P4686Discrimination of fibrotic myocardium from healthy myocardium patients with aortic stenosis: a radiomics approach with machine learning models

Abstract

Introduction: Cardiac fibrosis, as assessed by late gadolinium enhancement on cardiac MR images, is currently not examined or quantified in MR studies as a truly objective measure. Radiomics strategies that enable advanced quantitative characterization of a tissue of interest using medical images from existing datasets have recently gained interest throughout the medical imaging field. Purpose: To perform a quantitative assessment of the myocardium for automated identification of fibrosis in the myocardium using late gadolinium enhancement images in a heterogeneous group of patients with aortic stenosis compared to controls. Methods: 186 subjects (average age: 65.9, 70% male) were included in this analysis, of which 57 (31%) had late gadolinium enhancement on cardiac MRI (CMR). The myocardium was manually segmented from the CMR images. This segmentation was used for radiomic feature extraction and calculation of case-specific features, related to myocardial thickness. Univariate analysis was done with the area under the curve (AUC). Feature selection was done with mRMR feature selection, before multivariate analysis with random forest (RF), linear support vector machine (SVM) and a generalized linear model (GLM). A temporal validation was done that included 142 subjects (75%) for training and 47 subjects for testing. Clinical features were implemented in a separate clinical model. These features include peak aortic jet velocity, high-sensitivity Troponin-I, ECG-strain, age and gender. An external validation was done with a second dataset (n=101, average age: 70.2, 50% male), and did not include the clinical model. Results: 5639 features were extracted from the images. 344 features showed a moderate to good correlation in univariate analysis (FDR q-value<0.05, 0.6≤AUC≤1) with the presence of LGE. The models trained with radiomic features (GLM: 0.91, RF: 0.94, SVM: 0.90), showed a significantly higher performance than the model trained with clinical features (AUC: 0.78) in temporal validation. External validation showed a good performance AUC of 0.70, 0.68 and 0.73 for respectively the GLM, RF and linear SVM with CMR on a different scanner. Conclusion: This study demonstrates the potential of computer-aided diagnosis of medical images using a radiomics approach. Our algorithms were able to automatically discriminate between subjects with and without enhancement in a group of patients with relatively subtle fibrosis. Multivariate analysis of LGE could lead to a better classification of LGE than univariate analysis of features. Future studies are needed to demonstrate whether radiomics guided fibrosis analysis might be generalizable to other cardiac pathologies involving more patients but also more distinct scars.