Oral Abstract Session
SCMR 22nd Annual Scientific Sessions
Hossam El-Rewaidy, MSc, BSc
Research Assistant
Beth Israel Deaconess Medical Center
Ulf Neisius, MD, PhD
Research Fellow
Beth Israel Deaconess Medical Center and Harvard Medical School
Shiro Nakamori, MD
Research fellow
Beth Israel Deaconess Medical Center
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 studies have reported elevated myocardial global native T1 values in patients with hypertrophic cardiomyopathy (HCM) and dilated cardiomyopathy (DCM) [1–4]. Despite statistically significant differences in global native T1 among these cohorts and healthy subjects, a large overlap in measurements limits clinical utility of native T1 mapping for an individual patient (Figure 1.a) [5,6]. We therefore hypothesize that a radiomic approach to extract novel texture features from native T1 maps may provide incremental value for discriminating between different cardiomyopathies such as HCM, DCM, and healthy controls.
Methods: Myocardial T1 mapping images were acquired using 1.5T Philips Achieva system at five slices, from base to apex, for 321 subjects: 65 healthy, 116 HCM and 140 DCM patients using slice-interleaved T1 mapping sequence [7] . Left ventricular myocardium was manually delineated in each slice and the extracted myocardium were reshaped and stacked in rectangular shape to allow simultaneous multi-slice analysis (Figure 1). Four groups of texture descriptors were utilized to extract 152 features that capture spatially dependent and independent pixel statistics, as well as locally-repeated patterns from the T1 maps (Table 1). Seven features were selected using sequential forward selection technique as the most discriminative set of features among the three cohorts (Table 1). The classification accuracy, generalizability and reproducibility were tested on feature selection and independent validation datasets (3:1) using support vector machine (SVM) classifier. The selected features were combined in one index (Tx) using linear regression model.
Results:
The classification accuracy of the three cohorts: healthy, HCM, and DCM significantly improved from 51.1% (healthy, c=0.78, 95%-CI, 0.71 - 0.87; HCM, c=0.51, 95%-CI, 0.44 - 0.59; DCM, c=0.74, 95%-CI, 0.68 - 0.81) for global native T1 values to 85.2% (healthy, c=0.93, 95%-CI, 0.89 - 0.98; HCM, c=0.93, 95%-CI, 0.9 - 0.97; DCM, c=0.93, 95%-CI, 0.90 - 0.97) for the selected texture features using SVM on feature selection cohort (P0.9 for the important features).
Conclusion:
Texture features derived from native myocardial T1 mapping can provide incremental information that improves the discrimination among HCM, DCM patients, and healthy controls than the global native T1 values.