Radiomic analysis of computed tomography myocardial texture: a new tool to identify subclinical cardiac amyloidosis

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Background Cardiac amyloidosis (CA) is an increasingly diagnosed disease sharing several phenotypical features with aortic stenosis (AS). Purpose As diagnosing the two diseases has crucial prognostic and therapeutic implications, this study aims to identify a set of stable and discriminative radiomic features derived from cardiac computed tomography to differentiate them. Methods Forty-two patients were included in the study. For each patient, 107 radiomics features were evaluated by means of geometrical transformations (translations) to the region of interests (ROIs) and intra class correlation coefficient (ICC) computation. A stratified 7-fold cross (k=7) validation was performed to split data into learning, validation and test set (Figure 1). Three features selection methods (Wilcoxon signed rank-based method and/or LASSO regression) and five machine learning classifiers. Results Ninety radiomic features satisfied robustness criteria and 10 were kept after feature selection. The best results were obtained using logistic regression classifier combined with Wilcoxon signed rank and LASSO regression, obtaining an accuracy of 95 ± 7% and sensitivity and specificity equal to 95 ± 12% in the test set. The Receiver Operating Curve (ROC) for this model on the test set showed an excellent area under the curve (AUC) of 0.94 (Figure 2). Conclusions The application of radiomics shows promising results in distinguishing left ventricle hypertrophy caused by CA from AS and might be used as a non-invasive tool able to support clinical decision making.