Images acquired during traditional CMR (including short and long axis, T1/T2 weighted, initial perfusion pass, and late gadolinium enhancing sequences) are selected for radiomics analysis. Recent studies tend to preferentially select late gadolinium enhancement and cine sequences images [15, 16], while the use of the remaining sequences is gradually increasing [17, 18]. Image preprocessing methods include image resampling and data normalization. Image resampling is to change the image voxel resolution to 1 mm${}^3$ and the image size to 256${}^3$ voxels. It is used to eliminate the influence of the heterogeneity of MRI scanner models (such as image resolution, slice thickness, and patient position) [19]. Data standardization refers to unifying the indicators of different dimensional units into the same order of magnitude, eliminating the influence of different dimensional units on the results. For example, the original image maximum pixel matrix value is 256, and 1 is the minimum; after processing, the image value is comprised between an interval of 0-1 [20].

Image segmentation is used to outline ROIs on images for further feature selection and model building, and segmentation methods could be divided into three types consisting of manual, semi-automatic, and fully automatic segmentation algorithms [21]. Manual segmentation is time-costing but could promise accurate and reliable outcomes. The automatic algorithm could reduce the contour work for doctors dramatically with high operation repeatability. However, the automatic segmentation algorithm is not as popularized as the manual algorithm, and the accuracy of automatic segmentation is still to be improved. In CMR, the endocardium and epicardium of the left ventricle and endocardium of the right ventricle are often delineated on short-axis images, which define the boundaries of the left ventricular myocardium and the two ventricular blood pools. After the selection of ROIs, data can be extracted; data changes as ROIs change. Therefore, most studies are carried out based on the automatize selection and manually adjusting of ROIs, with better chances of repeatability [21] (Fig. 2).

Fig. 2.

On LGE images, we obtained the segmentation results of scar and myocardium of four different patients. The epicardial boundary was blue, the endocardial boundary was red and the scar was yellow. There were two segmentation methods: manual segmentation (previous line) and automatic segmentation (next row). Automatic image segmentation had better repeatability. Cited in [22].

Feature extraction is a crucial step for radiomics research. Routine images are converted into plenty of quantitative data. It is these quantitative features that combine image information with clinical practice. Radiomics features represent the geometry of ROIs, the distribution and the spatial correlation of signal strength of voxels in ROIs. These features can be divided into four categories:

a) first order features: derived from the ventricular contour to define voxel intensity distribution in ROIs (the principle is shown later), such as energy, kurtosis, skewness, and entropy (Table 2). In 1948, Shannon et al. [23] initially proposed that entropy could describe the quantitative information relative to the communication system. In CMR protocol, entropy is used to evaluate the heterogeneity of pixel signal strength composed by a mixture of fibrotic tissue and normal myocardium. Based on this application, many studies have reported the implication of entropy in the diagnosis and prognosis of common heart disease [24].

Table 2.According to the signal strength matrix, the first order features based on histogram can be constructed by simple arithmetic model. The abscissa is the intensity value of voxel signal, and the ordinate is the frequency of occurrence. The statistical data obtained from the histogram are summarized, and the specific first order features are listed.

Intensity
Mean	Energy	Kurtosis
Range	Entropy	Maximum
Median	Minimum	Skewness
Variance	Uniformity	10th percentile
Standard Deviation	Interquartile Range	90th percentile
Root Mean Squared (RMS)	Mean Absolute Deviation (MAD)	Robust Mean Absolute Deviation (rMAD)

b) shape features: determining the size and shape of ROIs in two or three dimensions. They include not only the conventional features like length, volume, and the surface of ROIs, but also features describing the overall shape of ROIs, such as compactness, sphericity, and elongation (Table 3). Shape features are independent of the gray intensity of the image, so they are only acquired through the original image [25].

Table 3.The left and right ventricular blood pools (red, yellow) and left ventricular myocardium (green) are segmented. The shape features of radiomics can be deduced from the contour of left and right ventricle, and the specific shape features can be listed.

	2D	3D
	Mesh Surface	Mesh Volume
	Pixel Surface	Voxel Volume
	Perimeter	Surface Area
	Perimeter to Surface ratio	Surface Area to Volume ratio
	Sphericity	Sphericity
	Spherical Disproportion	Compactness
	Maximum 2D diameter	Spherical Disproportion
	Major Axis Length	Maximum 3D diameter
	Minor Axis Length	Major Axis Length
	Elongation	Minor Axis Length
		Least Axis Length
		Elongation
		Flatness

c) texture features: describing the distribution and the interaction of voxels in ROIs and using digital quantization eigenvalues to describe the image information (Table 4). Quantitative texture features can be relatively objective parameters for diagnosis or prognosis prediction in the clinic, compared with conventional qualitative judgment based on a medical picture. The first step of texture feature analysis is to build the signal strength matrix. Each voxel within the ROIs is assigned to a value accordingly to the different intensity, and then a signal strength form is constructed to form a signal strength matrix. Next, we can directly calculate the signal histogram of the signal strength matrix, which forms the first order features mentioned above. However, the first order features represent the overall signal distribution within ROIs. More complex algorithms are needed to describe the correlation between each adjacent voxel. Through the complex calculation and reconstruction of the original signal strength matrix, a new matrix can be formed, such as gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), and gray level size zone matrix (GLSZM). GLCM is basically constructed according to the frequency of different signal strength pairs in the original signal strength matrix. Corresponding measures such as contrast (local variation) and entropy (confusion) can then be calculated; these features reflect the gray level changes of signal strength pairs and the degree of confusion in ROIs [26]. GLRLM is constructed by the number of consecutive runs along the specified direction in the signal strength matrix. It can evaluate the spatial correlation of any number of voxels (not just pairs). For example, run entropy reflects the degree of confusion in the gray distribution of the stroke length (higher value is more chaotic); high gray level run emphasis measures the distribution of higher gray values (higher value results in a greater concentration of higher gray values). GLSZM is reconstructed by quantifying the number of voxels within the same signal strength in the image, such as gray level non uniformity and gray level variance. Many scholars have confirmed the value of Gray level non uniformity (GLN) in the diagnosis and differential diagnosis of myocarditis [27].

d) high order feature: filter or high order image parameters are added to first order or texture features, such as Wavelet, Laplace of Gaussian, and autoregressive model. There may be redundant features or features not closely related to the results [28].

The number of extracted features can reach thousands. So feature selection (such as principal component analysis or clustering) will identify and delete redundant and unstable image features [29]. It will avoid the overfitting of the model.

The purpose of radiomics is to develop a model for diagnosis, prognosis, and treatment evaluation of medical diseases. A radiomics model is developed using specific classifiers for model training. There are kinds of classifiers, including random forest (RF), support vector machine (SVM), naive Bayes, and k-nearest neighbor method (K-NN) [30]. RF is composed of many decision trees; each decision tree comprises random samples and random features. The final judgment will be made through the summary of all the decision trees’ results [31]. Its advantage is that there will be no overfitting phenomenon if there are enough decision trees, while too many trees will slow down the algorithm speed. SVM uses a separation hyperplane to partition the training set correctly. It can tolerate some wrong samples in the boundaries, thus improving the robustness and generalization of the model [32]. Naive Bayes is based on probability theory and assumes that each feature is independent and suitable for small sample classification. The principle of K-NN is to take the k training samples nearest to the test samples. The category with the most frequent occurrence in these k samples is the classification result. The disadvantage of K-NN is that it has higher requirements for the balance of sample categories.

All these classifiers can be used for modeling, but the classifier which shows the best performance is preferable. Training sets, verification sets and testing sets are very important in the process of radiomics analysis. Generally, the data is distributed into several groups: training sets used to fit the model, and verification sets used to adjust parameters of the model but also for preliminary evaluation of the model, while the testing sets used to evaluate the final generalization effect of the model. For instance, K-NN was used to identify whether HCM patients occur ventricular arrhythmia (the Area Under Curve (AUC) = 0.92), but the research lacked testing sets. Hence, it was impossible to evaluate the generalization effect of the model, and further research was needed [30]. The model is representative only if the sample size is large enough; experts propose k-fold cross-validation for the defect that there are insufficient clinical research samples [33]. The training sets are split into k samples, and one of them is randomly selected as the verification sample, and the remaining k-1 is the training sample. The process is repeated k times to obtain the best performance. Usually, k is acceptable for sampling when its value is between 5 and 10. If k is less than 5, the underfitting problem could occur [34].

MI is one of the most serious cardiovascular diseases threatening human health, and the main pathophysiological mechanism is myocardial ischemia and necrosis due to coronary artery stenosis and occlusion [35]. Patients have the risk of ventricular arrhythmia and poor left ventricular remodeling [36]. Baessler et al. [37] studied the efficacy of texture features in diagnosing subacute and chronic myocardial infarction on CMR cine images. The results showed that the five texture features Teta1, Perc.01, Variance, Wavenhh.s-3, and S (5,5) SumEntrp had statistical significance in differentiating infarcted myocardium from normal myocardium. In the multiple Logistic regression models, diagnostic accuracy of MI was excellent (AUC = 0.93). Gibbs et al. [38] explored the relationship between the features based on the Laplacian of a Gaussian filter and scar heterogeneity and arrhythmia events. The results showed that the kurtosis (P¼0.005) and skewness (P¼0.046) of MI patients with arrhythmia events were significantly higher than those without arrhythmia events. Androulakis et al. [24] used the entropy based on CMR-LGE images to measure scar heterogeneity parameters in patients with myocardial infarction and the entire left ventricular myocardium. In the multivariate analysis results, the risk of ventricular arrhythmia increased by 1.9-fold for every unit of entropy increase in scar tissue (hazard ratio: 1.9; 95% confidence interval: 1.0-3.5; P = 0.042), indicating that entropy may be related to the fibrous scar causing the arrhythmia. For each unit increase in entropy throughout the left ventricle, the risk of death increased 3.2 times (hazard ratio: 3.2; 95% confidence interval: 1.1-9.9; P = 0.038), suggesting that the entropy of the entire left ventricle may reflect the fibrosis pattern associated with poor cardiac remodeling. However, there was no statistically significant difference in the patient’s entropy with or without ventricular arrhythmia (P = 0.07) in this study. Some scholars analyzed that it may be due to the loss of the spatial relationship between pixels and original pixels when the intensity signal value was input. So entropy could not be directly applied to image data analysis [39].

HCM is the most common primary cardiomyopathy, with a prevalence of around 1/500. The diagnostic criteria for HCM are end-diastolic left ventricular wall thickness $\ge$ 15 mm, or the patient with a family history of HCM in whose end-diastolic left ventricular wall thickness $\ge$ 13 mm and excluding hypertensive cardiomyopathy, valvular heart disease, and other cardiac diseases secondary to ventricular wall thickening [40]. Among the patients with genetically positive HCM, 23% had hypertension simultaneously [41], so the differential diagnosis of HCM and hypertensive heart disease is particularly important in clinical diagnosis and treatment. Neisius et al. [18] analyzed the texture of the images with native scanning T1-mapping and showed that these six features, namely, Run-length non-enrichment [RLN]-135${}^{\circ }$, [SRHGE]-135${}^{\circ }$, LBP-15, LBP-20, LBP-25, and LBP-28, can differentiate HCM and hypertensive heart disease, with the maximum diagnostic accuracy of 86.2%. Schofield et al. [42] tried to use these six first order features of the cine images to distinguish left ventricular hypertrophy causes on the unenhanced cine sequence images, which were mean value, standard deviation, entropy, kurtosis, average positive pixel, and skewness. The results showed that compared with the normal control group, HCM and non-compaction of the ventricular myocardium had the best performance, and the six texture features had significant statistical significance (P 0.001). In aortic stenosis disease, except entropy, the remaining five texture features had significant statistical significance (P 0.001). In hypertensive heart disease, except for skewness, the remaining five texture features were statistically significant. In the differential diagnosis of disease, these six texture features were statistically significant in the differential diagnosis of non-compaction of the ventricular myocardium and aortic stenosis, and the better result appeared between hypertrophic cardiomyopathy and aortic stenosis disease (AUC = 0.89). These studies demonstrated that radiomics could assist clinicians in the differential diagnosis of HCM and other diseases that cause left ventricular hypertrophy.

Most patients with HCM have a good prognosis, but a small number of patients have adverse cardiac events, such as adverse cardiac remodeling, malignant arrhythmia, sudden cardiac death. Some scholars believe that a key mechanism of adverse events is myocardial fibrosis. Since Choudhury et al. [43] first reported the evaluation of HCM fibrosis by delayed enhancement of LGE sequence in 2002, more and more studies have confirmed that the existence and degree of delayed enhancement were related to the occurrence of adverse cardiac events. In recent years, many studies have begun to explore the relationship between CMR-LGE texture features and cardiac adverse events. As far as we know, Amano et al. [44] reported, for the first time, the correlation between CMR-LGE image texture features of HCM and ventricular arrhythmia. The results showed that among the four texture features, the entropy LL of patients with arrhythmia was significantly smaller than that of patients without arrhythmia (P = 0.0058), the area under the curve of entropy LL was 0.72, and the area under the curve of delayed enhancement was 0.96. Although entropy LL was weaker than delayed enhancement in distinguishing HCM from ventricular tachycardia in this study, entropy LL was indeed a discriminating factor for HCM with or without ventricular tachycardia. Unlike previous scholars, Cheng et al. [45] magnified the focus from several specific texture features to 90 features extracted from images and studied the correlation between LGE-CMR texture features and poor prognosis of HCM patients with systolic dysfunction. The results showed that the higher X0_GLRLM_energy, lower X0_H_skewness and lower X0_GLCM_cluster_tendency were positively correlated with adverse events (P 0.05), suggesting that increased myocardial heterogeneity was associated with adverse cardiac events in HCM patients. Alis et al. [30] proposed whether texture analysis based on CMR-LGE image can identify ventricular arrhythmia in HCM patients. The results showed that ventricular arrhythmias in HCM can be identified using a K-NN classifier with 94% diagnostic accuracy (AUC = 0.92). This study’s highlight is not limited to a few texture features, but also through the method of radiomics to screen features and establish a model to evaluate the correlation between radiomics and cardiac adverse events. However, due to the small sample size, the research trains and verifies all the data without external testing, so it is impossible to evaluate generalization efficiency.

HCM is usually found in youth, so some patients need to have multiple CMR examinations during long-term follow-up. Gadolinium-based contrast agents are the most common contrast mediums used in CMR examination. However, gadolinium is a contraindication for patients with renal failure, and it will cause nephrogenic systemic fibrosis. This is because after receiving gadolinium-based contrast agents in patients with renal function impairment, the clearance rate of contrast medium in vivo is slowed down, and Gd${}^3$+ is more easily released from the chelate, which promotes the production of chemokines and growth factors by macrophages and monocytes; it can also stimulate the production of hyaluronic acid, fibronectin, and types I and III collagens. Therefore, patients with impaired renal function are prone to nephrogenic systemic fibrosis [46]. Accordingly, Baebler et al. [17] used texture analysis on T1 weighted images of HCM patients without contrast to identify the differences between HCM and normal myocardium. Results showed that the four texture features were gray level non -uniformity (P 0.001), energy of wavelet coefficients in low frequency sub-bands (P 0.001), fraction (P 0.001) and sum average (P = 0.007), which provided a new parameter for the evaluation of HCM by contrast free T1 weighted images. Neisius et al. [47] used texture analysis on T1 mapping images of HCM patients to identify a delayed enhancement on LGE sequences. The results showed that the five texture features, LRLGE-45${}^{\circ }$, LBP-22, LBP-16, homogeneity-4, and LBP-1, had the greatest contribution value to the DTE classifier. In the DTE classifier, the c-index of the training set was 0.75, and that of the test set was 0.74. The learning DTE classifier could identify some LGE negative patients, suggesting that about 1/3 of LGE negative patients could avoid using gadolinium-based contrast agents, consistent with nearly 50% of HCM patients without delayed enhancement in previous studies [48], that may contribute unnecessarily to the use of contrast media in the future. The above studies show that radiomics plays an important role in diagnosing, differential diagnosis, and prognosis evaluation of HCM, which provides a new technology and direction for future HCM research.

The etiology of DCM can be either primary or secondary. This disease is characterized by left ventricular or right ventricular dilatation with contractile dysfunction. The diagnosis of DCM remains a major clinical challenge because of the exclusion of ischemic cardiomyopathy and other non-ischemic cardiomyopathies that may produce a pattern similar to the left ventricular remodeling. Shao et al. [49] studied whether texture analysis on the T1-mapping image of SVM was helpful to DCM diagnosis. The results showed that the SVM classifier’s diagnostic accuracy based on a meaningful histogram of the T1-mapping MRI image and GLCM was 85%. This method may provide an auxiliary tool for objective quantitative evaluation of clinical DCM diagnosis. Accurate risk stratification of sudden cardiac death is essential for the prognosis of patients with DCM.

The accuracy of left ventricular ejection fraction (LVEF) used to be a risk stratification index is limited and is greatly affected by the operator. Therefore, scholars hope to explore a new parameter to assist the risk stratification assessment of DCM. Myocardial fibrosis in patients with DCM provides a material basis for ventricular arrhythmia [50]. Myocardial heterogeneity is formed by the coexistence of different types and degrees of fibrous tissue with viable myocardium. Muthalaly et al. [51] used radiomics methods and hypothesized that the left ventricular myocardium heterogeneity was considerable in DCM patients requiring primary prevention by CMR assessment; that is, the entropy value was enormous, and the risk of ventricular arrhythmia was also greater. It showed that when judging whether DCM patients had an arrhythmia, left ventricular entropy was statistically significant in univariate factor analysis and multivariate factor analysis (P 0.05), suggesting that entropy was an independent predictor of ventricular arrhythmia in DCM patients risk factors. After 2.4 years of follow-up, no arrhythmia occurred in patients with LVEF 35% and left ventricular entropy 4.46, indicating that left ventricular entropy can provide a stable and straightforward risk assessment method for DCM patients. At present, the incidence rate of radiomics studies in DCM is still low, probably because the DCM diagnosis is not clear enough, the incidence of DCM is low, and the DCM wall is thin, and it is not easy to draw ROIs.

The viral infection is the most common cause of myocarditis, with an incidence rate of 20/10000 and a mortality rate of 4.8/100000 [52, 53]. The disease pathological changes comprise myocardial edema, degeneration, necrosis, and lymphocyte infiltration [54, 55]. Myocarditis is asymptomatic in mild cases, but severe arrhythmia, acute cardiac insufficiency, and sudden cardiac death have been noticed in some cases. The “gold standard” for diagnosing myocarditis is the endomyocardial biopsy, but the examination is invasive and has low sensitivity, so it is not widely used in clinical practice [56, 57]. CMR provides specific value for noninvasive diagnosis of myocarditis. Noninvasive diagnosis of myocarditis is still one of the challenging problems in clinical practice. Baessler et al. [27] analyzed the T1 and T2 mapping images of patients with acute infarct myocarditis similar to acute infarct myocarditis using radiomics. They found that these three texture features were statistically different between patients with positive and negative endocardial biopsy (P 0.05), which were T2-rluni, T2 sum entropy, and T2-GLevNonU. In the classification model composed of two run-length matrix features (T2-rluni and T2-GLevNonU), the AUC was 0.88, so these two features could well diagnose patients with acute infarct myocarditis. Besides, Baessler et al. [58] evaluated the diagnostic value of imaging radiomics parameters for myocarditis in heart failure patients. They found that the combination of the two features of average T2 time and T2-GLNU had the highest diagnostic performance in patients with acute heart failure myocarditis (AUC = 0.76, 95% confidence interval: 0.43-0.95), the combination of T2_kurtosis and T1-GLevNonU in myocarditis patients with chronic heart failure had the highest diagnostic performance (AUC = 0.85, 95% confidence interval: 0.57-0.90). In patients with acute and chronic heart failure with myocarditis, the difference in texture features with the best performance may reflect the two diseases different pathological changes. In the diagnosis of myocarditis, the texture features based on T2 images perform well, mainly because myocarditis is closely related to myocardial edema. So the T2 sequence maybe the best choice to observe myocardial water content of CMR. In summary, it has particular potential research value in the diagnosis and differential diagnosis of myocarditis.