Robustness of T1 and ECV mapping radiomics features: a between-session evaluation in young athletes

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Introduction Radiomics of cardiac MRI T1, T2 and extracellular volume (ECV) maps has the potential to add biomarkers that can aid in the detection and diagnosis of myocardial diseases. Recently, the feasibility of CMR mapping based radiomics to classify various myocardial diseases was demonstrated [1-6]. However, reproducibility studies have reported sensitivity of radiomics to acquisition parameters and processing steps involved concluding that only a limited number of features may be reproducible [7-8]. As CMR mapping guidelines recommend to use site-specific normal values [9], radiomics features derived likely also need careful site-specific evaluation to benchmark disease-related feature alterations. Purpose We aimed to assess the between-session reproducibility of radiomics features in a longitudinal dataset of MOLLI T1 and ECV maps obtained in young athletes at 1.5T. Materials and methods This study included data from 17 healthy subjects (15-20y; informed consent obtained) with data acquired two years apart [10] considered for this purpose as test-retest data since a prior standard analysis showed near identical average T1 (t1: 977±16 ms, t2: 982±20ms) and ECV (t1: 23.4±1.3%, t2: 23.4±1.5%). T1 mapping data was acquired on a 1.5T system (Ingenia, Philips) using MOLLI 5s(3s)3s. After motion correction and T1 and ECV map calculation [11], the left ventricular myocardium was manually delineated by two readers independently (3D Slicer [12]). In total 44 images (short and long axis) were included for each time point. The radiomics analysis resulted in 96 features per image (7 feature families, ‘shape’ excluded; no filters applied; Pyradiomics, [13]). The concordance correlation coefficient (CCC) was calculated to assess reproducibility, and features with CCCs ≥ 0.7 were considered reproducible. A coefficient of variation (CV) below 15% was considered low. Results Only a limited number of radiomics features had high CCC (T1: 6/96 ECV 0/96) or a low CV (T1: 32/96, ECV:30/96) in the between-session analysis. The inter-reader evaluation showed that the effect of the delineation on the results was limited. Features that were most robust in the between-session analysis were ‘first order (total)energy’ for T1 maps and ‘glcm_Autocorrelation’ for ECV maps (table 1). These results in young healthy subjects confirm previous test-retest reports [9-10]. Features with low CCC levels or high CV may however still be useful when discriminating between patient with myocardial diseases if the difference is larger than the confidence interval assessed via this reproducibility analysis. Conclusion In these healthy subjects, a strong variability in reproducibility of radiomics features of T1 and ECV mapping can be noted. Nonetheless, these variability measures are informative to determine features that are likely most robust when discriminating between health and disease and can be used as a benchmark towards radiomics AI-based diagnostic approaches. Top ranked features for either T1 or ECV