(Tuesday, 7/16/2019) 9:30 AM - 10:00 AM
Room: Exhibit Hall | Forum 9
Purpose: The purpose of this work is to investigate the repeatability of radiomics features extracted from daily CBCT images for prostate cancer patients.
Methods: Two prostate patients who had daily CBCT images reconstructed with an iterative algorithm (iCBCT) were included in this study. Texture features were calculated from gray-level run length matrix (GLRLM), gray-level co-occurrence matrix (GLCOM), neighborhood gray-level difference matrix (NGTDM), and gray-level size zone matrix (GLSZM). For each patient, forty-two texture features were extracted from the gross tumor volumes (GTVs) defined on forty iCBCT images. First, image preprocessing was done using three quantization algorithms (Uniform, Equal, and Lloyd), two grey level normalizations (Collewet and Normal), and three bit depth level quantizations. Next, a linear fit was performed on each texture feature grouped into weekly data sets to mitigate the influence of texture changes during treatment. The root mean square error (RMSE) was derived from residuals of the linear regression for each texture feature as an estimate of the repeatability. RMSE was then normalized by the mean to make the texture features comparable. Texture features that correlate highly with prostate volume were removed.
Results: All image pre-processing methods provided highly repeatable texture features (median RMSE < 8%). For the most repeatable image processing method, thirty-six features have median RSME < 10%, and eight features have less than 1% RSME. Moreover, twenty-five features have linear fit R squared values > 0.5. Previous studies done with lung cancer CBCT also found many of these features to be highly repeatable (Fave et al. 2015).
Conclusion: Many of the radiomics features extracted from daily iCBCT images for prostate cancer patients were highly repeatable (median RSME<10%) and are thus robust for radiomics analysis.
Funding Support, Disclosures, and Conflict of Interest: This project was funded in part by a grant from the Varian Medical Systems, Inc.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
