CAN PREPROCESSING AFFECT COLLINEARITY AND DIMENSIONALITY REDUCTION IN RADIOMICS? A PILOT STUDY IN HYPERTROPHIC CARDIOMYOPATHY MR T1 AND T2 MAPPING

Background: Frequently, radiomic analyses present issues caused by high dimensionality and multicollinearity. Thus, redundant radiomic features are usually removed through correlation analysis [1]. In a typical radiomic workflow, different image preprocessing steps are commonly performed, and previous studies have investigated how they influence features’ extraction, showing an appreciable sensitivity of radiomic features estimate to preprocessing [2]. However, no previous work has assessed the dependence of correlation-based dimensionality reduction on image preprocessing. Material and Methods: The effect of voxel size resampling, discretization, and filtering on correlation-based dimensionality reduction was investigated for radiomic features from cardiac T1 and T2 maps of 26 patients with hypertrophic cardiomyopathy. A dimensionality reduction of features based on either Pearson’s or Spearman’s correlation coefficient, followed by the computation of the stability index [3], was performed. Preliminary Results: With varying resampling voxel size and discretization bin width, correlation-based dimensionality reduction produced a slightly different percentage of remaining features with a relatively high stability index. These preprocessing steps affected more textural than shape or first-order features. The remaining features’ stability was relatively low for different filters. Notably, correlation-based dimensionality reduction was less sensitive to preprocessing when considering radiomic features from T2 compared with T1 maps. Our pilot findings further confirm the non-negligible effect of preprocessing in radiomic analyses, with the consequent need to consider it toward standardization of methods and when comparing results from different clinical studies.