Epicardial adipose tissue radiomics to reduce the residual cardiovascular risk: from methodological challenges to clinical practice

Abstract

Background: The reproducibility of radiomic features is essential to improve the applicability of radiomic research. Cardiac radiomics, in particular, adipose tissue radiomics, may prove useful in predicting cardiovascular disease development. This study aims to explore the robustness of computed tomography-derived-epicardial adipose tissue radiomic features and identify the most stable features among segmentation techniques and image preprocessing filters. Methods: We computed radiomic features extracted from the epicardial adipose tissue segmentations. Comparisons were made for segmentations performed manually segmentation by the same operator (intra-observer variability) and between two different experts (inter-observer variability). A comparison was also made comparing manual segmentation and semi-automatic segmentation using the Cardiac Risk Assessment tool (Siemens HealthCare). We also assessed the effects of image resampling and normalization on radiomic features' robustness. Dice coefficient was used to assess the similarity between segmentation masks. For each comparison, we reported the most stable radiomic feature family as determined by higher intra-class correlation. Results: Reproducibility of segmentations was excellent when performed by the same operator (DSC = 0.95), but smaller when performed by other operators (DSC = 0.92) and another method (DSC = 0.87 for the automatic segmentation). Regarding the reproducibility of radiomic features, intra-observer reproducibility of radiomic features is excellent (mean ICC = 0.94) but decreases significantly when radiomic features are compared for segmentations performed by different operators (mean ICC for inter-observer comparison = 0.84) and by different techniques (mean ICC for manual vs. automatic segmentation = 0.81). We reported the most stable radiomic feature families for each comparison. Similarly, the reproducibility of radiomic features varies with image alterations, such as image resampling and normalization (mean difference for ICC between the original image and transformed images = 0.02). Conclusion: The stability of radiomic features was mainly affected by segmentation techniques. Whilst the manual vs. automatic segmentation technique was statistically different from the inter-observer comparison for manual segmentations, the difference is not clinically relevant (mean difference for ICC = 0.03). A similar finding is reported for image transformations. Thus, an automatic tool could replace manual segmentation.