Recently, some studies have applied radiomic analysis to CMR imaging in order to assess whether this tool could reveal myocardial phenotypic alterations in HCM patients18–21,31−36. For instance, Baeßler et al.31 have performed ML classification based on texture analysis of T1-weighted CMR images, and they have found, through logistic regression analysis, a model which discriminated between HCM patients and control subjects, with a sensitivity/specificity of 91%/93%. Alis et al.35 have reported that texture analysis of LGE images could classify HCM patients with and without ventricular tachyarrhythmia. They have evaluated the diagnostic performances of four ML-based methods (i.e., random forest, k-nearest neighbors, support vector machine, and naïve Bayes), reaching diagnostic accuracies (assessed by 10-fold cross-validation) ranging from 82.3–94.1%. Amano et al.18, using radiomic features from T2-weighted STIR images, have found that a GLRLM-GrayLevelNonUniformity value of 64.7 was the best discriminator between HCM patients and controls with an area under the receiver operating characteristic curve of 0.93. Notably, only a few pilot HCM studies19–21, 36 have applied radiomic analysis to T1 maps. Two studies19,21 have focused on the diagnostic capability of radiomic analysis to discriminate between HCM and HHD patients. In particular, Neisius et al.19, through a support vector machine classifier with 10-fold cross-validation, have identified six texture features (two GLRLM features and four local binary pattern features) which yielded a diagnostic accuracy of 80% in classifying the two groups of patients. Shi et al.21, using two combinations of texture features estimated from T1 maps, have obtained (by means of logistic regression models with 8-fold cross-validation) an area under the receiver operating characteristic curve of 0.97 and of 0.80 in differentiating patients with HCM/HHD from healthy subjects and HCM from HHD patients, respectively. Moreover, the study by Neisius et al.36 has shown that radiomic analysis applied on T1 maps could be used as a potential decision support tool prior to gadolinium administration. Indeed, they have employed a decision tree ensemble with 10-fold cross-validation and identified five texture features (one GLCM feature, one GLRLM feature, and three local binary pattern features) which were able to distinguish between LGE + and LGE-, with an area under the receiver operating characteristic curve of 0.74 assessed in a testing cohort. Wang et al.20 have assessed the genotypic-phenotypic association in HCM patients through radiomic and ML analyses of T1 maps. They have reported no significant differences between two subgroups of patients (MYH7 and MYBPC3) when using conventional CMR imaging analyses (i.e., cardiac function, LGE, and native T1). Nonetheless, radiomic analysis of T1 maps reached a discrimination accuracy of 85.5% when employing a linear support vector machine classifier on a test validation dataset.
Previous studies have investigated the effect of image preprocessing on radiomic features estimation as a function of various factors, which include imaging technique/modality and anatomical region40–43, 47,48, 51–55,72. Nonetheless, to the best of our knowledge, so far, no previous study has assessed the effect of both image preprocessing and image filtering on radiomic features estimated from CMR T1 and T2 maps. Thus, in a group of HCM patients, we performed a rather comprehensive analysis, which considered multiple elements such as resampling voxel size, image discretization, and image filters.
In radiomic studies, voxel size resampling is a recommended and employed preprocessing step when analyzing data with different acquisition protocols or from different scanners, which can result in different acquisition voxel sizes59. For both T1 and T2 maps, we found that textural features belonging to GLCM and NGTDM classes showed moderate relative variability (0.75 < ICC ≤ 0.9) when varying resampling voxel size, except for a limited number of radiomic features (Fig. 1). Conversely, the estimate of textural features of GLRLM, GLSZM, and GLDM classes showed a higher relative variability when varying resampling voxel size, with few exceptions (e.g., GrayLevelVariance from GLRLM, GLSZM, and GLDM, which presented ICC > 0.9 for both T1 and T2 maps). In general, radiomic features belonging to shape and first order classes were characterized by higher ICC values than textural features (Fig. 1), indicating that resampling voxel size has less impact on their estimate. In particular, shape radiomic features had ICC values > 0.9 (Elongation was the only exception, showing 0.75 < ICC ≤ 0.9) and median CV values within 0.8%-2.7% (Table 2) for both T1 and T2 maps. Most of first order radiomic features presented ICC values > 0.75, with only few features showing considerable (i.e., Energy from T1 and T2 maps, Kurtosis from T1 maps, and Skewness from T2 maps) or high (i.e., Kurtosis from T2 maps) relative variability. Furthermore, we revealed that even some radiomic features from T1 and T2 maps with ICC > 0.75 can have a non-negligible absolute variability in terms of CV (up to 30% or more, for single subjects) when varying resampling voxel size. In this regard, we found that T1 maps presented an appreciably higher number of radiomic features with ICC > 0.75 and median CV > 15% than T2 maps (Figs. 4 and 5b). For several textural features, a significant linear correlation between their estimates and resampling voxel sizes was observed (Supplementary Tables S1 and S3).
Discretization of image intensity is another important step before radiomic features estimation, which allows simplifying rather complex computational operations. Given that T1 and T2 represent quantitative physical properties of tissues, we chose to apply a fixed bin width approach59. Bin width ranges of variation (i.e., 3.6–6.4 ms and 0.49–0.57 ms for T1 and T2 maps, respectively) were selected using the same criterion (i.e., median number of quantization levels between 30 and 13043,65,66), allowing a comparison between the results of T1 and T2 maps. A remarkable difference between T1 and T2 maps in sensitivity of radiomic features estimation to discretization was found. In general, textural features estimated from T1 maps showed higher variability than their estimates from T2 maps (ICC > 0.75). Only a limited number of T1 maps textural features were characterized by ICC > 0.75 (Fig. 2) and some of them still had median CV > 15% (Fig. 5a). Overall, CV values were greater for T1 maps (up to 75%) than for T2 maps (less than 20%) (see Table 2). As expected, shape and intensity-based statistical features (i.e., all first order features but Entropy and Uniformity) yielded ICC = 1 and CV = 0%. Indeed, these radiomic features are estimated by PyRadiomics, according to the IBSI recommendation, prior to discretization59 and hence this preprocessing step does not affect their estimate. The repeated measures correlation analysis showed that all radiomic textural features (with few exceptions) have a significant linear correlation between their estimates and bin width (Supplementary Tables S2 and S4).
Digital image filters can be applied before radiomic features extraction to detect and emphasize tissue characteristics different from those usually obtained from original (i.e., unfiltered) images. In this regard, the IBSI has proposed a new reference manual, in order to define and standardize the implementation of image filters in radiomics software60. We observed a relevant sensitivity of both first order and textural features estimates to image filter, with ICC ≤ 0.5 (Fig. 3). This result, which was confirmed by the corresponding high CV values (Table 3), was not unexpected for first order radiomic features, given that image filters can strongly modify image intensity. However, the entity of this effect was not a priori predictable for textural radiomic features.
The revealed effect of resampling voxel size and bin width on the estimate of myocardial radiomic features from T1 and T2 mapping, albeit limited in many cases, sustains the importance of reporting/describing in detail these aspects in clinical and research studies. In this regard, while previous studies involving myocardial T1 mapping have described the radiomic features extraction process quite exhaustively19–21, 36, some information about image preprocessing was still missing. For instance, none of these studies have reported whether and how image discretization was performed (notably, Neisius et al.19 have specified the number of discretized intensity levels but only for the radiomic features belonging to the GLRLM class, while no discretization was indicated for the GLCM class). Furthermore, the studies conducted by Neisius et al.19 and Shi et al.21 have acquired images with different in-plane spatial resolution (i.e., 2.1 and 1.3 mm, respectively) and employed slightly different preprocessing steps (e.g., range re-segmentation and intensity outlier filtering, respectively59). This could partly explain the obtained different results in terms of discriminative radiomic features and accuracy.
So far, previous HCM radiomic studies19–21, 36 have exploited only T1 mapping, mainly because of its capability of revealing myocardial fibrosis. However, T2 maps are considered the gold standard for the evaluation of myocardial edema, which is a well-known negative prognostic factor in HCM15,16. Moreover, although T2 mapping is not per se capable of evaluating myocardial fibrosis, texture analyses might theoretically overcome this limitation, unveiling myocardial structural heterogeneity due to myofibrillar disarray and fibrosis18,27. Therefore, given also that, overall, radiomic features from T2 maps have proven to be characterized by a lower sensitivity to image preprocessing than radiomic features from T1 maps (especially when varying bin width), radiomics from T2 mapping might be exploited to obtain a complementary characterization of HCM, albeit this lower sensitivity does not necessarily imply higher discriminative or predictive power. In addition, it should be noted that, for both T1 and T2 mapping, the sensitivity of several radiomic features to resampling voxel size and bin width resulted rather independent of bin width and resampling voxel size, respectively. Moreover, only a limited number of radiomic features showed high relative variability (ICC ≤ 0.5) associated with different resampling voxel size or bin width. Some radiomic features with ICC > 0.75 still presented high variability (> 30%) in terms of CV for single subjects, suggesting that such radiomic data from different centers should be compared or pooled with caution.