Texture Analysis of Non-Contrast T2-Weighted Cardiovascular Magnetic Resonance Imaging to Discriminate Between Cardiac Amyloidosis and Hypertrophic Cardiomyopathy

Shan Huang West China Hospital of Sichuan University Ke Shi West China Hospital of Sichuan University Yi Zhang West China Hospital of Sichuan University Wei-feng Yan West China Hospital of Sichuan University Ying-kun Guo West China Second University Hospital of Sichuan University Yuan Li West China Hospital of Sichuan University Zhi-gang Yang (  yangzg666@163.com ) West China Hospital of Sichuan University


Introduction
Cardiac amyloidosis (CA) is a progressive and in ltrative cardiomyopathy that frequently leads to heart failure and cardiac death [1]. CA is characterized by left ventricular hypertrophy (LVH) and preserved left ventricular ejection fraction (LVEF) [2]. In certain clinical scenarios, it can still be a challenge to differentiate CA from hypertrophic cardiomyopathy (HCM) [3]. However, it is of utmost importance to diagnose CA in an early stage to guide the management and improve the prognosis of these patients [4].
At present, late gadolinium enhancement (LGE) of cardiovascular magnetic resonance imaging (CMR) has remarkable diagnostic performance in diagnosing CA and HCM in clinical settings [5]. However, CA patients frequently suffer from impaired renal function, either due to amyloid deposition in the kidneys or reduced cardiac output from heart failure (HF). And the use of a gadolinium contrast agent is known to have the potential to cause renal sclerosis [6]. Thus, contrast agents are contraindicated in quite a few CA patients. In addition, both CA and HCM may not present with typical LGE patterns. A diagnosis based on LGE cannot be very de nitive. Thus, using a non-contrast method for diagnosing CA and avoiding contrast-induced renal injury seems extremely important.
T2-weighted imaging (T2WI) is widely used in the clinic to identify pathological lesions in many organs, such as the brain, liver and muscles. Abnormal T2 signal intensity is correlated with edema, cellular proliferation, and vessel densities [7,8]. In the heart, T2-weighted sequences are often used to identify myocardial edema [9,10]. However, due to the low contrast between normal and edematous myocardium, it is often a challenge to determine the presence and extent of the lesion [11].
Texture analysis (TA) quanti es the texture of an image based on spatial distributions of pixel signal intensity and relationships of values between neighboring pixels [12]. TA has the capability to overcome the limitations of traditional subjective visual interpretation of images and recognize lesions that are imperceptible to the human eye. CMR-based TA has been reported to have great performance for several clinical applications. A previous study found that TA of T2 mapping presented high sensitivity and speci city for the diagnosis of acute infarct-like myocarditis [13].
To the best of our knowledge, this is the rst study to use texture features derived from T2WI to discriminate CA from HCM. We retrospectively included 100 CA and 217 HCM patients, aiming to elucidate the value of TA in detecting and differentiating myocardial tissue alterations on non-contrast T2-weighted CMR images of patients with CA and HCM.

Study population
We retrospectively included 317 subjects (CA: n = 100, HCM: n = 217) who underwent CMR scans between January 2016 and June 2020. This study was approved by the institutional review board and in accordance with the Declaration of Helsinki. Informed consent was obtained from all participants.
CA was diagnosed on the basis of positive myocardial biopsy or extracardiac biopsy in conjunction with a mean LV wall thickness (septum and posterior wall) ≥ 12 mm on CMR [14]. The diagnosis of HCM was based on the presence of unexplained LV wall thickness ≥ 15 mm or ≥ 13 mm with a family history of HCM or apical hypertrophy in the absence of other conditions capable of producing a similar degree of hypertrophy [15]. HCM patients with previous septal ablation or myectomy were excluded.
The exclusion criteria for all subjects were valvular heart disease (greater than mild stenosis or greater than moderate regurgitation), signi cant coronary artery disease, and other con rmed systemic diseases. Images of poor quality were excluded. Hypertensive patients with concentric hypertrophy were also excluded.
Cardiovascular magnetic resonance CMR imaging was performed in accordance to a standard protocol, as previously published by our group [16]. All CMR images were performed using a 3.0T scanner (MAGNETOM Skyra, Siemens Healthcare, Erlangen, Germany) with a 30-channel phased-array receiver coil. Cine images were obtained using a balanced steady-state free-precession sequence in consecutive short-axis views covering the entire LV (from the level of the mitral valve annulus to the LV apex) with the following parameters:

Imaging analysis
Conventional CMR parameters including LV end-diastolic volume (EDV), LV end-systolic volume (ESV), LVEF, LV mass, maximal LV wall thickness (MWT) and LGE extent were calculated using commercially available software (CVI 42 ; Circle Cardiovascular Imaging, Inc., Calgary, Canada). The extent of LGE was quanti ed by 5 standard deviations (SDs) above the signal intensity of remote normal myocardium [17].
Texture analysis was performed on non-contrast T2-weighted CMR scans using 3D Slicer based on the Pyradiomics library [18]. The regions of interest (ROIs) were manually delineated in the basal septum of the left ventricle by a radiologist with 4 years of experience in cardiovascular imaging who was blinded to the patients' information. ROI delineation was repeated twice in a subset of 30 randomly selected patients by the same radiologist for intraobserver analysis and by another radiologist with 15 years of experience for interobserver analysis.

Feature extraction and selection
A total of 837 features were extracted from T2-weighted images during the process of image ltering and feature extraction. Stepwise feature selection and dimension reduction were performed due to the high number of texture features. First, the intraclass correlation coe cient (ICC) was calculated to assess the intra-and interobserver reproducibility of the selected features. Features with ICCs < 0.75 were excluded, ICCs ranging from 0.75 to 1 considered "excellent" [19]. The Boruta algorithm [20], corrplot by carret [21] and the least absolute shrinkage and selection operator (LASSO) with 10-fold cross-validation [22] were performed in a stepwise manner for dimension reduction. Furthermore, both the CA and HCM groups were randomly divided into a training dataset and a testing dataset (7:3). To select variables that allow for the discrimination of myocardial tissue alterations in HCM and CA patients, the classi cation tree model [23], a commonly used machine learning algorithm, was employed to calculate discriminative performance in the training cohort and validated in the testing cohort.
After the radiomics signature was established, the diagnostic e ciency and accuracy of the model was validated in patients with similar hypertrophy matched by LV mass index as well as age, sex and maximum wall thickness (MWT).

Statistics analysis
Statistical analysis during the construction of the radiomics signature was performed in R (version 4.0.1; R Foundation for Statistical Computing, Vienna, Austria) [24] with RStudio (version 1.3.959; RStudio, Boston, Mass) [25]. The R packages used for statistical analyses are described in the supplementary materials. Other statistical analyses were conducted with SPSS (Version 19; IBM, Armonk, NY). The normality of the data distribution was determined using the Kolmogorov-Smirnov test. Continuous data are expressed as the means ± SDs or medians with interquartile ranges. The t-test or the Mann-Whitney Utest was conducted, as appropriate. The diagnostic accuracy of the optimal radiomic parameters was evaluated by the area under the curve (AUC) from receiver operating characteristic (ROC) analyses. The diagnostic sensitivity, speci city, positive predictive value (PPV), negative predictive value (NPV) and accuracy were also calculated. ROC curves were compared using DeLong's test. P < 0.05 was considered statistically signi cant.
The patients with CA were older and had higher NYHA functional classes than the HCM patients. The HCM group had a higher body mass index. Moreover, more hypertensive patients were found in the HCM group than in the CA group. The NT-proBNP and troponin T levels were markedly higher in the CA group than in the HCM group (log NT-proBNP: 8.5 ± 1.1 vs. 7.0 ± 1.0; log troponin T: 4.6 ± 0.7 vs. 3.0 ± 0.9; p < 0.001 for both). Multistep texture feature selection and dimension reduction The multistep texture feature selection and dimension reduction process is described in Fig. 1. In total, we extracted 837 texture features from 6 feature groups, including rst-order features, gray level cooccurrence matrix (GLCM), gray level dependence matrix (GLDM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), and neighboring gray tone dependence matrix (NGTDM). De nitions and interpretations of the texture features are presented in the supplementary materials. Based on the ICCs for intra-and interobserver reproducibility, 614 features with ICCs < 0.75 were rst excluded. The Boruta algorithm, based on the random forest machine learning algorithm, further con rmed 51 important features. Then, a correlation matrix was calculated for these features to eliminate collinearity at the level of |rho|≥0.9, leading to a further reduction from 51 to 20 features. These 20 features were nally fed into the LASSO algorithm, resulting in the 7 most important and independent texture features used for model tting.

Discussion
Texture analysis is a postprocessing method to identify subtle tissue alterations and can be applied to standard and routinely acquired clinical CMR sequences. Our study demonstrated that TA was feasible and reproducible for detecting myocardial tissue alterations on non-contrast T2-weighted images. The radiomics model constructed with texture features derived from this sequence had a great performance in differentiating between CA and HCM patients and was comparable to LGE. Thus, TA of T2WI can help eliminate the need for contrast agent administration in these patients. Radiomics utilizes TA to comprehensively and elaboratively analyze the spatial distributions of pixel gray levels in images, which further derives substantial quantitative texture features characterizing the underlying tissue texture [26]. These texture features, in combination with robust mathematical models, could represent reliable diagnostic tools [27].
Our study indicated that the optimal combination of texture features had an accuracy of 86% for differentiating between CA and HCM. HCM often shows extremely and heterogeneously increased wall thickness. We validated the diagnostic capacity of the TA model in patients with similar hypertrophy matched by LV mass index, age, sex, and MWT. This subgroup analysis showed that the TA model still had a great discriminative capacity with high sensitivity and speci city for CA and HCM. Delong's tests of the AUCs showed that the diagnostic capacity of the TA model was comparable to that of LGE.
LGE is the most popular and useful CMR sequence for the diagnosis and differentiation of CA and HCM in the clinical setting. However, the required administration of a gadolinium agent has limited some patients from undergoing this examination. There are many researchers seeking to develop novel and gadolinium-free techniques for the characterization of myocardial diseases. For example, Neisius et al. utilized texture features of the myocardium through native T1 mapping to discriminate between HCM and hypertensive heart disease[28]. Baessler et al. indicated that TA of non-contrast cine images allowed for the diagnosis of subacute and chronic ischemic scars with high accuracy [29]. Our study made an effort to apply TA to non-contrast T2 images and achieved a comparable result to LGE in discriminating CA from HCM.
T2WI is a routinely applied and widely accessible sequence for CMR in most institutions. A high signal intensity on T2WI is indicated to be correlated with myocardial edema and considered helpful for the diagnosis of myocarditis and acute myocardial infarction [30]. Hen et al. had previously reported that a high T2 signal was an independent predictor of life-threatening arrhythmic events in HCM patients [31]. Kotecha et al. demonstrated that myocardial edema was present in CA by histology and CMR T2 mapping. T2 signal was an independent predictor of death in light-chain amyloidosis (AL), suggesting that in addition to amyloidosis in ltration, myocardial edema possibly caused by amyloidosis bril toxicity would be an additional mechanism contributing to mortality [32]. Our study demonstrated that texture features derived from T2WI could re ect tissue alterations of the myocardium in CA and HCM.
The most important texture feature for discriminating between CA and HCM on T2WI in our study was GLRLM-SRE LHL. A greater value of GLRLM-SRE represents ner textural textures, while a greater value of GLRLM-LRE indicates coarser structural textures. Both features are derived from the run-length feature matrix, which describes the number of gray level runs of various lengths. A gray level run is de ned as a group of pixels that have the same gray-level value in a given direction. The length is the number of pixels within the run. The run-length matrix elements describe the number of runs of a speci c gray level value, and a particular run length can be observed in the ROI [33]. Thus, the two features GLRLM-SRE and GLRLM-LRE can serve as a measure of tissue homogeneity. In our study, the CA group showed lower GLRLM-LRE and higher GLRLM-SRE values than the HCM group, suggesting that CA has a ner textural texture on T2WI. The possible reason for this might be correlated with the more pronounced myocardial edema present in CA [32]. The coarse texture of HCM might re ect tissue inhomogeneity, such as myocardial disarray and brosis. Thus, TA seems to be a useful quantitative tool for the identi cation of tissue alterations.

Conclusion
In conclusion, our study demonstrated that texture analysis based on non-contrast T2-weighted images can feasibly differentiate CA from HCM, even in patients with similar hypertrophy. The selected nal texture features could achieve a comparable diagnostic capacity to LGE. TA could help eliminate the use of contrast agents and additional sequences to discriminate between these two groups of patients in clinical settings. Availability of data and materials The datasets analyzed in the current study are available from the corresponding author on reasonable request.

Limitations
This study still had several limitations. First, this was a single-center study. Second, since this was a retrospective study, texture analysis of novel sequences, such as T1 mapping and T2 mapping, could not be performed at this time. However, we will conduct prospective studies with these novel imaging techniques in the future. In addition, the sub-analysis of patients with similar hypertrophy was limited by the small sample size. Further well-designed prospective studies are necessary to determine the utility of these TA parameters for a more general application.

Funding
This work was supported by 1·3·5 project for disciplines of excellence, West China Hospital, Sichuan University (ZYGD18013).