1350907 - Field Strength Dependency and Reproducibility of Radiomic Texture Features in Paired CMR T1 Maps

Radiomic texture feature (RTF) analysis in cardiovascular magnetic resonance (CMR) imaging has the potential to quantify and characterize the various types of fibrosis when T1 mean values cannot[1]. RTFs have helped assess myocardial infarction and differentiate between fibrotic cardiomyopathies[2]. Adopting RTF analysis in CMR and other clinical practices has been limited as studies have reported high sensitivity to imaging parameters[3]–[4]. To leverage RTFs for more accurate diagnosis, multicenter scans must be pulled together. This will be made possible by understanding the effects of imaging parameters on RTFs in CMR. In this study, we examined the reproducibility of RTFs of T1 maps in paired CMR scans collected at 1.5T and 3T.

Methods:

We performed a retrospective analysis of paired CMR T1 maps from 3 short-axis slices collected from healthy volunteers (n=15)  on a 1.5T and a 3T scanner. Each T1 map was segmented in three contouring runs performed blindly on successive weeks. Then, 93 texture features were extracted per slice using pyradiomics from ten image preprocessing filters (wavelet, gradient, LBP, etc.) [5].

We computed the agreement between feature pairs as the Intraclass Correlation Coefficient (ICC) for each corresponding 930-dimensional feature vector pair (1.5T vs. 3T for a given slice location). We categorized RTFs based on the following criteria for reproducibility: ICC≥0.75 = good to excellent, 0.75 >ICC≥0.5 = moderate, ICC< 0.5 = poor) and reported RTF percentage per category. We also examined the effect of segmentation (the extent of overlap measured by the Dice coefficient) on the repeatability of texture features. Here, features extracted from two of the three contouring runs on each image were compared using Lin's concordance correlation coefficient (CCC). Repeatability was categorized as good = CCC≥0.95, moderate = 0.95 >CCC≥0.9, or poor  = CCC< 0.9.

Results:

Of the 930 RTFs, 36.8%±4.6% had moderate or better reproducibility, 23.7 % to 26.8% of features were reproducible across all runs, and 19.9% to 17.2% were reproducible across slices. 13.44% of all RTF were reproducible across all runs and slices (Table 1). 
A sub-analysis of run1 reveals that LBP-2D is the most reproducible image preprocessing filter while LGDM and GLRLM are the most reproducible feature classes (Figure 1). With a relatively high average overlap between corresponding contours in runs 1 and 2 (DC: 0.88-0.90), only 35.55% to 48.60% of RTFs were repeatable. Our subanalysis (Figure 2) shows that the most repeatable feature classes are GLDM and GLRLM, while the gradient Filter produces the most repeatable features across multiple runs.

Conclusion: Our study shows LV RTFs are sensitive to field strength and image segmentation. Only about 13% of features had moderate to good reproducibility ICC >0.75. We also learned from multiple runs that high precision, high accuracy segmentation methods could increase both repeatability and reproducibility.