ABSTRACT
With increasing use of MR imaging for the prostate, a variety of radiomic approaches have been proposed for machine learning-based disease diagnosis, prognosis, and response prediction. Popular radiomic features quantify localized intensity and filter responses to capture variations in signal intensity values within a region of interest. However, T2w MR intensity values do not exhibit tissue-specific meaning between patients, scanners, and institutions, likely impacting the robustness of MRI-based radiomic features. We present a case study on correcting inherent intensity drift on the original T2w signal intensity values as well as associated radiomic features, in a multi-site setting. This study utilized 147 T2-weighted prostate MRI datasets curated from across 4 different sites. 131 radiomic texture features were extracted from within expert-annotated tumor and non-tumor regions. A unique measure of feature instability was utilized to quantify cross-site reproducibility, between the pre- and post-standardization MRI datasets. Standardization resulted in tumor and non-tumor region intensities becoming reproducible in over 99% of all cross-site comparisons, indicating an improved numeric consistency in standardized T2w intensities across sites. Only 8% of all 131 radiomic features exhibited worse cross-site reproducibility after standardization, which were localized complex co-occurrence and macro-scale wavelet features within non-tumor regions alone. Intensity standardization may be a critical post-processing step for more reproducible radiomic characterization of prostate T2w MRI data, across multiple sites.
