Quick Fire Session
SCMR 22nd Annual Scientific Sessions
Ulf Neisius, MD, PhD
Research Fellow
Beth Israel Deaconess Medical Center and Harvard Medical School
Hossam El-Rewaidy, MSc, BSc
Research Assistant
Beth Israel Deaconess Medical Center
Shiro Nakamori, MD
Research fellow
Beth Israel Deaconess Medical Center
Jennifer Rodriguez, BA
Clinical Research Coordinator
Beth Israel Deaconess Medical Center and Harvard Medical School
Warren Manning, MD
Professor
Beth Israel Deaconess Medical Center
Reza Nezafat, PhD
Associate Professor of Medicine
Department of Medicine (Cardiovascular Division) Beth Israel Deaconess Medical Center
Background: Several cardiac diseases are associated with left ventricular hypertrophy (LVH). The majority of these disease is characterized by an increase in global native T1 (1). However, there may be differences in spatial localization of T1 intensity levels that differentiate related etiologies. Radiomic texture analysis (TA) can quantify spatial distributions of pixel intensity levels (2) and as such could provide diagnostic information in patients with LVH. In this study, we sought to investigate if TA can provide diagnostic information comparing the most prevalent diseases with LVH, hypertensive heart disease (HHD) and hypertrophic cardiomyopathy (HCM).
Methods: We retrospectively identified 232 subjects (53 HHD, 108 HCM, 71 healthy controls) who consecutively were imaged using free-breathing slice-interleaved native T1 mapping at five short axis locations (3). HHD was defined as increased LVWT (≥12 mm) associated with the diagnosis of arterial hypertension in the absence of severe chronic kidney disease, LV cavity dilatation, and cardiac disease that could result in a similar magnitude of hypertrophy. HCM was diagnosed in accordance with contemporary guidelines (4,5). Myocardial regions of each T1 image were reshaped to rectangles of standard size and stacked to provide a single map with whole heart coverage (Figure 1). Four sets of texture descriptors were applied to capture spatially dependent and independent pixel statistics (Table 1). Each disease group was randomly split 4:1 (feature selection: independent validation). To test feasibility, interobserver agreement was tested in the independent cohort (n=30) by calculating the interclass correlation coefficient (ICC) for a 2-way mixed effects model with absolute agreement. To test for accuracy, six texture features (RLN 135°, SRHGE 135°, LBP 15, LBP 20, LBP 25, LBP 28; Table 1) were sequentially selected with best discriminatory capacity between HHD and HCM and tested in the feature selection cohort using Support Vector Machine (SVM) classifier. To test reproducibility, the same texture features were applied to the independent validation dataset.
Results:
Eighty-three out of 152 original features had excellent (ICC > 0.75) interobserver agreement. The selected texture features provided in combination the maximum diagnostic accuracy of 86.2% (c-statistic 0.820, CI 0.769-0.903). For the independent validation dataset, the diagnosis accuracy of the selected features remained high at 80.0% (c-statistic 0.89, CI 0.77-1.0). The selected features also differentiated between controls and HCM patients.
Conclusion: TA of native T1 images yields novel imaging markers that can discriminate between HHD and HCM patients.