Click here to

Session: Multi-Disciplinary: Data Science/Radiomics [Return to Session]

Treatment Response Prediction for MRI-Guided Adaptive Radiation Therapy of Pancreatic Cancer Using Multiscale Wavelet-Based Delta-Radiomics

H Nasief*, W Hall, X Chen, E Paulson, B Erickson, X Li, Medical College of Wisconsin, Milwaukee, WI

Presentations

MO-IePD-TRACK 4-4 (Monday, 7/26/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: To investigate whether delta radiomics (change of radiomic features) based on daily MRI acquired during MR-guided adaptive radiation therapy (MRgART) of pancreatic cancer can predict treatment response.

Methods: Daily motion-average MRI sets derived from 4D-MRI acquired on a 1.5T MR-Linac during MRgART for 18 pancreatic cancer patients undergoing chemoradiation therapy (CRT) with 30-35 Gy in 5-fractions were acquired. CA19-9 response, pathological response for the cases treated with pre-operative CRT, and metastasis status were correlated with radiomic data. Wavelet decompositions were utilized to overcome challenges with inter- and intra-patient MRI intensity variations and to provide a robust way to extract radiomics data. On each daily MRI set, 438 wavelet-based delta-radiomics features (WDRF) were calculated from multiscale features extracted from different decomposition levels. Good or bad response groups were defined based on either pathological treatment response or distant metastasis free progression, and 50% reduction in CA19-9. Spearman correlations were applied to rule out redundant information. T-test, regression models, and linear mixed effect model were used to identify independent WDRFs with significant changes that correlate to response. Bayesian classifier with a leave-one-out cross validation method was used to determine WDRF combination with highest outcome predictions. The performance of the model to predict response was measured using AUC of ROC curve.

Results: Spearman correlation showed that 67 features from different wavelet decomposition were not highly correlated and can be used for the analysis. Of these features, 30 WDRFs showed trends correlating to response, with 8 WDRFs passing t-test and linear mixed effect (p<0.05) and demonstrated significant changes by 3rd or 4th fraction during treatment. Best performing model was a 3-feature combination that can predict response with an AUC of 0.94.

Conclusion: Machine learning identified wavelet delta-radiomics features from daily MRI in MRgART that can predict treatment response during chemoradiation therapy for pancreatic cancer.

ePosters

    Keywords

    MRI, Radiation Therapy, Texture Analysis

    Taxonomy

    IM- MRI : Radiomics

    Contact Email

    Share: