Early Myocardial Infarction Detection over Multi-view Echocardiography

JOURNAL OF L A T E X 1 Early Myocardial Infarction Detection o v er Multi-vie w Echocardiography A ysen Degerli, Serkan Kiran yaz, T ahir Hamid, Rashid Mazhar , and Moncef Gabbouj Abstract —Myocardial infar ction (MI) is the leading cause of mortality in the world that occurs due to a blockage of the coronary arteries feeding the myocardium. An early diagnosis of MI and its localization can mitigate the extent of myocardial damage by facilitating early therapeutic interventions. F ollowing the blockage of a coronary artery , the regional wall motion abnormality (RWMA) of the ischemic myocardial segments is the earliest change to set in. Echocardiography is the fundamental tool to assess any R WMA. Assessing the motion of the left ventricle (L V) wall only fr om a single echocardiograph y view may lead to missing the diagnosis of MI as the R WMA may not be visible on that specific view . Theref ore, in this study , we pr opose to fuse apical 4-chamber (A4C) and apical 2-chamber (A2C) views in which a total of 12 myocardial segments can be analyzed for MI detection. The proposed method first estimates the motion of the L V wall by Activ e Polynomials (APs), which extract and track the endocardial boundary to compute my ocardial segment displacements. The features are extracted from the A4C and A2C view displacements, which are concatenated and fed into the classifiers to detect MI. The main contrib utions of this study are 1) cr eation of a new benchmark dataset by including both A4C and A2C views in a total of 260 echocardiography record- ings, which is publicly shared with the research community , 2) improving the performance of the prior work of threshold- based APs by a Machine Learning based approach, and 3) a pioneer MI detection appr oach via multi-view echocardiograph y by fusing the information of A4C and A2C views. Experimen- tal results show that the proposed method achieves 90.91% sensitivity and 86.36% precision f or MI detection ov er multi- view echocardiography . The software implementation is shared at https://github .com/degerliaysen/MultiEchoAI. Index T erms —Active Polynomials, Echocardiograph y , Machine Learning, Motion Estimation, My ocardial Infarction. I . I N T RO D U C T I O N M Y OCARDIAL infarction (MI) is caused by the death of myocardial cells subsequent to ischemia due to the blockage of coronary arteries. Presentation of MI is gener - ally evident with shortness of breath, pain around the chest, shoulders, back, and arms [1]. Howe ver , these symptoms may not occur in the early stages of MI. Due to the blockage of the coronary artery and depriv ation of blood supply , there is progressiv e damage to the af fected part of the myocardium. Hence, it is critical to mak e an early detection of MI, to limit This work was supported by the NSF-Business Finland Center for V isual and Decision Informatics (CVDI) Advanced Machine Learning for Industrial Applications (AMaLIA) project under Grant 4183/31/2021. A ysen Degerli (aysen.de gerli@tuni.fi) and Moncef Gabbouj (mon- cef.gabbouj@tuni.fi) are with the Faculty of Information T echnology and Communication Sciences, T ampere University , T ampere, Finland. Serkan Kiranyaz (mkiranyaz@qu.edu.qa) is with the Department of Elec- trical Engineering, Qatar University , Doha, Qatar . T ahir Hamid and Rashid Mazhar are with the Hamad Medical Corporation, Doha, Qatar . and pre vent death & disability . Currently , the diagnosis of MI is based upon a time-consuming method of serial observations of electrocardiography (ECG), blood le vel of cardiac enzymes, and imaging techniques [2]. At the outset of MI, ECG is an insensitiv e tool with 0 . 77 predictiv e value in ruling out MI [3]. Moreov er , human error leads to misdiagnosis of ischemic ECG changes in 12 − 16% cases of MI [4]–[6]. Furthermore, cardiac biomarkers take time to ev olve to a diagnostic le vel. After the onset of MI, the high sensitivity cardiac troponin (hs-cTn) starts to rise in 3 hours and it needs a repeat sample at least 6 hours after onset of chest pain to quantify according to the American Heart Association definition of MI [7]. Therefore, the most con venient tool to diagnose and assess MI in its early stages is echocardiography , which has easy accessibility , low cost, and lo west risk compared to other cardiac imaging options [8], [9]. T wo-dimensional (2D) echocardiography was first intro- duced in the late 1950 s, which is a non-in vasi ve ultrasound imaging technique that monitors the heart in real-time [9], [10]. The early detection of MI can be performed by e valu- ating the regional wall motion abnormality (R WMA) in 2D echocardiography , where the abnormalities caused by MI can be detected as a re gion of weaker motion of the myocardium. Howe ver , the assessment of R WMA is highly subjecti ve and variant among experts [11]. Moreover , the echocardiography recordings are generally subject to a high lev el of noise with low image quality , where the left ventricle (L V) wall is mostly unrecognizable. Thus, visual assessment of the R WMA highly depends on the expertise of the echocardiographist and the quality of the echocardiography recordings. Therefore, in order to achieve reliable MI detection, computer-aided diagnosis techniques are de veloped to help cardiologists in the diagnosis. Consequently , motion estimation algorithms are utilized to assess and quantify the R WMA in echocardiography . Se veral approaches that are popular for estimating myocardial motion are optical flo w methods, deformation imaging, and activ e contours. Gradient-based optical flow methods estimate the motion by capturing the flow of pixels with constant intensity ov er time. The optical flo w can be described as the v elocity distrib utions of the bright pix el movements in the image that can be ap- proximated by the partial deriv ativ es with respect to spatial and temporal coordinates. Several studies [12]–[15] hav e utilized gradient-based optical flow methods by adding constraints to regularize the myocardial motion in 2D echocardiography . Howe ver , the ultrasound imaging is highly v ariant with the angle and depth of the ultrasound beam, which results in many artifacts, such as noise, shadowing, and dropouts on the image JOURNAL OF L A T E X 2 Fig. 1: The chambers of the heart in A2C and A4C view echocardiography . that cause optical flow methods to fail at estimating large L V wall displacements [16], [17]. Moreover , the noisy nature of echocardiography de grades the performance of the optical flo w algorithms. Deformation imaging is widely used in echocardiography to perform strain analysis of the myocardium [18]–[27]. The strain is defined by the length of the L V wall that is measured via the speckle tracking method, which is kno wn also as the region-based optical flo w or block-matching method in echocardiography . The speckle tracking method searches for a similar block of pixels through consecutiv e frames in a specified search windo w . The strain measurements refer to deformation of the myocardium, e.g., if the v elocity of the L V wall segments is non-uniform, then the myocardium is infarcted. Even though deformation imaging is a promising method to detect MI, it suffers sev erely from the weakness of the optical flo w methods, which are not rob ust to the noisy nature of echocardiography . Thus, there is a need for high- quality echocardiography recordings with 50 − 70 frames per second (fps) in order to tackle the issues raised by the speckle tracking in deformation imaging [28]. As a result, the clinical usage of deformation imaging is limited. The activ e contour (snak e) is introduced by Kass et al. [29] that e volves iterativ ely to minimize the energy curves to e xtract the edges, lines, or boundaries in images. They are used in studies [30]–[32] to extract and track the endocardial boundary of the L V w all in echocardiography . Howe ver , the endocardial boundary is often discontinuous in echocardiography due to the high level of noise. These occlusions and indentations on the L V wall cause snake to fail at e xtracting to the true boundary [33]. In this study , in order to overcome the aforementioned limitations of the motion estimation algorithms, we use the Activ e Polynomials (APs) [34] that constrain the activ e con- tours to achiev e rob ust se gmentation and tracking of the endocardial boundary of the L V wall. In our previous study [34], we proposed a single-view MI detection approach by thresholding the maximum displacement of APs in A4C view echocardiography . Ho wev er, setting a fix ed threshold is nev er guaranteed to be optimal for decision-making. Moreover , in the literature, many studies ha ve also used single-view , high- quality , or simulated echocardiographic data [30]. Addition- ally , in cardiology , sev eral studies [35]–[37] diagnose MI using conv entional and Deep Learning methods. Ho wever , one major limitation is that they require large datasets for training. Therefore, the reliability and performance of the previously proposed methods may significantly v ary as the clinical data is usually scarce, in low quality/resolution, and subject to a high lev el of noise. Hence, our first objectiv e in this study is to improve the performance of our pre vious work [34] by bringing an intelligent diagnosis via Machine Learning ov er the hand-crafted features to surpass threshold-based diagnosis in single-view echocardiography . A common and major drawback of all the prior studies in this domain is that the proposed MI detection methods all rely on single-vie w , mostly over the apical 4-chamber vie w . Only certain segments can be analyzed on a single-view and this brings an ine vitable problem of missing the ongoing MI if the R WMA is not present on those se gments. In other words, regardless of their accurac y , if the segment(s) that show the abnormal motion is not visible on that particular echocardiog- raphy view , they are bound to fail the MI detection. Therefore, the second objecti ve of this study is to diagnose MI on the L V wall by using multi-view echocardiography , which includes apical 4-chamber (A4C) and apical 2-chamber (A2C) views, where all the chambers of the heart, and only left atrium and L V are visible, respecti vely as it is depicted in Fig. 1. In order to improv e robustness and generalization of MI diagnosis, in this study , we propose a multi-view Machine Learning (ML) approach over the maximum displacement features as depicted in Fig. 2. As the pioneer MI diagnosis study in the literature over multi-vie w echocardiography , we aim to determine the best ML approach for this purpose. Therefore, we perform an e xtensive set of comparative ev alu- ations among se veral ML methods including Decision Tree (DT), Random F orest (RF), k-Nearest Neighbour (k-NN), Support V ector Machine (SVM), and 1D-Con volutional Neural Networks (1D-CNN). Accordingly , the contributions of this study are summarized as follows: • W e improve the performance of threshold-based APs [34] ov er single-view echocardiography . Additionally , APs are adapted for the A2C vie w echocardiography for the first time. • W e propose a pioneer algorithm for MI diagnosis using multi-view echocardiography . Noting that the ground- truth labels of single-view and multi-view echocardiogra- phy are different, direct comparison is not viable. Hence, our study re veals the results of multi-view echocardiogra- phy for the first time to perform a reliable MI diagnosis by considering more myocardial segments in the analysis phase. • An extended benchmark dataset, HMC-QU 1 is created that includes 260 echocardiography recordings of 130 MI patients and healthy subjects from A4C and A2C vie ws. 1 The benchmark HMC-QU dataset is publicly shared with the research community at the repository https://www .kaggle.com/aysendegerli/hmcqu- dataset. JOURNAL OF L A T E X 3 Fig. 2: The diagram of the proposed MI detection approach using multi-view echocardiography . The endocardial boundary is first extracted by the APs method. Then, the defined myocardial segments are tracked through one-cardiac cycle to form the displacement curves. The maximum displacements are generated from each segment to define the features that are then concatenated and fed into the classifier to detect MI. The rest of the article is or ganized as follows. In Section II, we give the details of the proposed approach. In Section III-A, we introduce the HMC-QU dataset, and in Sections III-C and III-D we report the experimental results. Finally , we conclude the paper and suggest topics for future research in Section IV. I I . M E T H O D O L O G Y In this section, the proposed approach will be described in detail. As it is depicted in Fig. 2, in the first step, the endo- cardial boundary of the L V wall is extracted by APs. Then, the boundary is divided into myocardial segments from which the displacement curv es are generated. Lastly , the features are extracted from the displacements of each myocardial segment, which are then used as the input for the classifiers for MI diagnosis. A. Endocar dial Boundary Extr action by Active P olynomials Accurate e xtraction of the L V wall is crucial to obtain the true motion of the myocardium. In order to overcome the limi- tations of the activ e contours [29] in echocardiography , we use Activ e Polynomials [34] to extract the endocardial boundary of the L V wall. Echocardiography is usually subjected to a high lev el of noise, and during acquisition, some parts of the chamber walls might be missing or out of vie w . APs provide a robust and reliable segmentation and tracking of the L V wall, where their formation is illustrated in Fig. 4. A brief summary will be presented next and the details of the method can be found in [34]. In the first stage of the APs formation, the Ridge Polyno- mials (RPs) are formed on the L V wall. In echocardiography , the L V wall may partially be missing or in visible due to lo w quality . Thus, in the ev olvement process of the acti ve contours, the contour may escape from the chamber causing inaccurate segmentation of the endocardial boundary . Therefore, the RPs are first created to constrain the acti ve contours as illustrated in Fig. 3. In the second stage of the proposed method, we initialize an activ e contour from inside the chamber . The initial mask for the contour is located in the middle of the L V as Fig. 3: The comparison of the original and constrained activ e contours. In both A4C and A2C vie ws, the constrained acti ve contours can extract the endocardial boundary more accurately . JOURNAL OF L A T E X 4 Fig. 4: The APs method for the endocardial boundary of the L V wall extraction consists of three stages: 1) the RPs on the L V wall are formed in input echocardiography , 2) the acti ve contour is ev olved from inside of the L V and constrained by the RPs, and 3) the APs are formed by fitting 4 th − order polynomials on the ev olved activ e contour . a mini-version of the current frame’ s RPs. The aim is to ev olve an activ e contour to detect and e xtract the endocardial boundary of the L V wall. A typical edge detector is e xpressed as follows: lim z →∞ g ( z ) = 0 , (1) where g is a function with positiv e and decreasing values, z is an image, and the edges of z are detected at the locations where the gradient is zero. Howe ver , detecting the edges of images with rough and discontinuous objects is challenging with the gradient method since generally gradient is not zero on that particular edges. Thus, Chan-V ese [38] activ e contour method is utilized since its stopping criteria do not depend on the gradient. Therefore, it is suitable especially for echocardiography , where there are discontinuities (e ven though it has been minimized by RPs) and rough edges on the L V wall due to the high le vel of noise and acquisition. Once the activ e contour has con verged to the endocardial boundary , the APs can then be formed o ver the ev olved acti ve contour . As shown in Fig. 3, the acti ve contour may be noisy with sev ere discontinuities on the L V wall. In order to achieve a smooth endocardial boundary se gmentation, the e volv ed active contour is divided into two sections. The left part of the contour corresponds to the activ e contour points from start to apex, Fig. 5: The myocardial segments of A4C and A2C views echocardiography based on the 17 − segment model. whereas the right part is from apex to end. After the di vision, we compose 4 th − order smooth polynomials each of which is fitted to the equally distanced 9 points from both right and left parts to form APs that are the final form of the endocardial boundary . Thus, APs provide a robust and smooth endocardial boundary for MI diagnosis. B. Myocar dial Segment Displacements After extracting the endocardial boundary by APs in each frame of the echocardiography recordings, the boundary is tracked and its displacement is measured in one-cardiac cycle. In the diagnosis, the L V w all is segmented into 17 myocardial segments, which is the recommendation of the American Fig. 6: The myocardial se gment names and numbers are sho wn for both A4C and A2C views echocardiography at the top row . The b ull e ye’ s plot of the 17 − segment model is illustrated, where each color -coded and numbered segment corresponds to coronary arteries at the bottom ro w . JOURNAL OF L A T E X 5 Fig. 7: The displacement curves of A4C and A2C vie w echocardiography recordings of a patient, where the frames consist of one-cardiac cycle. Heart Association Writing Group on Myocardial Segmentation and Registration for Cardiac Imaging [39]. The myocardial segments on the L V wall for both A4C and A2C vie ws can be seen in Fig. 5 with a total of 12 distinct myocardial segments. It is recommended that segment − 17 should be removed if the wall motion or regional strain are analyzed using the 17 − segment model [40]. Thus, in the analysis, we hav e excluded segment − 17 shown as the white myocardial segment in the APs block in Fig. 4. Consequently , analyzing the motion of the 12 myocardial segments in multi-vie w echocardiography yields information regarding all the coronary arteries feeding the heart muscle since they cover most of the heart area as illustrated in Fig. 6. The myocardial segments are divided based on the length of the APs that are formed at the end of the endocardial boundary extraction process. The length of the APs are considered individually as previously explained left and right parts, where the length of the left part is represented as L , whereas the right part’ s length is R . Accordingly , in A2C and A4C vie ws, the length of apical myocardial segments are R/ 7 and L/ 7 , whereas other segments hav e 2 R/ 7 and 2 L/ 7 for right and left parts, respectively . The displacement curves are plotted for each myocardial segment in A2C and A4C views echocardiography . The dis- placements are calculated ov er one-cardiac c ycle echocardio- graphy recordings, where the reference frame is at time t = 0 . At each time instance, the displacement is defined as follo ws: D s κ ( t ) = 1 N N X i =1 q ( x i s κ ( t = 0) − x i s κ ( t )) 2 + ( y i s κ ( t = 0) − y i s κ ( t )) 2 , (2) where D s κ is the av erage displacement measure of the κ numbered myocardial segment s κ at a time instant t , N is the number of points equally taken on the myocardial segment, i.e., N = 5 in our implementation, and ( x, y ) is the coordinate of each point taken. Accordingly , at the reference frame, the displacement measurement is equal to zero. It is expected that the displacement measures of each segment would gradually increase from end-diastole to middle of the cycle; on the other hand, gradually decrease from middle of the cycle to end- systole as illustrated in Fig. 7. C. F eatur e Engineering In the feature engineering stage, we extract information from the displacement curves by quantitativ ely imitating the ev aluation of cardiologists. A cardiologist visually assesses the R WMA by correlating the infarction to the displacement measurement of a myocardial segment. Accordingly , the larger the displacement measure a myocardial se gment has, the less the chance of it being infarcted. Thus, the maximum displacement of each myocardial se gment is extracted as the features. Howe ver , a myocardial segment displacement cannot directly be compared since the displacements decrease gradually from valve to apical cap due to the structure of the heart. Therefore, the displacements of the myocardial se gments are normalized by dividing the maximum displacement of a segment with the minimum interv al between the segment and the other segment at the opposite side. For example, the interval between se gment − 3 and its opposite segment − 6 is greater than the interval between segment − 14 and its opposite segment − 16 . Thus, we bring each displacement to the same lev el for a fair comparison. The interv al measurement used in the normalization is defined as follows: I ( s κ ,s ε ) ( t ) = 1 N N X i =1 | x i s κ ( t ) − x i s ε ( t ) | + | y i s κ ( t ) − y i s ε ( t ) | , (3) where I ( s κ ,s ε ) is the av eraged Manhattan distance of N = 5 number of equally taken points on the tw o opposite se gments s κ , s ε at time t , and κ , ε are the segment numbers. Accord- ingly , we form the features of each myocardial segment as follows: f s κ = max ( D s κ ) min ( I ( s κ , s ε ) ) , (4) JOURNAL OF L A T E X 6 where f is the displacement feature of segment s κ that is the maximum displacement divided by the minimum interv al between its opposite segment s ε . For the displacement calcu- lation, we ha ve used the Euclidean distance, whereas, for the interval measurement, the Manhattan distance is utilized to scale the features into [0 , 1] more precisely as adapted by the threshold-based APs [34] approach. Consequently , in single- view echocardiography , where we only use whether A4C or A2C view echocardiography recording, we extract feature vectors Φ 1 , Φ 2 ∈ R 6 × 1 in one-cardiac cycle, respectiv ely defined as follows: Φ 1 =         f s 3 f s 9 f s 14 f s 16 f s 12 f s 6         , and Φ 2 =         f s 4 f s 10 f s 15 f s 13 f s 7 f s 1         , (5) where s denotes the numbered myocardial segment features as illustrated in Fig. 6 and calculated in Eq. (4). On the other hand, in multi-view echocardiography , we concatenate the feature vectors to form F =  Φ 1 T Φ 2 T  ∈ R 1 × 12 . D. MI Detection The MI detection is performed via binary classification task, where the extracted features are fed into se veral classifiers as follows: Support V ector Machine (SVM), k-Nearest Neighbors (k-NN), Decision Tree (DT), Random F orest (RF), and 1D- Con volutional Neural Networks (1D-CNN). The training is performed over K number of samples { f j train , L j train } K j =1 , where f and L are the data and ground-truths, respectiv ely . Support V ector Machines. The binary classification task via SVM is performed by separating the data with a hyperplane [41]. The best-fitting hyperplane is determined by maximizing the inter-class and minimizing the intra-class dif ferences. In order to impose non-linearity to SVM models, kernel-based methods are used to construct non-linear features that map the data into higher dimensions to perform an easier class separation. Thus, the performance of classification can be improv ed and the ov erfitting can be av oided. Decision T ree. The hierarchical structure of DT performs a classification task by feeding the data to nodes that are divided into branches for transferring the input to the most suitable class label [42]. The tree is formed by selecting the nodes as they are divided into branches, and whene ver the stopping criterion is satisfied, the final node is assigned to a class. DT models are suitable for small datasets, and computationally less expensi ve compared to other models used in this study . Random F orest. As an ensemble version of the DT , the RF prevents the overfitting issue that occurs due to the tight- fitting of the model to the training data. Overfitting causes the generalization capability of the model to deteriorate. Therefore, the RF model overcomes the ov erfitting issue by constructing individual trees by minimizing their correlation in the classification task. After the majority voting, the best model is selected to be used for the task. k-Nearest Neighbors. The k-NN method classifies data by assigning a sample to the same class as its k-nearest Fig. 8: The proposed 1D-CNN structure consists of two 1D- con volutional (1D-Con v) and two max-pooling (MaxP .) layers. The input, filter , and fully connected layer sizes are denoted as A , B , and D , respectiv ely . neighbours [43]. It is popular due to its robustness to noise and the simplicity of the algorithm. Moreover , it requires a few parameters to tune, which mak es the cross-v alidation process straightforward [44]. The performance of the k-NN method improv es as more data is used to train it. Howe ver , more training data increases its computational cost and memory consumption since k-NN stores the training data in order to calculate the distance between samples to classify a test sample. 1D-Con volutional Neural Netw orks. The most popular ML method during the last decade is Con volutional Neural Networks (CNNs) that are feed-forward models consisting of input, output, and hidden layers [45]. Their difference to Artificial Neural Networks is that con volution operations are performed in the hidden layers. In one-dimensional signal processing applications, 1D-CNNs are preferred due to their feasibility to one-dimensional con volution operations and low computational complexity compared to 2D-CNNs [45]. The 1D-CNN model proposed in this study maps the input feature, F to the corresponding class label, L : L ← − P ϑ,χ ( F ) . The first block ϑ ∈ { b j , w j } M j =1 consists of M = 2 number of 1D-conv olutional layers, Rectified Linear Unit (ReLU) activ ation function, and max-pooling layers by the size of 2 , respectiv ely . The second block χ consists of a fully connected layer , ReLU activ ation function, an output layer , and softmax activ ation function, respectiv ely . The filter and kernel sizes of the con volutional layers are presented in Section III-B. Accordingly , the block diagram of the proposed 1D-CNN is illustrated in Fig. 8, where its compact structure avoids ov erfitting. I I I . E X P E R I M E N T A L R E S U LT S In this section, we detail the HMC-QU dataset and report the experimental results for single-view and multi-vie w echocar- diography . A. HMC-QU Dataset The cardiologists of Hamad Medical Corporation, and re- searchers from Qatar University and T ampere University hav e compiled the HMC-QU dataset that includes 2D echocardiog- raphy recordings for MI detection. This benchmark dataset has been approv ed for usage by the local ethics board of the hospital in February 2019 . The dataset consists of 260 JOURNAL OF L A T E X 7 T ABLE I: The number of myocardial segments corresponds to the ground-truth labels of the HMC-Q U dataset. Myocardial Se gments MI P atients non-MI Subjects Segment − 1 29 101 Segment − 3 26 104 Segment − 4 29 101 Segment − 6 16 114 Segment − 7 40 90 Segment − 9 46 84 Segment − 10 31 99 Segment − 12 28 102 Segment − 13 47 83 Segment − 14 64 66 Segment − 15 53 77 Segment − 16 53 77 recordings from A2C and A4C vie ws of 130 subjects. The MI term indicates any sign of R WMA, whereas subjects without RMW A are labeled as non-MI in the dataset. All the MI patients had first-time acute MI and were admitted with ECG and cardiac enzymes evidence. The patients were treated with coronary angiography/angioplasty whose echocardiogra- phy recordings are taken within 24 hours after admission or before the operation. Non-MI subjects were not diagnosed with MI according to their echocardiography recordings, but they underwent a required health check in the hospital. The number of myocardial segments with respect to ground- truth labels are presented in T able I. The A4C view includes 80 MI and 50 non-MI recordings, whereas 68 MI and 62 non- MI recordings are from the A2C view . Accordingly , MI ratios are 61 . 54% and 52 . 3% in A4C and A2C views, respecti vely . The ratios differ from each other since only 60 patients hav e their MI visible in both views. Therefore, in multi-vie w echocardiography , the ground-truth labels correspond to 88 MI patients and 42 non-MI subjects, where the ground-truth labels are formed as MI if any of the vie ws depict R WMA, whereas non-MI if no sign of R WMA is visible in both views. Thus, the ov erall MI ratio is 67 . 69% in multi-view echocardiography . In T able II, a detailed ground-truth label formation with respect to views is presented. In each echocardiography recording, the myocardial seg- ments on the L V wall are categorized into five dif- ferent stages: 1 − normal or hyperkinesia, 2 − hypokinesia, 3 − akinesia, 4 − dyskinesia, and 5 − aneurysm as the sev erity of MI ascends, respectiv ely . In this study , we perform a binary classification task to simplify the problem. Therefore, we hav e downsized the ground-truth labels to 1 − non-MI (normal), and T ABLE II: The number of patients with respect to their corresponding ground-truth labels from A4C and A2C views. Ground-truths # of P atients A4C view A2C view MI MI 60 non-MI non-MI 42 MI non-MI 20 non-MI MI 8 (2 , 3 , 4 , 5) − MI. The ultrasound machines used for acquisition are Phillips and GE V i vid from GE Healthcare (United States). The spatial resolution of the echocardiography recordings varies from (422 × 636) to (768 × 1024) , and the temporal resolution is 25 frames per second (fps). B. Experimental Setup The detection models are ev aluated ov er the dataset in a stratified 5 -fold cross-validation scheme with a ratio of 80% training, and 20% test (unseen data) sets considering a balanced ratio of classes. The confusion matrices are formed by the elements: true positiv e ( T P ), true neg ativ e ( T N ), false positiv e ( F P ), and false negati ve ( F N ). Thus, the standard performance metrics are calculated as follows: S ensitiv ity = T P T P + F N , (6) where the sensitivity (recall) is the ratio of correctly detected MI patients to all MI patients in the dataset, S pecif icity = T N T N + F P , (7) where the specificity is the ratio of correctly classified non-MI subjects to all non-MI subjects in the dataset, P r ecision = T P T P + F P , (8) where the precision refers to the number of correctly detected MI patients over the total number of samples detected as positiv e class in the dataset, Accur acy = T P + T N T P + T N + F P + F N , (9) where the accuracy is the ratio of correctly detected samples in the dataset, F ( β ) = (1 + β 2 ) P r ecision × S ensitiv ity β 2 × P r ecision + S ensitiv ity , (10) where the F 1 − Score and F 2 − Score are calculated as the weighting parameter β = 1 and β = 2 , respecti vely . The F 1 − Score refers to the harmonic average of precision and sensitivity metrics. On the other hand, F 2 − Score emphasizes the sensiti vity metric with a higher β value. Consequently , the objecti ve of the detection phase is to maximize sensitivity with a preferable specificity to avoid missing MI patients. Moreov er , F 2 − Score is tar geted to be maximized with a reasonable F 1 − Score value. The implementation of the detection models is performed on Python using the T ensorFlo w library [46] and Scikit-learn library [47], whereas the feature engineering of the proposed method is implemented on MA TLAB version R 2019 a. For the experiments, we hav e used a PC with Intel® i 7 − 8665 U CPU 32 GB system memory , and a workstation with NV idia® GeForce R TX 2080 Ti GPU card 128 GB system memory . In the training phase of each classifier , we have performed a grid search o ver a 5 -fold cross-validation scheme that is an exhausti ve search of specified parameter values for each model in order to set the best hyper -parameters for the testing phase. The grid search is performed by the GridSear chCV function JOURNAL OF L A T E X 8 T ABLE III: A verage MI detection performance results (%) computed from 5-folds in single-vie w echocardiography . MI Ratios Model Sensiti vity Specificity Precision F 1 − Score F 2 − Score Accurac y A4C 61 . 54% DT 85 . 00 68 . 00 80 . 95 82 . 93 84 . 16 78 . 46 RF 86 . 25 84.00 89.61 87.90 86 . 90 85.38 SVM 82 . 50 70 . 00 81 . 48 81 . 99 82 . 29 77 . 69 k-NN 88.75 78 . 00 86 . 59 87 . 65 88.31 84 . 62 1D-CNN 83 . 75 78 . 00 85 . 90 84 . 81 84 . 17 81 . 54 APs [34] 86 . 25 77 . 08 86 . 25 86 . 25 86 . 25 82 . 81 A2C 52 . 30% DT 67 . 65 58 . 06 63 . 89 65 . 71 66 . 86 63 . 08 RF 66 . 18 77.42 76 . 27 70 . 87 67 . 98 71 . 54 SVM 76.47 66 . 13 71 . 23 73 . 76 75.36 71 . 54 k-NN 72 . 06 77.42 77.78 74.81 73 . 13 74.62 1D-CNN 64 . 71 75 . 81 74 . 58 69 . 29 66 . 47 70 . 00 APs [34] 69 . 12 59 . 68 65 . 28 67 . 14 68 . 31 64 . 62 of the Scikit-learn library . The best parameters are selected according to the highest F 1 − Score over the v alidation data, which is extracted from the training set of each fold. Thus, in the testing phase, the best parameters are set for each fold differently . Accordingly , we search the best parameters of the classifiers as follows: DT has searched the function of Gini impurity and entr opy for measuring the quality of a split, the maximum number of features that are selected for the best split is defined by the auto , log 2 , and squar e r oot of the number of features in the training set, the nodes are separated by the supported strategies that are set to random and best , and the performance of the model is e v aluated on the test set by checking the scoring of each standard performance metrics. RF classifier has bootstrap parameter set to false and true that indicates the data usage as b uilding the trees, the class weights are determined by balanced and balanced subsample mode, the quality of splits are measured by Gini impurity and entr opy functions, the maximum number of features that are selected for the best split is defined by the auto , log 2 , and the squar e r oot of the number of features in the training set, the warm start parameter is set to false and true , the number of trees in the forest is searched in [5 , 50] with a gap of 5 increasing at each step and the performance on the test set is ev aluated by checking the scoring of each performance metrics. SVM classifier has radial basis function (rbf) and linear kernel functions with the re gularization parameter searched in [1 , 1000] with a gap of × 10 increasing at each step. The kernel coefficients are determined in [10 − 1 , 10 − 6 ] with a decrease of 10 − 1 at each step and the scoring parameters for the testing phase are searched ov er each performance metric. k-NN decides the best algorithm for computing the nearest neighbors automatically or with brute-search, BallT ree, and KDT ree algorithms by weighting each neighborhood uniformly and in verse of their distance. The number of neighbors is determined in [5 , 30] with a gap of 5 increasing at each step, the metric used for computing the distances between neighbors are Manhattan and Euclidean , and the scoring parameters are selected as the performance metric with the highest scoring value. 1D-CNN is trained by Adam optimization algorithm [48] along with categorical cross-entrop y loss function with a learning rate of [10 − 1 , 10 − 7 ] decreasing at each step by 10 − 1 . The filter sizes of [4 , 8 , 12 , 16 , 24 , 32] and the kernel sizes of [3 , 5 , 7 , 9 , 11 , 13 , 15] are searched to train the model with [25 , 50 , 75 , 100] epochs by setting the scoring parameter to performance metric with the highest value. C. Single-view Experimental Results In Fig. 9, some examples of the APs formation can be depicted. The figure rev eals that the APs can successfully represent the true endocardial boundary e ven for lo w quality A4C and A2C vie ws. W e present the performance of each classifier in single- view (A4C or A2C vie w) echocardiography , indi vidually . The MI detection results are presented in T able III. In A4C view Fig. 9: A4C and A2C view frames for the endocardial bound- ary extraction process by the APs method. The sample images at the first row are subjected to artifacts, noise, or low contrast. JOURNAL OF L A T E X 9 T ABLE IV: A verage MI detection performance results (%) of state-of-the-art models, threshold-based APs, and our proposed approach computed from 5-folds in A4C vie w echocardiography . Method Model Sensiti vity Specificity Precision F 1 − Score F 2 − Score Accuracy Kusunose et al. [35] ResNet-50 74 . 68 54 . 84 68 . 95 71 . 59 73 . 38 66 . 25 DenseNet-121 73 . 39 63 . 52 73 . 48 72 . 46 72 . 79 69 . 38 InceptionResNetv2 72 . 34 65 . 05 73 . 56 72 . 26 72 . 15 69 . 38 Inception-v3 67 . 95 66 . 70 73 . 71 70 . 17 68 . 72 67 . 50 Xception 85 . 61 47 . 36 69 . 26 76 . 24 81 . 49 69 . 38 Kiranyaz et al. [34] APs [34] 86 . 25 77 . 08 86 . 25 86 . 25 86 . 25 82 . 81 Ours DT 85 . 00 68 . 00 80 . 95 82 . 93 84 . 16 78 . 46 RF 86 . 25 84.00 89.61 87.90 86 . 90 85.38 SVM 82 . 50 70 . 00 81 . 48 81 . 99 82 . 29 77 . 69 k-NN 88.75 78 . 00 86 . 59 87 . 65 88.31 84 . 62 1D-CNN 83 . 75 78 . 00 85 . 90 84 . 81 84 . 17 81 . 54 echocardiography , the prior approach with the threshold-based APs method [34] achie ves 86 . 25% sensitivity with a specificity lev el of 77 . 08% . The results indicate that imposing ML into the algorithm generally outperforms the threshold-based APs method in [34] by the classifiers utilized in this study . In the A4C vie w , the k-NN classifier achie ves the highest sensitivity lev el of 88 . 75% , whereas the highest specificity of 84% is obtained by the RF classifier . The performance of prior work in [34] for A2C vie w echocardiography is 69 . 12% sensitivity with 59 . 68% speci- ficity as shown in T able III. Once again, it was generally outperformed by the proposed approach with the ev aluated classifiers. The SVM classifier achiev ed the highest sensitivity lev el of 76 . 47% , whereas the highest F 1 − Score is obtained by the k-NN classifier with 74 . 81% in A2C vie w echocardiogra- phy . For further inv estigation, we compare our proposed ap- proach with an end-to-end solution using Deep Learning models. Accordingly , we adapted the method proposed by Kusunose et al. [35] that utilizes Deep Con volutional Neural Networks (DCNNs) to detect R WMAs over circular view echocardiography recordings. For the comparison, we used A4C view echocardiography recordings of HMC-QU dataset. From state-of-the-art DCNNs, we use the models ResNet-50 [49], Inception-v3 [50], DenseNet-121 [51], Xception [52], and InceptionResNetv2 [53] with transfer learning, where the networks are initialized with ImageNet datasetweights. W e resized the echocardiography frames and selected the three frames corresponding to a one-cardiac-cycle: end-diastole to end-systole and end-systole to end-diastole to use as the input of pre-trained models with (224 × 224 × 3) image sizes. Then, we performed a stratified 5 -fold cross-v alidation scheme and applied data augmentation using the Image Data Generator in K eras to the training set of each fold to av oid overfitting with an increased number of images up to 2000 . Accordingly , we compare the average MI detection performances between Kusunose et al. [35] (DCNNs), Kiranyaz et al. [34] (threshold- based APs), and our proposed method with Machine Learning- based APs, where the performance of each model is computed ov er a cross-v alidation scheme with the same train/test sets for each fold. Consequently , T able IV rev eals that our pro- posed approach outperforms the end-to-end network solution proposed by Kusunose et al. [35] and achie ves the highest performance for each performance metric. D. Multi-view Experimental Results In multi-view echocardiography , we mer ge the single- view information by concatenating the features as F =  Φ 1 T Φ 2 T  ∈ R 1 × 12 . Alternati vely , we utilize both of the single-view echocardiography results to detect MI in multi- view echocardiography by simply mer ging the A4C and A2C view detection results with the ” OR ” operator as a straight- forward solution. Accordingly , if either of the single-view detection outcomes is MI, the multi-view detection outcome will also be MI. In T able VI, the MI detection performances for multi-view echocardiography are presented. The results indicate that the proposed approach that concatenates the single-vie w features has proximate performance compared to the alternati ve solu- tion in multi-vie w MI detection. Accordingly , the proposed multi-view approach with the SVM classifier achieves an elegant sensiti vity level of 90 . 91% . On the other hand, the alternativ e approach with the RF classifier has the highest precision by 86 . 36% . Accordingly , the confusion matrices of the RF model that giv es the highest accuracy in multi-vie w echocardiography are sho wn in T able V, where both solutions T ABLE V: The confusion matrices of MI detection in multi- view echocardiography by the RF model, where the symbol ? indicates the concatenated features. (a) Multi-vie w ? Multi-view ? Predicted non-MI MI Ground T ruth non-MI 26 16 MI 11 77 (b) Multi-vie w Multi-view Predicted non-MI MI Ground T ruth non-MI 30 12 MI 12 76 JOURNAL OF L A T E X 10 T ABLE VI: A verage MI detection performance results (%) computed from 5-folds in multi-view echocardiography , where the symbol ? indicates the concatenated features. Echocardiography V iew Model Sensitivity Specificity Precision F 1 − Score F 2 − Score Accuracy Multi-view ? DT 84 . 09 64 . 29 83 . 15 83 . 62 83 . 90 77 . 69 RF 87 . 50 61 . 90 82 . 80 85.08 86 . 52 79.23 SVM 90.91 42 . 86 76 . 92 83 . 33 87.72 75 . 38 k-NN 86 . 36 59 . 52 81 . 72 83 . 98 85 . 39 77 . 69 1D-CNN 76 . 14 73.81 85.90 80 . 72 77 . 91 75 . 38 Multi-view DT 90.91 45 . 24 77 . 67 83 . 77 87 . 91 76 . 15 RF 86 . 36 71.43 86.36 86 . 36 86 . 36 81.54 SVM 88 . 64 47 . 62 78 . 00 82 . 98 86 . 28 75 . 38 k-NN 89 . 77 64 . 29 84 . 04 86.81 88.57 81.54 1D-CNN 84 . 19 64 . 29 83 . 15 83 . 62 83 . 90 77 . 69 are compared. The F 1 − Score performance metrics of each ML classifier is plotted in Fig. 10 for single-vie w (A4C and A2C) and multi-view echocardiography . As it can be seen from the figure, the detection of MI is successful with the proposed approach. Due to the different myocardial segments appearing on each vie w , a direct comparison of the experimental results of multi-vie w and single-vie w is not viable, e.g., consider the case that MI detection from a single-vie w will always fail if the infarcted se gment is not one of the se gments visible on that view . Therefore, the results of single-view indicate the MI detection performance over the infarcted se gments visible on that view only . Similarly , the results of multi-view indicate the MI detection performance over the infarcted segments visible on one of the views. Therefore, with fewer cases, the single-view performance may be higher than the one for multi-view . Hence, the reliable way to perform the diagnosis would be to use multi-view echocardiography that includes more information regarding the myocardial segment motion from both A4C and A2C views. Fig. 10: The F 1 − Scores of the ML classifiers are plotted for single-view (A4C and A2C) and multi-view (the proposed feature concatenation) echocardiography . E. Computational complexity The computational complexity of the proposed multi-view MI detection method is the total computational complexity that arises from each individual block depicted in Fig. 2. The endocardial boundary extraction, myocardial segment displacement, and feature engineering blocks of the method are from the prior work, where their computational complexities are detailed in [34]. Howe ver , the time elapsed for ex ecuting the algorithm is doubled since both A4C and A2C vie ws are used in this study . On the other hand, the MI detection stage has a computational complexity that differs with respect to the utilized classifiers. Accordingly , the classifiers hav e the computational complexities in the prediction phase as follows: DT as O ( V ) , RF as O ( V n tree ) , SVM as O ( V n sv ) , and k-NN as O ( V n train ) , where the length of the feature vector , number of trees, number of support vectors, and number of training samples are denoted as V , n tree , n sv , and n train , respectively . Furthermore, the con volutional layer computations of 1D-CNN are as follows: C = L X l =1 N ( l − 1) N l V ( l − 1) K 2 ( l − 1) + L − 1 X l =0 N ( l +1) N l ( K l + V l ) K 2 l + L − 1 X l =0 N ( l +1) N l K l ( K l + V l ) 2 , (11) where C in Eq. (11) is the multiplication operations of L number of layers at each back propagation iteration, N number of connections between layers, and K sized filter . Thus, 1D- CNN has the time complexity O ( C ) . T able VII shows the a verage time elapsed in seconds ( s ) during the inference of each step of the proposed algorithm in multi-vie w echocardiography that is illustrated in Fig. 2. Accordingly , the most time-consuming stage arises from APs, where around 60 seconds have passed for its ex ecution. On the other hand, the f astest block of the algorithm is MI detection with real-time execution. Ov erall, the proposed algorithm requires 61 . 0318 seconds on av erage to process A4C and A2C echocardiography views with one-cardiac cycle each JOURNAL OF L A T E X 11 T ABLE VII: The av erage time elapsed for e xecuting the algorithm stages in multi-vie w echocardiography . Algorithm Stage Proposed Methods Elapsed Time ( s ) Endocardial Boundary Extraction APs 59 . 0696 Myocardial Segment Displacement Maximum Displacements 0 . 0712 Feature Engineering Scaled Displacements 0 . 0120 MI Detection DT 3 . 023 × 10 − 6 RF 5 . 250 × 10 − 5 SVM 7 . 592 × 10 − 6 k-NN 3 . 725 × 10 − 5 1D-CNN 2 . 665 × 10 − 3 ( ≈ 30 − 50 frames in total). I V . C O N C L U S I O N S The early detection of MI is a crucial task to prevent further tissue damages or even death. In this study , we propose to detect MI over multi-view echocardiography by merging the information extracted from A4C and A2C views. Contrary to the recent studies proposed for single-view , this is the first study that accomplishes a multi-vie w MI detection for a reliable and rob ust diagnosis. Moreov er , this study shows that the threshold-based APs method in [34] can significantly be impro ved by using an ML-based approach even for single- view MI detection. The experimental results sho w that the detection performance has increased with the proposed ap- proach in single-view echocardiography by 2 . 50% and 7 . 35% for the sensitivity metric in A4C and A2C views, respectiv ely . Furthermore, in multi-view echocardiography , the proposed approach has achiev ed a sensitivity lev el of 90 . 91% and an F 2 − Score of 87 . 72% . The proposed method can be clinically used as an assisti ve tool to help cardiologists and technicians to prevent subjective and operator -dependent assessments by accurately measuring the L V myocardial displacements and plotting the color-coded myocardial segments. Finally , another major contribution of this study is the formation of the multi-view HMC-QU dataset that is publicly shared with the research community . W e plan to extend our approach to other views in order to detect MI due to the blockage of any coronary artery . W ith this accom- plishment, we will be able to identify the blocked arteries and also predict the location of the blockage simultaneously for the benefit of localizing the rev ascularization tar gets. R E F E R E N C E S [1] K. Thygesen, J. S. Alpert, and H. D. White, “Universal definition of myocardial infarction, ” J. Am. Coll. Cardiol. , vol. 50, no. 22, pp. 2173– 2195, 2007, http://dx.doi.org/10.1016/j.jacc.2007.09.011. [2] K. Thygesen, J. S. Alpert, A. S. Jaffe, M. L. Simoons, B. R. Chaitman, and H. D. White, “Third univ ersal definition of myocardial infarction, ” Cir culation , vol. 126, no. 16, pp. 2020–2035, 2012, http://dx.doi.org/10. 1161/CIR.0b013e31826e1058. [3] P . Grande, C. Christiansen, A. Pedersen, and M. S. Christensen, “Opti- mal diagnosis in acute myocardial infarction. a cost-effecti veness study . ” Cir culation , vol. 61, no. 4, pp. 723–728, 1980, http://dx.doi.org/10.1161/ 01.CIR.61.4.723. [4] L. Goldman, E. F . Cook, D. A. Brand, T . H. Lee, G. W . Rouan, M. C. W eisberg, D. Acampora, C. Stasiulewicz, J. W alshon, G. T erranov a et al. , “ A computer protocol to predict myocardial infarction in emergency department patients with chest pain, ” N. Engl. J. Med. , vol. 318, no. 13, pp. 797–803, 1988, http://dx.doi.org/10.1056/NEJM198803313181301. [5] A. Qamar, C. McPherson, J. Babb, L. Bernstein, M. W erdmann, D. Y a- sick, and S. Zarich, “The goldman algorithm revisited: prospecti ve ev aluation of a computer-deriv ed algorithm versus unaided physician judgment in suspected acute myocardial infarction, ” Am. Heart J. , vol. 138, no. 4, pp. 705–709, 1999, https://doi.org/10.1016/S0002- 8703(99) 70186- 9. [6] F . A. Masoudi, D. J. Magid, D. R. Vinson, A. J. Tricomi, E. E. L yons, L. Crounse, P . M. Ho, P . N. Peterson, and J. S. Rumsfeld, “Implications of the failure to identify high-risk electrocardiogram findings for the quality of care of patients with acute myocardial infarction: results of the emergenc y department quality in myocardial infarction (edqmi) study , ” Cir culation , v ol. 114, no. 15, pp. 1565–1571, 2006, https://doi.org/10. 1161/CIRCULA TIONAHA.106.62. [7] A. R. MacRae, P . A. Kavsak, V . Lustig, R. Bhargav a, R. V andersluis, G. E. P alomaki, M.-J. Y erna, and A. S. Jaffe, “ Assessing the require- ment for the 6-hour interval between specimens in the american heart association classification of myocardial infarction in epidemiology and clinical research studies, ” Clin. Chem. , vol. 52, no. 5, pp. 812–818, 2006, https://doi.org/10.1373/clinchem.2005.059550. [8] Y . S. Chatzizisis, V . L. Murthy , and S. D. Solomon, “Echocardiographic ev aluation of coronary artery disease, ” Cor on Artery Dis. , vol. 24, no. 7, pp. 613–623, 2013, https://doi.org/10.1097/MCA.0000000000000028. [9] J. S. Gottdiener, J. Bednarz, R. De vereux, J. Gardin, A. Klein, W . J. Manning, A. Morehead, D. Kitzman, J. Oh, M. Quinones et al. , “ American society of echocardiography recommendations for use of echocardiography in clinical trials: A report from the american society of echocardiography’ s guidelines and standards committee and the task force on echocardiography in clinical trials, ” J. Am. Soc. Echocar diogr . , vol. 17, no. 10, pp. 1086–1119, 2004, http://dx.doi.org/10.1016/j.echo. 2004.07.013. [10] I. Edler and K. Lindstr ¨ om, “The history of echocardiography , ” Ul- trasound in Med. & Biol. , vol. 30, no. 12, pp. 1565–1644, 2004, https://doi.org/10.1016/S0301- 5629(99)00056- 3. [11] T . R. Porter , S. L. Mulvagh, S. S. Abdelmoneim, H. Becher, J. T . Belcik, M. Bierig, J. Choy , N. Gaibazzi, L. D. Gillam, R. Janardhanan et al. , “Clinical applications of ultrasonic enhancing agents in echocar- diography: 2018 american society of echocardiography guidelines up- date, ” J . Am. Soc. Echocar diogr . , vol. 31, no. 3, pp. 241–274, 2018, https://doi.org/10.1016/j.echo.2017.11.013. [12] M. Alessandrini, A. Basarab, H. Liebgott, and O. Bernard, “Myocardial motion estimation from medical images using the monogenic signal, ” IEEE T rans. Image Pr ocess. , vol. 22, no. 3, pp. 1084–1095, 2013, = https://doi.org/10.1109/TIP .2012.2226903. [13] V . Behar, D. Adam, P . Lysyansk y , and Z. Friedman, “The combined ef- fect of nonlinear filtration and window size on the accuracy of tissue dis- placement estimation using detected echo signals, ” Ultrasonics , vol. 41, no. 9, pp. 743–753, 2004, https://doi.org/10.1016/j.ultras.2003.09.003. [14] M. Suhling, M. Arigovindan, C. Jansen, P . Hunziker , and M. Unser, “Myocardial motion analysis from b-mode echocardiograms, ” IEEE T rans. Image Pr ocess. , v ol. 14, no. 4, pp. 525–536, 2005, https://doi. org/10.1109/TIP .2004.838709. [15] B. D. Lucas, T . Kanade et al. , “ An iterati ve image registration technique with an application to stereo vision, ” in Proc. Intl. Joint Conf. on Artificial Intelligence (IJCAI) , vol. 2. V ancouver , British Columbia, 1981, pp. 674–679. [16] J. Bosch, S. Mitchell, B. Lelieveldt, F . Nijland, O. Kamp, M. Sonka, and J. Reiber , “ Automatic segmentation of echocardiographic sequences by activ e appearance motion models, ” IEEE T rans. Med. Imaging , v ol. 21, no. 11, pp. 1374–1383, 2002, https://doi.org/10.1109/TMI.2002.806427. [17] N. Ouzir , A. Basarab, H. Liebgott, B. Harbaoui, and J.-Y . T ourneret, “Motion estimation in echocardiography using sparse representation and dictionary learning, ” IEEE T rans. Image Pr ocess. , vol. 27, no. 1, pp. 64– 77, 2018, https://doi.org/10.1109/TIP .2017.2753406. [18] M. Ledesma-Carbayo, J. Kybic, M. Desco, A. Santos, M. Suhling, P . Hunziker, and M. Unser , “Spatio-temporal nonrigid registration for ul- trasound cardiac motion estimation, ” IEEE T rans. Med. Imaging , v ol. 24, no. 9, pp. 1113–1126, 2005, https://doi.org/10.1109/TMI.2005.852050. [19] V . Behar, D. Adam, P . Lysyansk y , and Z. Friedman, “Improving mo- tion estimation by accounting for local image distortion, ” Ultrasonics , vol. 43, no. 1, pp. 57–65, 2004, https://doi.org/10.1016/j.ultras.2004.02. 022. JOURNAL OF L A T E X 12 [20] B. H. Amundsen, T . Helle-V alle, T . Edvardsen, H. T orp, J. Crosby , E. L yseggen, A. St ˜ A¸ylen, H. Ihlen, J. A. Lima, O. A. Smiseth, and S. A. Sl ˜ A¸rdahl, “Noninv asive myocardial strain measurement by speckle tracking echocardiography: V alidation against sonomicrometry and tagged magnetic resonance imaging, ” J . Am. Coll. Car diol. , vol. 47, no. 4, pp. 789–793, 2006, https://doi.org/10.1016/j.jacc.2005.10.040. [21] B. Heyde, S. Cygan, H. F . Choi, B. Lesniak-Plewinska, D. Barbosa, A. Elen, P . Claus, D. Loeckx, K. Kaluzynski, and J. D’hooge, “Regional cardiac motion and strain estimation in three-dimensional echocardiog- raphy: a validation study in thick-walled univ entricular phantoms, ” IEEE T rans Ultr ason F err oelectr F req Contr ol , v ol. 59, no. 4, pp. 668–682, 2012, https://doi.org/10.1109/TUFFC.2012.2245. [22] R. A. Mukaddim, N. H. Meshram, C. C. Mitchell, and T . V arghese, “Hierarchical motion estimation with bayesian regularization in cardiac elastography: Simulation and in v iv o validation, ” IEEE T rans Ultrason F err oelectr F req Contr ol , v ol. 66, no. 11, pp. 1708–1722, 2019, https: //doi.org/10.1109/TUFFC.2019.2928546. [23] M. Bansal, L. Jeffriess, R. Leano, J. Mundy , and T . H. Marwick, “ Assessment of myocardial viability at dob utamine echocardiography by deformation analysis using tissue velocity and speckle-tracking, ” J ACC Cardiovasc. Ima ging , vol. 3, no. 2, pp. 121–131, 2010, https: //doi.org/10.1016/j.jcmg.2009.09.025. [24] F . Jamal, J. Strotmann, F . W eidemann, T . Kukulski, J. D’hooge, B. Bi- jnens, F . V an de W erf, I. De Scheerder , and G. R. Sutherland, “Nonin- vasi ve quantification of the contractile reserve of stunned myocardium by ultrasonic strain rate and strain, ” Cir culation , v ol. 104, no. 9, pp. 1059–1065, 2001, https://doi.org/10.1161/hc3501.093818. [25] M. Leitman et al. , “T wo-dimensional strain–a novel software for real- time quantitative echocardiographic assessment of myocardial function, ” J. Am. Soc. Echocardio gr . , vol. 17, no. 10, pp. 1021–1029, 2004, https: //doi.org/10.1016/j.echo.2004.06.019. [26] J. H. Rif fel, M. G. Keller , M. Aurich, Y . Sander, F . Andre, S. Giusca, F . aus dem Siepen, S. Seitz, C. Galuschky , G. Korosoglou et al. , “ Assessment of global longitudinal strain using standardized myocardial deformation imaging: a modality independent software approach, ” Clin. Res. Car diol. , vol. 104, no. 7, pp. 591–602, 2015, https://doi.org/10. 1007/s00392- 015- 0822- 7. [27] A. M. Shah and S. D. Solomon, “Myocardial deformation imaging: cur- rent status and future directions, ” Cir culation , vol. 125, no. 2, pp. e244– e248, 2012, https://doi.org/10.1161/CIRCULA TIONAHA.111.086348. [28] M. Dandel, H. Lehmkuhl, C. Knosall a, N. Suramelashvili, and R. Hetzer , “Strain and strain rate imaging by echocardiography-basic concepts and clinical applicability , ” Curr . Car diol. Rev . , v ol. 5, no. 2, pp. 133–148, 2009, https://doi.org/10.2174/157340309788166642. [29] M. Kass, A. Witkin, and D. T erzopoulos, “Snakes: Active contour models, ” Int. J. Comput. V ision. , vol. 1, no. 4, pp. 321–331, 1988, https://doi.org/10.1007/BF00133570. [30] G. Jacob, J. Noble, C. Behrenbruch, A. Kelion, and A. Banning, “ A shape-space-based approach to tracking myocardial borders and quantifying regional left-v entricular function applied in echocardiogra- phy , ” IEEE T rans. Med. Ima ging , vol. 21, no. 3, pp. 226–238, 2002, https://doi.org/10.1109/42.996341. [31] V . Chalana, D. Linker, D. Haynor, and Y . Kim, “ A multiple acti ve contour model for cardiac boundary detection on echocardiographic sequences, ” IEEE T rans. Med. Imaging , vol. 15, no. 3, pp. 290–298, 1996, https://doi.org/10.1109/42.500138. [32] M. Landgren, N. C. Overgaard, and A. Heyden, “Segmentation of the left heart ventricle in ultrasound images using a region based snake, ” in Med. Imaging 2013: Image Pr oc. , vol. 8669, 2013, p. 866945, https: //doi.org/10.1117/12.2001284. [33] A. De gerli, M. Zabihi, S. Kiran yaz, T . Hamid, R. Mazhar , R. Hamila, and M. Gabbouj, “Early detection of myocardial infarction in low-quality echocardiography , ” IEEE Access , vol. 9, pp. 34 442–34 453, 2021, https: //doi.org/10.1109/A CCESS.2021.3059595. [34] S. Kiranyaz, A. Degerli, T . Hamid, R. Mazhar, R. E. Fadil Ahmed, R. Abouhasera, M. Zabihi, J. Malik, R. Hamila, and M. Gabbouj, “Left ventricular wall motion estimation by active polynomials for acute myocardial infarction detection, ” IEEE Access , vol. 8, pp. 210 301– 210 317, 2020, https://doi.org/10.1109/A CCESS.2020.3038743. [35] K. Kusunose, T . Abe, A. Haga, D. Fukuda, H. Y amada, M. Harada, and M. Sata, “ A deep learning approach for assessment of regional wall motion abnormality from echocardiographic images, ” JA CC Cardiovasc. Imaging , vol. 13, no. 2, pp. 374–381, 2020, https://doi.org/10.1016/j. jcmg.2019.02.024. [36] H. A. Omar , A. Patra, J. S. Domingos, P . Leeson, and A. J. Noblel, “ Automated myocardial wall motion classification using handcrafted features vs a deep cnn-based mapping, ” in Proc. 40th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC . IEEE, 2018, pp. 3140–3143, https://doi.org/10.1109/EMBC.2018.8513063. [37] K. S. V idya, E. Ng, U. R. Acharya, S. M. Chou, R. San T an, and D. N. Ghista, “Computer-aided diagnosis of myocardial infarction using ultrasound images with dwt, glcm and hos methods: a comparative study , ” Comput. Biol. Med. , vol. 62, pp. 86–93, 2015, https://doi.or g/ 10.1016/j.compbiomed.2015.03.033. [38] T . F . Chan and L. A. V ese, “ Activ e contours without edges, ” IEEE T rans. Image Pr ocess. , v ol. 10, no. 2, pp. 266–277, 2001, https://doi.org/10. 1109/83.902291. [39] R. M. Lang, M. Bierig, R. B. De vereux, F . A. Flachskampf, E. Foster , P . A. Pellikka, M. H. Picard, M. J. Roman, J. Seward, J. S. Shanewise et al. , “Recommendations for chamber quantification: a report from the american society of echocardiography’ s guidelines and standards committee and the chamber quantification writing group, developed in conjunction with the european association of echocardiography , a branch of the european society of cardiology , ” J. Am. Soc. Ec hocar diogr . , vol. 18, no. 12, pp. 1440–1463, 2005, https://doi.org/10.1016/j.echo. 2005.10.005. [40] R. M. Lang, L. P . Badano, V . Mor-A vi, J. Afilalo, A. Armstrong, L. Ernande, F . A. Flachskampf, E. F oster , S. A. Goldstein, T . Kuznetsov a et al. , “Recommendations for cardiac chamber quantification by echocar- diography in adults: an update from the american society of echocar- diography and the european association of cardiovascular imaging, ” J. Am. Soc. Echocar diogr . , vol. 16, no. 3, pp. 233–271, 2015, https: //doi.org/10.1016/j.echo.2014.10.003. [41] N. Cristianini and J. Sha we-T aylor , An Introduction to Support V ector Machines and other kernel-based learning methods . Cambridge, U.K.: Cambridge Uni v . Press, 2000. [42] L. Breiman, J. Friedman, R. Olshen, and C. Stone, Classification and Re gr ession T rees. Statistics/Probability Series . Belmont, CA, USA: W adsworth Publishing Company , 1984. [43] K. Gowda and G. Krishna, “The condensed nearest neighbor rule using the concept of mutual nearest neighborhood (corresp.), ” IEEE Tr ans. Inf. Theory , vol. 25, no. 4, pp. 488–490, 1979, https://doi.or g/10.1109/TIT . 1979.1056066. [44] J. Goldberger , G. E. Hinton, S. Roweis, and R. R. Salakhutdinov , “Neighbourhood components analysis, ” Adv . Neural Inf. Pr ocess. Syst. , vol. 17, pp. 513–520, 2004, https://proceedings.neurips.cc/paper/2004/ file/42fe880812925e520249e808937738d2- Paper .pdf. [45] S. Kiranyaz, O. A vci, O. Abdeljaber, T . Ince, M. Gabbouj, and D. J. Inman, “1d con volutional neural netw orks and applications: A survey , ” Mech. Syst. Signal Pr ocess. , vol. 151, p. 107398, 2021, https://doi.org/ 10.1016/j.ymssp.2020.107398. [46] M. Abadi, A. Agarwal, P . Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Da vis, J. Dean, M. Devin et al. , “T ensorflow: Large-scale machine learning on heterogeneous distributed systems, ” arXiv preprint arXiv:1603.04467 , 2016. [47] F . Pedregosa, G. V aroquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P . Prettenhofer , R. W eiss, V . Dubourg, J. V an- derplas, A. Passos, D. Cournapeau, M. Brucher , M. Perrot, and ´ Edouard Duchesnay , “Scikit-learn: Machine learning in Python, ” J. Mach. Learn. Res. , v ol. 12, no. 84, pp. 2825–2830, 2011, http://jmlr .org/papers/v12/ pedregosa11a.html. [48] D. P . Kingma and J. Ba, “ Adam: A method for stochastic optimization, ” in 3rd International Conference on Learning Repr esentations, ICLR 2015, San Die go, CA, USA, May 7-9, 2015, Conference T rack Pr oceedings , Y . Bengio and Y . LeCun, Eds., 2015. [Online]. A vailable: http://arxiv .or g/abs/1412.6980 [49] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition, ” in Pr oc. IEEE Conf. Comput. V is.P attern Recognit. (CVPR) , 2016, pp. 770–778, https://doi.org/10.1109/CVPR.2016.90. [50] C. Szegedy , V . V anhoucke, S. Ioffe, J. Shlens, and Z. W ojna, “Rethinking the inception architecture for computer vision, ” in Pr oc. IEEE Conf. Comput. V is.P attern Recognit. (CVPR) , 2016, pp. 2818–2826, https:// doi.org/10.1109/CVPR.2016.308. [51] G. Huang, Z. Liu, L. V an Der Maaten, and K. Q. W einberger , “Densely connected conv olutional networks, ” in Pr oc. IEEE Conf. Comput. V is.P attern Recognit. (CVPR) , 2017, pp. 2261–2269, https://doi.org/10. 1109/CVPR.2017.243. [52] F . Chollet, “Xception: Deep learning with depthwise separable con vo- lutions, ” in Pr oc. IEEE Conf. Comput. V is.P attern Recognit. (CVPR) , 2017, pp. 1800–1807, https://doi.org/10.1109/CVPR.2017.195. [53] C. Szegedy , S. Ioffe, V . V anhoucke, and A. Alemi, “Inception-v4, inception-resnet and the impact of residual connections on learning, ” in Pr oc. of the AAAI Conf . on Artificial Intelligence , vol. 31, no. 1, 2017, https://doi.org/10.1609/aaai.v31i1.11231.