Assessing Left Ventricle Radiomic Features Robustness by Segmentation Perturbations

Radiomics extracts quantitative features from images and uses them to develop diagnostic and prognostic prediction models. Radiomics is largely employed in oncology, but promising results have been recently achieved also in the cardiovascular field. Despite the emerging potential of radiomics, some issues still need to be investigated, such as assessing features robustness. This step represents a key point in the features selection process, ensuring radiomic model performance in the clinical setting. To address this challenge, several techniques have been employed but most of them result time-consuming or require additional resources. In the current study, an alternative approach involving perturbations of the region of interest (ROI), to assess features robustness, is proposed. The aim is to simulate variations in ROI segmentation that may occur in clinical practice. Perturbations were applied to left ventricles previously segmented by clinicians on cardiac computed tomography (CCT) images. Features extraction was performed and the intraclass correlation coefficient was computed for robustness assessment. Features stability was tested through ROI erosion, dilation, and contour randomization while ROI translation was employed to evaluate features discrimination capacity. Among the 107 radiomics features extracted, 79 were identified as robust and submitted to further features selection steps. Non-redundant and relevant features were, thus, employed as input of a logistic regression algorithm, trained to differentiate cardiac amyloidosis from hypertrophic cardiomyopathy reaching a mean accuracy of 81%.