume criteria. SSIM was extracted for each daily MRI set with respective to its reference (e.g., planning) MRI in six regions enclosed by: 50%, 80%, 100%, 50–80%, 50–100%, and 80–100% isodose surfaces, respectively. A series of histogram features, i.e., mean, median, moment (2nd order), skewness's, kurtosis, and full width at half maximum (FWHM), along with a new parameter MMF ratio=Mean/(Moment*FWHM)*10^5, were extracted from each SSIM. A linear regression, t-SNE, and 5-fold cross-validation ensemble classifier was determined to predict if OLAR is necessary for a daily MRI set.

Results: Higher significant difference was observed for SSIM features extracted from the 50–100% isodose surface regions based on initial linear regression analysis (p = 0.007) and t-SNE projection between OLAR vs. reposition. The developed ensemble classifier in term of MMF ratio can predict 90% repostion and 95.9% OLAR with five-fold cross validation. The prediction accuracy was 94.2% and AUC was 0.97. A simplified decision tree with six-nodes was created.

Conclusion: An ensemble-based classifier using SSIM was developed, allowing to automatically and objectively determine if OLAR is necessary for a treatment fraction during MRgART.

Funding Support, Disclosures, and Conflict of Interest: Funding support from NIH R01CA247960

Handouts

    Keywords

    MRI, Radiation Therapy, Image-guided Therapy

    Taxonomy

    IM/TH- MRI in Radiation Therapy: MRI for treatment planning

    Contact Email

    Share: