Effect of Inter-User Segmentation Differences on Ischemic Stroke Radiomics from CTA and NCCT

Abstract

Radiomics is emerging as a promising tool for analyzing variations in signal intensities among different imaging modalities. For this technique, variations in medical images are quantified into a high dimensional space using automated data-characterization algorithms. Such radiomic features (RFs) are then used in advanced mathematical analyses of medical images for the prediction of treatment outcomes, disease prognoses, and pathology detection. In the field of acute ischemic stroke intervention and management, the procedural outcomes of mechanical thrombectomy have been associated with RF subsets according to several published studies. However, sensitivity of these features to key radiomics parameters in the determination of the RFs and the effect of inter-user segmentation accuracy remains unexplored but is an important consideration to the standardization of radiomics-based image biomarkers. In this study, we collected clots and corresponding non-contrast CT (NCCT) and CT angiography (CTA) images from 17 patients undergoing mechanical thrombectomy for large vessel stroke. Clot image regions were then segmented by 3 observers and radiomics feature were extracted for each. In total, 200 RFs were extracted. Sensitivity analysis was conducted across 4 binwidths (2, 4, 8, and 16) for all RFs, and a binwidth of 2 was found to maximum agreeability between users. Interrater reliability was calculated using the interclass correlation coefficient (ICC) for RFs from the 3 segmentations. Observers showed lower reliability in RFs for CTA compared to NCCT RFs. However, observers had good agreement with ICC>0.75 for 67 and 43 RFs from NCCT and CTA clot regions respectively, several of which have been shown to be predictive of thrombectomy outcomes in previous studies.