ORIGINAL RESEARCH published: 29 October 2021 doi: 10.3389/fcvm.2021.764312 Radiomics-Based Classification of Left Ventricular Non-compaction, Hypertrophic Cardiomyopathy, and Dilated Cardiomyopathy in Cardiovascular Magnetic Resonance Cristian Izquierdo 1*, Guillem Casas 2,3,4,5 , Carlos Martin-Isla 1 , Victor M. Campello 1 , Andrea Guala 3,4 , Polyxeni Gkontra 1 , Jose F. Rodríguez-Palomares 2,3,4,5 and Karim Lekadir 1 1 Artificial Intelligence in Medicine Lab (BCN-AIM), Departament de Matemàtiques i Informàtica, Universitat de Barcelona, Barcelona, Spain, 2 Department of Cardiology, Hospital Universitari Vall d’Hebron, Barcelona, Spain, 3 Vall d’Hebron Institut de Recerca (VHIR), Barcelona, Spain, 4 CIBER-CV, Instituto de Salud Carlos III, Madrid, Spain, 5 Departament de Medicina, Universitat Autonoma de Barcelona, Barcelona, Spain Edited by: Left Ventricular (LV) Non-compaction (LVNC), Hypertrophic Cardiomyopathy (HCM), Márton Kolossváry, Massachusetts General Hospital and and Dilated Cardiomyopathy (DCM) share morphological and functional traits that Harvard Medical School, increase the diagnosis complexity. Additional clinical information, besides imaging data United States such as cardiovascular magnetic resonance (CMR), is usually required to reach a Reviewed by: Lennart Tautz, definitive diagnosis, including electrocardiography (ECG), family history, and genetics. Charité - Universitätsmedizin Berlin, Alternatively, indices of hypertrabeculation have been introduced, but they require Germany tedious and time-consuming delineations of the trabeculae on the CMR images. In Maria Aurora Morales, National Research Council (CNR), this paper, we propose a radiomics approach to automatically encode differences in Italy the underlying shape, gray-scale and textural information in the myocardium and its *Correspondence: trabeculae, which may enhance the capacity to differentiate between these overlapping Cristian Izquierdo c.izquierdo@ub.edu conditions. A total of 118 subjects, including 35 patients with LVNC, 25 with HCM, 37 with DCM, as well as 21 healthy volunteers (NOR), underwent CMR imaging. Specialty section: A comprehensive radiomics characterization was applied to LV short-axis images to This article was submitted to Cardiovascular Imaging, quantify shape, first-order, co-occurrence matrix, run-length matrix, and local binary a section of the journal patterns. Conventional CMR indices (LV volumes, mass, wall thickness, LV ejection Frontiers in Cardiovascular Medicine fraction—LVEF—), as well as hypertrabeculation indices by Petersen and Jacquier, Received: 25 August 2021 were also analyzed. State-of-the-art Machine Learning (ML) models (one-vs.-rest Accepted: 08 October 2021 Published: 29 October 2021 Support Vector Machine—SVM—, Logistic Regression—LR—, and Random Forest Citation: Classifier—RF—) were used for one-vs.-rest classification tasks. The use of radiomics Izquierdo C, Casas G, Martin-Isla C, models for the automated diagnosis of LVNC, HCM, and DCM resulted in excellent Campello VM, Guala A, Gkontra P, Rodríguez-Palomares JF and one-vs.-rest ROC-AUC values of 0.95 while generating these results without the Lekadir K (2021) Radiomics-Based need for the delineation of the trabeculae. First-order and texture features resulted Classification of Left Ventricular to be among the most discriminative features in the obtained radiomics signatures, Non-compaction, Hypertrophic Cardiomyopathy, and Dilated indicating their added value for quantifying relevant tissue patterns in cardiomyopathy Cardiomyopathy in Cardiovascular differential diagnosis. Magnetic Resonance. Front. Cardiovasc. Med. 8:764312. Keywords: radiomics, machine learning, left-ventricle non-compaction, dilated cardiomyopathy, hypertrophic doi: 10.3389/fcvm.2021.764312 cardiomyopathy Frontiers in Cardiovascular Medicine | www.frontiersin.org 1 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification 1. INTRODUCTION TABLE 1 | Cohort size for each specific disease/control and clinical criteria for inclusion. Cardiomyopathies (CMs) are defined as primary myocardial Disease Cohort size Clinical inclusion criteria disorders in the absence of other conditions that may affect the structural or functional properties of the heart’s muscle DCM 37 Depressed LVEF with increased LV (1). CMs are divided into distinct morphologic phenotypes volumes. Usually normal LV mass, wall (1), including hypertrophic cardiomyopathy (HCM) and dilated thickness and asymmetry. cardiomyopathy (DCM) as two of the most prevalent CMs. HCM 25 Increased LV mass with wall thickness > 15mm and/or asymmetry > 1.3. HCM is characterized by an increase in left ventricular (LV) wall Ussually preserved LVEF. thickness (2), DCM by LV (or biventricular) systolic dysfunction LVNC 35 Jacquier ratio > 20% and Petersen ratio and dilatation (3), while both disorders are unexplained by > 2.3. LVEF, LV volumes, LV mass and loading conditions. Left ventricular non-compaction (LVNC) is a wall thickness can be normal or not. recently defined and poorly understood condition, characterized NOR 21 Normal conventional CMR values. No by prominent LV trabeculae, a thin compacted myocardial layer, history of relevant cardiovascular disease and deep inter-trabeculae recesses (4). or systemic diseases. Cardiac magnetic resonance (CMR) is current the gold standard imaging modality for the clinical assessment of CMs, as well as to identify and differentiate the different phenotypes. combine multiple radiomics features into a novel discriminative CMR is widely used in the diagnosis of HCM (5), DCM model of LVNC, HCM, and DCM. Subsequently, the obtained (6), and LVNC (7–9). However, some LVNC features can radiomics model is evaluated in great detail and its performance overlap with those of other CMs and LVNC patients might compared to the one obtained based on CMR indices, including present with morphological findings of HCM and/or DCM the existing, manually estimated trabecular indices of LVNC. The (10). Furthermore, hypertrabeculation may also occur in the results described in this paper show the promise of the proposed healthy population, which makes it challenging to differentiate radiomics approach for achieving state-of-the-art LVNC, HCM, physiologic from pathological hyper-trabeculation forms (11) and DCM differential diagnosis more efficiently, while removing by CMR. The difficulties to differentially and timely diagnose the need for the delineation of the LV trabeculae. LVNC in clinical practice has motivated the development of new imaging indices, in particular, the Petersen (7) and Jacquier (8) coefficients, which estimate the level of hypertrabeculation in the 2. DATA AND METHODOLOGY LV myocardium. However, these coefficients, while they improve 2.1. Dataset LVNC diagnosis (9), are challenging and tedious to estimate in The study cohort consists of 118 subjects, including 37 DCM, practice, as they require expert and accurate identification and 25 HCM, and 35 LVNC patients, as well as 21 healthy control delineation of the trabeculae on the CMR images. This is a time- (NOR) subjects. HCM and DCM populations were available consuming task that is furthermore subject to inter-observer from the 2020 M&Ms MICCAI Challenge dataset (18). All variability given the inherent complexity of the trabeculae. patients for this study were assessed at the Hospital Universitari Radiomics is an emerging image analysis technique for deeper Vall d’Hebron (HUVH) following standard CMR protocols. phenotyping of cardiovascular health and disease in CMR Table 1 summarizes the clinical diagnostic criteria for each (12). It enables the examination of a large pool of advanced disease. In short, HCM, DCM, and LVNC diagnoses were imaging features that describe a wide range of complex, as well established by expert cardiologists based on currently accepted as subtle traits of the cardiac tissues at different scales and imaging criteria (5–8) combined with other clinical data, such as locations. Compared to existing cardiac indices such as those electrocardiography, family history, and genetics. The mean age listed above, radiomics features encode multivariate information of the cohort was 49.4 ±17.97 and 76 subjects (65% of the cohort) by capturing and combining heterogeneous morphological were men (see Tables 2, 3 for more detailed information). (e.g., sphericity, compactness) and appearance (e.g., entropy, coarseness) properties of the tissues. Hence, in the last years, 2.2. CMR Clinical Indices several works have shown its potential for identifying new All patients underwent a standard CMR protocol. In brief, all imaging signatures that can be leveraged for enhanced cardiac scans were performed with a 1.5 Tesla scanner (Avanto, Siemens disease understanding (13, 14) and quantification (15–17). In Healthcare, Erlangen, Germany), with typical cine parameters addition to providing comprehensive indicators of cardiac health as follows: TR/TE (repetition time/echo time) = 3.2/1.5 ms, and disease, CMR radiomics features are easier to calculate as voxel size 1.4 × 1.4 × 8 mm, and a slice gap of 2.0 mm. The they only require the segmentation of the myocardial boundaries, temporal resolution was interpolated to 25 phases per cardiac and even this segmentation process can be automatized (18). cycle (28–37 ms). The protocol includes a complete cine short- This work is the first to develop and evaluate a radiomics axis ventricular stack with the base to apex coverage acquired model for automatically differentiating LVNC, DCM, and HCM using balanced steady-state free procession (bSSFP) with one phenotypes in CMR. Based on a clinical dataset comprising breath-hold per image slice. Short axis cine images were obtained different CM subgroups as well as healthy subjects from routine and analyzed. Semi-automatic contouring of LV endocardial and clinical practice, a machine learning pipeline is implemented to epicardial end-diastolic (ED) and end-systolic (ES) borders was Frontiers in Cardiovascular Medicine | www.frontiersin.org 2 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification TABLE 2 | Cohort characteristics. Characteristics Full cohort DCM HCM LVNC NOR Sample size 118 37 25 35 21 Age (years) 49.39 ± 17.97 50.24 ± 15.12 60.88 ± 17.84 44.57 ± 17.46 40.85 ± 16.99 Sex (M/F) 76 / 42 26 / 11 18 / 7 19 / 16 13 / 8 Male percentage (%) 64.41% 70.27% 72.00% 54.28% 61.90 % BMI (Kg/m2 ) 26.37 ± 3.65 27.12 ± 3.89 26.32 ± 2.59 25.24 ± 3.45 25.67 ± 4.38 EDLVV (ml) 169.02 ± 63.88 225.40 ± 67.72 126.54 ± 34.58 159.07 ± 49.61 136.87 ± 24.47 ESLVV (ml) 97.03 ± 66.49 158.19 ± 75.45 51.34 ± 21.01 87.36 ± 45.56 57.65 ± 12.52 LV Mass (g) 122.29 ± 46.06 153.04 ± 51.70 139.05 ± 46.65 95.53 ± 26.37 102.52 ± 18.70 LVEF (%) 47.18 ± 16.03 32.47 ± 12.72 60.23 ± 9.79 47.26 ± 14.14 58.07 ± 4.09 Petersen coefficient 1.77 ± 0.90 1.61 ± 0.53 0.80 ± 0.39 2.86 ± 0.48 1.45 ± 0.54 Jacquier (%) 17.47 ± 10.08 16.70 ± 4.80 12.32 ± 2.72 24.30 ± 15.66 14.10 ± 3.19 ISV (mm) 11.15 ± 4.86 9.40 ± 1.70 19.12 ± 4.41 8.48 ± 1.61 9.09 ± 1.57 PW (mm) 7.26 ± 1.42 7.48 ± 1.46 8.25 ± 1.49 6.51 ± 1.10 6.95 ± 0.92 Assymetry (IVS/PW) 1.51 ± 0.57 1.27 ± 0.18 2.34 ± 0.75 1.30 ± 0.18 1.30 ± 0.16 The table presents the average and standard deviation for the characteristics of the entire cohort, including age, sex, percentage of men, and BMI. On the right are presented the characteristics of each disease group. TABLE 3 | Mann Whitney U statistical test for demographics and clinical characteristics group comparison. Demographics DCM-HCM DCM-LVNC DCM-NOR HCM-LVNC HCM-NOR LVNC-NOR Age (years) 0.011 0.08 0.05 0.0011 < 0.001 0.26 BMI (Kg/m2 ) 0.354 0.144 0.208 0.378 0.415 0.432 CMR indices EDLVV (ml) < 0.0001 < 0.0001 < 0.0001 < 0.0036 0.08 0.031 ESLVV (ml) < 0.0001 < 0.0001 < 0.0001 < 0.0001 0.028 < 0.0014 LV mass (g) 0.02 < 0.0001 < 0.0001 0.0011 0.013 0.1214 LVEF (%) < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.099 0.0005 Petersen coefficient < 0.0001 < 0.0001 0.094 < 0.0001 < 0.0001 < 0.0001 Jacquier (%) < 0.0001 0.014 0.019 < 0.0001 0.040 0.0005 IVS (mm) < 0.0001 0.017 0.304 < 0.0001 < 0.0001 0.09 PW (mm) 0.014 0.0007 0.058 < 0.0001 0.0006 0.046 Asymetry < 0.0001 0.11 0.12 < 0.0001 < 0.0001 0.4934 P-values provided for existing CMR indices involved in classification. IVS, Inter-ventricular septum; PW, Posterior wall. performed with Circle 42 (CVi 42) software (Calgary, Canada). 2.3. Radiomics Extraction A total of 9 existing CMR indices were quantified, including From the region of interest provided by the LV endocardial LV ejection fraction (LVEF), end-diastolic and end-systolic LV and epicardial contours, radiomics were extracted from end- volumes (EDLVV, ESLVV), LV mass, inter-ventricular septum diastole (ED) and end-systole (ES) phases, following a pre- (IVS), posterior wall (PW) thickness, asymmetry (IVS/PW), established pipeline from the open-source Python [(20), version and Petersen and Jacquier coefficients (7, 8). Right ventricle 3.7.9] PyRadiomics library [(21), version 3.0]. A set of 420 contours were not provided for the study; thus features were radiomics features were extracted from the LV cavity (LV) not considered. However, these diseases are predominantly and LV myocardium (LVMYO) within the original filter, related to the LV cavity, therefore missing information from the including different types: 52 shape, 72 first-order, and 296 texture right ventricle was not considered relevant. Fractal dimensions features (see Supplementary Material for a full list of radiomics (9) were not included in the experiments due to its limited extracted). We perform a radiomics features characterization applicability in daily clinical routine and its lack of prognostic by considering both the ED and ES phases to be able to correlation (19). Additionally, Petersen and Jacquier coefficients identify disease-specific changes over a heartbeat cycle of the (7–9) were considered more validated for this experiments. All radiomics features. CMR analyses were performed by an expert cardiologist with Shape radiomics features quantify and measure several years of experience in the field. the morphological traits in segmented contours, Frontiers in Cardiovascular Medicine | www.frontiersin.org 3 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification independently of the gray-level intensity distribution. to play an important role in identifying grayscale changes in These subset of features are easily interpretable, as they the LV or LVMYO tissue, a relevant trait of trabeculations in are closely related to surfaces and volumes calculated LVNC subjects. with existing CMR indices in clinical routine. Some of these features include simple metrics such as volume, elongation, or surface area, and more advanced metrics 2.4. Machine Learning Scheme like sphericity. From the previous section, a total of 420 radiomics features First-order radiomics features are obtained from signal were extracted and were potential candidates for inclusion in intensity characteristics based on the histogram. Specifically, the targeted radiomics model for disease classification. However, first-order statistics describe the distribution of pixel/voxel not all of these features will have predictive power, and hence intensities within the image region defined by the mask feature selection will be first applied to select the most optimal through commonly used and basic metrics, regardless of the features for the classification task. We separated this feature spatial relationship (21). First-order radiomics features include selection process into 2 steps. First, we identify those that are easily interpretable features such as median or mean, and highly correlated, for each feature, and remove them from the more mathematically advanced metrics such entropy, energy, radiomics set as they carry a similar predictive signal. For this or kurtosis. purpose, we estimated Pearson correlation between all features, Finally, texture radiomics features capture subtle changes and those above 0.9 were considered redundant. It is well-known in the gray-scale pixel distribution, identifying tendencies, that radiomics are highly redundant thus, with this first step and neighboring gray-scale changes through advanced we only aimed to remove the most correlated ones, and be matrix calculations. Texture features can be grouped in: further reduced with a more sophisticated feature selector. The Gray Level Co-occurrence Matrix (GLCM), Gray Level Size Zone procedure resulted in a reduction from 420 radiomics features Matrix (GLSZM), Gray Level Run Length Matrix (GLRLM), to only 120. This reduced subset is introduced in the Pipeline Neighboring Gray Tone Difference Matrix (NGTDM), and Gray function from Python Sci-kit learn package [(22), version 0.24.2] Level Dependence Matrix (GLDM) (21). that has three different steps, including the additional feature Shape features are expected to capture morphological cardiac selection method mentioned above: characteristics normally associated with each disease. On the 1. Normalization: Data normalization is required before other hand, both first-order and texture features are expected introducing the data into the machine learning models FIGURE 1 | Machine learning validation: 10-fold outer loop Nested CV and a three-fold CV inner loop scheme. Frontiers in Cardiovascular Medicine | www.frontiersin.org 4 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification because different scales in the variables measured represent building section. For this purpose, the SelectKBest function that features have different contributions to the model from Python’s Sci-kit learn library (22) was performed. The fitting, and this may introduce bias. We applied the algorithm works by selecting the best features based on MinMaxScaling function from the Sci-kit Python library univariate statistical tests. It selects the features according to (22). The StandardScaler was also considered, although no two different parameters: highest score function and number of significant difference was found with a min-max scaling. features (k). These parameters had to be defined beforehand. 2. Feature selection: The number of features remaining was The score function parameter selected was f_classif, which still large and had to be reduced before reaching the model computes the ANOVA f -value for the sample and provides the associated p-value for each feature based on the correlation with the class label. The k parameter is defined as the number of features selected by the feature selector. Prior the analysis, TABLE 4 | Summary table of the testing performance of the selected models. we do not know the optimal number of features to be selected, therefore the k-value before the analysis had to be defined CMR indices Radiomics p-value within the range of parameters from 5 to 120 (i.e., number of Models Mean STD Mean STD possible features it may select), to be later introduced in the One vs. Rest SVM 0.972 0.03 0.942 0.03 > 0.05 Random forest 0.978 0.03 0.964 0.01 > 0.05 Logistic regression (multinomial) 0.970 0.03 0.956 0.03 > 0.05 The table provides AUC values for CMR existing indices and radiomics, along with the p-value. Bold values represent the highest AUC value obtained in test. TABLE 5 | Generic and differential diagnosis AUC testing metrics for Random Forest model. CMR indices Radiomics p-value General model Mean STD Mean STD Random Forest 0.978 0.03 0.964 0.01 > 0.05 Differential diagnosis models DCM-vs.-Rest 0.97 0.02 0.93 0.03 > 0.05 HCM-vs.-Rest 1.00 0.00 0.99 0.03 > 0.05 FIGURE 3 | In our grid search scheme, we analyzed incrementally the number NOR-vs.-Rest 0.95 0.02 0.97 0.02 > 0.05 of features selected for all the hyper-parameter combinations and we show LVNC-vs.-Rest 0.96 0.04 0.92 0.03 > 0.05 that validation AUC was not increasing significantly once reached 30–40 features, with the lowest STD at 40 features. AUC when selecting under 10 P-values are presented in the table for statistical significance analysis and prove they are features reduced drastically. comparable. Bold values represent the highest AUC value obtained in test. FIGURE 2 | Roc curve comparison for differential diagnosis. The left subfigure represents the CMR indices ROC curve. The right subfigure represents the radiomics ROC curve. Frontiers in Cardiovascular Medicine | www.frontiersin.org 5 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification hyper-parameter optimization Grid Search CV. The number average in a single list, we analyzed the features selected by each of features selected is tested iteratively, selecting the best k of the 10 models and selected for representation those features radiomics according to the score function and tested further that remained constant across the 10-folds and were finally with the model. sorted by feature importance score. With this, we create a highly 3. Model building: Three different machine learning algorithms were trained and tested: One vs. Rest Support Vector Machine (SVM), Multi-class Random Forest (RF), and Multi-class Logistic regression (LR) for classification (22). According TABLE 6 | Top-10 best performing radiomics for each differential diagnosis, to a recent review (23), both SVM and RF models were divided across subsection. the most used techniques for conventional ML for image- based diagnosis. Additionally, prior knowledge from a recent Radiomics feature Type Region Phase Weight (%) publication (24) proved also the reliability of SVM and RF DCM vs. Rest when dealing with radiomics. Finally, we decided to include Minor axis Shape LVMYO ES 7.0 Logistic Regression as one of the most-used techniques Volume Shape LV ES 6.7 in statistical analysis for the purpose of increasing the Least axis Shape LVMYO ES 5.8 comparison benchmark. Max2D diameter slice Shape LV ES 5.5 For evaluation, the experiment is validated in a nested CV Least axis Shape LV ED 5.4 scheme (25). We performed a 10-fold outer loop and a three- Least axis Shape LVMYO ED 5.1 fold inner loop (see Figure 1 for a more graphical description). Long run high gray level emphasis GLRLM LV ES 5.1 This represents that for each fold in the outer loop, 90% of the Minor axis Shape LVMYO ED 5.0 data is kept for training and validation, while the remaining 10% Volume Shape LV ED 5.0 will be held for testing. The same procedure was performed in Sphericity Shape LVMYO ES 5.0 the inner loop for each fold. The remaining train and validation HCM vs. Rest data were split into 66% for training and 33% for validation. Surface area to volume ratio Shape LVMYO ES 7.4 All the splits in our scheme were performed with Stratified K- Sphericity Shape LVMYO ES 7.2 fold sci-kit learn function (22) to keep classes balanced. The Large dependence high gray level E. GLDM LV ES 6.9 normalization and feature selection steps were performed in each Long run high gray level E. GLRLM LV ES 6.7 fold of the nested CV scheme to avoid data leakage. This means Sphericity Shape LVMYO ED 6.2 that no knowledge of the held test set was introduced into the Skewness First order LV ES 5.3 training stage, which could corrupt the learning process and its Gray level non-uniformity GLSZM LV ES 5.1 posterior generalization. Autocorrelation GLCM LV ES 5.0 Models’ performance are dependent on the hyper-parameters Energy First order LV ES 4.9 selected (26). For this purpose, a Grid Search CV (22) was applied Surface area to volume ratio Shape LV ES 4.8 in the inner loop, thus ensuring the optimal hyper-parameters NOR vs. Rest were selected. Grid Search CV (22) is an optimizer algorithm Gray level non-uniformity GLRLM LV ES 6.5 that calculates the model’s performance for each combination of Run length non-uniformity GLRLM LVMYO ES 6.0 hyper-parameters and keeps the one that achieved the highest Sphericity Shape LVMYO ES 5.9 prediction metric, to be later tested on the testing held data (see Low gray level run E. GLRLM LVMYO ED 5.7 Supplementary Table 1 in Supplementary Material for the full Dependence variance GLDM LVMYO ES 5.7 list of hyper-parameters used). Max2D diameter slice Shape LVMYO ES 5.2 Paired t-test on both distributions of testing AUC Gray level non-uniformity GLSZM LVMYO ES 5.2 performances was performed to analyze the statistical Long run high gray level emphasis GLRLM LVMYO ED 5.2 significance for each general machine learning model across CMR indices and radiomics, as well as for each differential Max2D diameter slice Shape LV ES 5.0 diagnosis (i.e., identifying a single disease class from the whole Minor axis Shape LV ED 4.9 cohort) and prove they were comparable. Additionally, Receiver LVNC vs. Rest Operating Characteristic (ROC) curves were calculated to Large dependence low gray level E. GLDM LVMYO ES 6.6 provide a better representation of the true and false-positive Surface area to volume ratio Shape LVMYO ED 6.5 rates for each differential diagnosis. Due to the architecture Inverse variance GLCM LVMYO ES 6.1 of our Nested CV scheme, we obtain 10 different tests AUC, Long run emphasis GLRLM LVMYO ES 5.9 one per fold (10-fold outer loop, see Figure 1, left side). This Large area low gray level E. GLSZM LVMYO ES 5.7 means that each of the 10 models resulted from the Grid Search Large dependence low gray level E. GLDM LVMYO ED 5.6 CV in the inner loop might be different (i.e., the combination Large area emphasis GLSZM LVMYO ED 5.2 of hyper-parameters and the number of features selected may Zone percentage GLSZM LVMYO ED 5.1 vary depending on different characteristics of the training and Long run emphasis GLRLM LVMYO ED 5.1 validation set). Thus, to present the most relevant features on Percentile 90 First order LVMYO ES 4.8 Frontiers in Cardiovascular Medicine | www.frontiersin.org 6 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification approximated list of the most relevant radiomics features for hyper-parameter combinations and we showed in Figure 3 that each classification. the AUC values did not increase significantly after integrating Since it is possible that different iterations selected a different 30–40 radiomics features in the model. number of features (i.e., for example, fold 1 could select 30 To further illustrate the predictive power of the radiomics features and fold 2, 40 features), we analyzed the validation features, Table 6 presents across subsections the 10 best AUC values for each number of features (k = 10, 20, 30, 40) performing radiomics for each differential diagnosis, sorted by across all the combination of hyper-parameters to see the effect their weighted feature importance (in percentage). Moreover, of increasing the number of features on AUC (see Figure 3). Table 7 shows the 10 best radiomics features involved for the Finally, to provide a more clinical perspective, we analyzed the general RF model. implications of feature type (shape, first-order, or texture), region By looking at Figure 4, we can observe that shape features play (LVMYO, LV cavity), and phase (ED, ES) for each differential an important role when classifying DCM against the rest of the diagnosis, and linked them with the existing clinical knowledge. diseases. Alternatively, texture features have a higher impact in the classification of HCM and healthy subjects, and even a higher impact in the differential diagnoses of LVNC. 3. RESULTS Left side of Figure 5 shows the distribution of the selected 3.1. Classification Performance radiomics features across the cardiac structures (i.e., LVMYO In the first experiment, we evaluated and compared the vs. LV cavity). For the identification of DCM, most of the performance of the different machine learning models (namely radiomics features (65%) pertained to the LV cavity, while the RF, OVR-SVM, and LR). As it can be seen in Table 4, the highest remaining 35% belonged to the LVMYO. Conversely, features AUC values were obtained by the RF technique, for all models. extracted from the LV cavity and the LVMYO participated Hence, the RF technique is used as the baseline models in the equally to the identification of HCM and healthy subjects. Finally, remainder of the experiments. Table 5 presents the in-depth the largest difference in terms of region importance can be results of the RF and each differential diagnosis. found for LVNC, where almost 96% of the features belonged Subsequently, we performed a comparison between the AUC to LVMYO, with only a 4% to LV cavity (see Figure 5, left scores obtained by the existing CMR indices (i.e., standard image). This last finding is in line with the existing clinical model) and radiomics models for classification. As it can be knowledge. Normally, papillary muscles are considered inside seen from Table 4, the radiomics model had a comparable the LV cavity and not as a myocardial mass, according to the performance to the standard model and there was no statistically guidelines and clinical consensus among cardiologists. But in significant differences between the two models (p > 0.05). patients with LVNC conditions, trabeculae and papillary muscles However, the radiomics model was obtained without the expert are quantified [within Jacquier coefficient (8)] outside the LV delineation of the trabeculae. In more detail, the ROC curves for cavity and included in the myocardial mass (LVMYO). For this the classification models are presented in Figure 2. reason, contrary to what the name itself suggests, the assessment 3.2. Radiomics Signatures In this section, we provide more details on the contributions of the different radiomics features to the classification models. We analyzed incrementally the number of features selected for all the TABLE 7 | Top-10 best performing radiomics for RF model. Radiomics feature Type Region Phase Weight (%) General RF model Sphericity Shape LVMYO ES 4.9 Long run high gray level emphasis GLRLM LV ES 4.6 Sphericity Shape LVMYO ED 3.6 Surface area to volume ratio Shape LVMYO ES 3.6 Gray level non-uniformity GLSZM LV ES 3.4 Minimum First order LVMYO ED 3.4 Least axis Shape LV ED 2.5 Large area low gray level emphasis GLSZM LVMYO ES 3.0 Volume Shape LV ED 2.9 Coarseness NGTDM LV ES 2.2 FIGURE 4 | Differential diagnosis overlapping radar plot, comparing the Please check Supplementary Table 2 in Supplementary Material for the full approximate distribution of the selected radiomics across types of radiomics. list of radiomics selected. Frontiers in Cardiovascular Medicine | www.frontiersin.org 7 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification FIGURE 5 | Distribution of the selected radiomics features across the cardiac structures [left image, i.e., LV myocardium (LVMYO) vs. LV blood pool (LV)] and cardiac cycle phase (right image). For this analysis, all the radiomics selected by the RF model were used. between LVNC and the rest of cardiomyopathies is determined myocardial fibrosis (see Table 7). Specifically, Long Run by changes or differences in the LVMYO. High Gray Level Emphasis is a texture feature that explains Regarding the contribution of cardiac cycle phases, ES longer contrasted gray level strings/regions, which can be related radiomics were more important than ED features for the to the existence and prominence of myocardial trabeculations, classification of DCM, HCM, and healthy subjects. However, ES typically associated with LVNC. and ED phases play a similar role when assessing LVNC (see Regarding the radiomics features for each differential Figure 5, right image). diagnosis, differences were found among the various diseases. Finally, we compared the time needed to obtain the diagnoses Concretely, we can observe how the contribution of the using the standard as well as the proposed radiomics models. radiomics features to the prediction are distributed, by type With the existing CMR indexes, our clinical experts spent (Figure 4), region and phase (Figure 5). While shape features approximately 9–12 min to delineate the trabeculae and then seem to play the most important role in DCM classification, derive an LVNC diagnosis, on average. In contrast, for the texture features are more important when classifying HCM and proposed radiomics-based approach, the time to assess one LVNC subjects. This finding is in line with clinical knowledge: subject was reduced to 10–20 s depending on the image DCM is defined primarily by a dilation of the left ventricle, while characteristics [i.e., volume, slice images, or bin width (21)]. myocardial fibrosis has a pivotal role in HCM diagnosis and LVNC is defined by the presence of myocardial trabeculations. 4. DISCUSSION 4.2. Limitations and Future Work 4.1. Summary of Findings The findings presented in this paper must be considered in Machine learning-based radiomics models showed excellent light of the study limitations, and future work may take performance for differentiating between hypertrophic different directions. Firstly, while the radiomics performance for cardiomyopathy, dilated cardiomyopathy, left ventricular automated diagnosis is promising, these results were obtained non-compaction, as well as healthy subjects. According to the based on a single-center small-size clinical cohort. To confirm results presented in Table 4 and Figure 2, radiomics and existing these promising results, future studies should be extended CMR indices resulted in similar performances. toward multi-center studies. Furthermore, this work relied on The 10-most significant radiomics features for the general RF semi-automated, manually controlled, delineations of the LV model comprise a combination of radiomics types, regions, and endo and epi-cardial contours on the short-axis images, before heart cycle phases (see Table 7). Looking in detail, myocardium the extraction of the existing indices and radiomics features. sphericity seems to play an important role in the overall However, automatic segmentation of the ventricular boundaries classification, occupying the first and third spots in terms of has been extensively investigated using deep learning (18) and predictive power for the ES and ED phases, respectively. This can these models could be extended to segment pathological cases in be explained by the remodeling of the heart that affects the global particular for LVNC. Finally, this paper focused on the diagnosis and regional structure of the left ventricle. of LVNC and related CMs. In clinical practice, subsequently, Moreover, texture features were found to add substantial prediction of LVNC related events before they occur would information, positioning in the second top position of the enable early and personalized prevention. Our plan is to extend ranking, which underlines the widely accepted diagnostic the proposed radiomics models to enable patient-specific risk importance of myocardial tissue characteristics such as prediction and prognosis estimation after LVNC diagnosis. Frontiers in Cardiovascular Medicine | www.frontiersin.org 8 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification 5. CONCLUSIONS AUTHOR CONTRIBUTIONS CMR radiomics constitutes a promising approach to CI, GC, AG, JR-P, PG, and KL: design of work. CI, VC, CM-I, differentially diagnose overlapping and complex conditions and PG: experimental setup and coding. CI, GC, and AG: such as HCM, DCM, and LVNC. The classification performance interpretation of results. CI and GC: manuscript draft. VC, of radiomics models are in-line with the one obtained by using CM-I, GC, AG, JR-P, and KL: manuscript review. All authors existing CMR indexes but the diagnoses can be reached fully contributed to the article and approved the submitted version. automatically without the need for expert delineation of the trabeculae as in previous works. FUNDING DATA AVAILABILITY STATEMENT This publication was partially funded by the European Union’s Horizon 2020 research and innovation euCanSHare project The original contributions presented in the study are included under grant agreement No 825903. KL received funding from in the article/Supplementary Material, further inquiries can be the Spanish Ministry of Science, Innovation and Universities directed to the corresponding author/s. under grant agreement RTI2018-099898-B-I00. AG has received funding from the Spanish Ministry of Science, Innovation and ETHICS STATEMENT Universities (IJC2018-037349-I). Ethical review and approval was not required for the study SUPPLEMENTARY MATERIAL on human participants in accordance with the local legislation and institutional requirements. Written informed consent for The Supplementary Material for this article can be found participation was not required for this study in accordance with online at: https://www.frontiersin.org/articles/10.3389/fcvm. the national legislation and the institutional requirements. 2021.764312/full#supplementary-material REFERENCES 9. Captur G, Muthurangu V, Cook C, Flett AS, Wilson R, Barison A, et al. Quantification of left ventricular trabeculae using fractal analysis. J Cardiovasc 1. Elliott P, Andersson B, Arbustini E, Bilinska Z, Cecchi F, Charron P, Magnet Reson. (2013) 15:36. doi: 10.1186/1532-429X-15-36 et al. Classification of the cardiomyopathies: a position statement from the 10. Oechslin E, Jenni R. Left ventricular noncompaction. J Am Coll Cardiol. (2018) European society of cardiology working group on myocardial and pericardial 71:723–6. doi: 10.1016/j.jacc.2017.12.031 diseases. Eur Heart J. (2007) 29:270–6. doi: 10.1093/eurheartj/ehm342 11. Oechslin E, Klaassen S. Left ventricular noncompaction. J Am Coll Cardiol. 2. Elliott P, Anastasakis A, Borger MA, Borggrefe M, Cecchi F, Charron P, (2019) 73:1612–5. doi: 10.1016/j.jacc.2018.11.064 et al. 2014 ESC Guidelines on diagnosis and management of hypertrophic 12. Raisi-Estabragh Z, Izquierdo C, Campello VM, Martin-Isla C, Jaggi A, cardiomyopathy: the Task Force for the Diagnosis and Management of Harvey NC, et al. Cardiac magnetic resonance radiomics: basic principles Hypertrophic Cardiomyopathy of the European Society of Cardiology (ESC). and clinical perspectives. Eur Heart J Cardiovasc Imaging. (2020) 21:349–56. Eur Heart J. (2014). 35:2733–79. doi: 10.1093/eurheartj/ehu284 doi: 10.1093/ehjci/jeaa028 3. Pinto YM, Elliott PM, Arbustini E, Adler Y, Anastasakis A, Böhm M, et al. 13. Cetin I, Raisi-Estabragh Z, Petersen SE, Napel S, Piechnik SK, Neubauer S, Proposal for a revised definition of dilated cardiomyopathy, hypokinetic non- et al. Radiomics signatures of cardiovascular risk factors in cardiac MRI: dilated cardiomyopathy, and its implications for clinical practice: a position results from the UK biobank. Front Cardiovasc Med. (2020) 7:591368. statement of the ESC working group on myocardial and pericardial diseases. doi: 10.3389/fcvm.2020.591368 Eur Heart. (2016) 37:1850–8. doi: 10.1093/eurheartj/ehv727 14. Amano Y, Suzuki Y, Yanagisawa F, Omori Y, Matsumoto N. Relationship 4. Jenni R, Oechslin EN, van der Loo B. Isolated ventricular non- between extension or texture features of late gadolinium enhancement and compaction of the myocardium in adults. Heart. (2007) 93:11–5. ventricular tachyarrhythmias in hypertrophic cardiomyopathy. Biomed Res doi: 10.1136/hrt.2005.082271 Int. (2018) 2018:1–6. doi: 10.1155/2018/4092469 5. Cardim N, Galderisi M, Edvardsen T, Plein S, Popescu BA, D’Andrea 15. Aerts HJWL, Velazquez ER, Leijenaar RTH, Parmar C, Grossmann P, A, et al. Role of multimodality cardiac imaging in the management Carvalho S, et al. Decoding tumour phenotype by noninvasive imaging of patients with hypertrophic cardiomyopathy: an expert consensus of using a quantitative radiomics approach. Nat Commun. (2014) 5:5644. the European Association of Cardiovascular Imaging Endorsed by the doi: 10.1038/ncomms5644 Saudi Heart Association. Eur Heart J Cardiovasc Imaging. (2015) 16:280. 16. Neisius U, El-Rewaidy H, Nakamori S, Rodriguez J, Manning WJ, Nezafat doi: 10.1093/ehjci/jeu291 R. Radiomic analysis of myocardial native T1 imaging discriminates 6. Donal E, Delgado V, Bucciarelli-Ducci C, Galli E, Haugaa KH, Charron P, et al. between hypertensive heart disease and hypertrophic cardiomyopathy. JACC Multimodality imaging in the diagnosis, risk stratification, and management Cardiovasc Imaging. (2019) 12:1946–54. doi: 10.1016/j.jcmg.2018.11.024 of patients with dilated cardiomyopathies: an expert consensus document 17. Cheng S, Fang M, Cui C, Chen X, Yin G, Prasad SK, et al. LGE-CMR-derived from the European Association of Cardiovascular Imaging. Eur Heart J texture features reflect poor prognosis in hypertrophic cardiomyopathy Cardiovasc Imaging. (2019) 20:1075–93. doi: 10.1093/ehjci/jez178 patients with systolic dysfunction: preliminary results. Eur Radiol. (2018) 7. Petersen SE, Selvanayagam JB, Wiesmann F, Robson MD, Francis JM, 28:4615–24. doi: 10.1007/s00330-018-5391-5 Anderson RH, et al. Left ventricular non-compaction. J Am Coll Cardiol. 18. Campello VM, Gkontra P, Izquierdo C, Martin-Isla C, Sojoudi A, (2005) 46:101–5. doi: 10.1016/j.jacc.2005.03.045 et al. PMF. Multi-centre, multi-vendor and multi-disease cardiac 8. Jacquier A, Thuny F, Jop B, Giorgi R, Cohen F, Gaubert JY, et al. Measurement segmentation: the M &Ms challenge. IEEE Trans Med Imaging. (2021). of trabeculated left ventricular mass using cardiac magnetic resonance doi: 10.1109/TMI.2021.3090082. [Epub ahead of print]. imaging in the diagnosis of left ventricular non-compaction. Eur Heart J. 19. Ivanov A, Dabiesingh DS, Bhumireddy GP, Mohamed A, Asfour (2010) 31:1098–104. doi: 10.1093/eurheartj/ehp595 A, Briggs WM, et al. Prevalence and prognostic significance of left Frontiers in Cardiovascular Medicine | www.frontiersin.org 9 October 2021 | Volume 8 | Article 764312 Izquierdo et al. Cardiac Pathologies’ Radiomics-Based Classification ventricular noncompaction in patients referred for cardiac magnetic 26. Probst P, Boulesteix AL, Bischl B. Tunability: importance of hyperparameters resonance imaging. Circ Cardiovasc Imaging. (2017) 10:e006174. of machine learning algorithms. J Mach Learn Res. (2019) 20:1934–65. doi: 10.1161/CIRCIMAGING.117.006174 20. Van Rossum G, Drake FL. Python 3 Reference Manual. Scotts Valley, CA: Conflict of Interest: The authors declare that the research was conducted in the CreateSpace (2009). absence of any commercial or financial relationships that could be construed as a 21. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, potential conflict of interest. et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. (2017) 77:e104–7. doi: 10.1158/0008-5472.CAN-17-0339 Publisher’s Note: All claims expressed in this article are solely those of the authors 22. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, and do not necessarily represent those of their affiliated organizations, or those of et al. Scikit-learn: machine learning in Python. J Mach Learn Res. (2011) the publisher, the editors and the reviewers. Any product that may be evaluated in 12:2825–30. this article, or claim that may be made by its manufacturer, is not guaranteed or 23. Martin-Isla C, Campello VM, Izquierdo C, Raisi-Estabragh Z, Baeßler B, Petersen SE, et al. Image-based cardiac diagnosis with machine learning: a endorsed by the publisher. review. Front Cardiovasc Med. (2020) 7:1. doi: 10.3389/fcvm.2020.00001 24. Rauseo E, Morcillo CI, Raisi-Estabragh Z, Gkontra P, Aung N, Lekadir K, et al. Copyright © 2021 Izquierdo, Casas, Martin-Isla, Campello, Guala, Gkontra, New imaging signatures of cardiac alterations in ischaemic heart disease and Rodríguez-Palomares and Lekadir. This is an open-access article distributed under cerebrovascular disease using CMR radiomics. Front Cardiovasc Med. (2021) the terms of the Creative Commons Attribution License (CC BY). The use, 8:716577. doi: 10.3389/fcvm.2021.716577 distribution or reproduction in other forums is permitted, provided the original 25. Cawley GC, Talbot NLC. On over-fitting in model selection and subsequent author(s) and the copyright owner(s) are credited and that the original publication selection bias in performance evaluation. J Mach Learn Res. (2010) in this journal is cited, in accordance with accepted academic practice. No use, 11:2079–107. distribution or reproduction is permitted which does not comply with these terms. Frontiers in Cardiovascular Medicine | www.frontiersin.org 10 October 2021 | Volume 8 | Article 764312