Purpose: Magnetic resonance-guided radiotherapy (MRgRT) provides radiation oncology departments a modality that acquires high contrast images ideal for anatomical visualization for target localization, motion tracking, and gating. The purpose of this work is to explore the predictive potential of delta radiomics texture features (i.e., changes in radiomic texture features over time) extracted from daily set up images of pancreas patients treated with MRgRT.
Methods: The gross tumor volume (GTV) was delineated on daily set up images of 27 pancreas patients treated with 33-50 Gy in 5 fractions. Gray level intensities for each GTV were quantized to 64 levels using the histogram equalization method and 39 second order texture features were extracted. Two sets of delta radiomics features were calculated: absolute change (DF(ABS)) and normalized change (DF(NORM)). DFABS is the absolute difference between fraction 1 and fractions 2-5. DF(NORM) features are DF(ABS) normalized by fraction 1 feature values. Delta radiomics features were ranked by predictive importance using the Gini Index obtained during training of a random forest (RF) model. The RF models constructed using DF(ABS) and DF(NORM) from fraction 1 and 2 produced in the highest area under (AUC) the receiver operating characteristic curve and selected the same two DF(ABS) and DF(NORM) features as most important. Gray level co-occurrence matrix-based sum average and gray level size zone matrix-based large zone low gray level emphasis were used to construct RF models for evaluation of predictive performance using leave-one-out (LOO) analysis to obtain AUC.
Results: The RF using DF(ABS) features achieved an AUC, with a 95% confidence interval, of 0.77(0.58–0.96) and the RF with DF(NORM) achieved an AUC = 0.78(0.59–0.97).
Conclusion: Delta radiomics texture features calculated from fraction 1 and 2 set up images during SBRT may provide valuable prediction of treatment response with time to modify the treatment and improve outcomes.
Not Applicable / None Entered.