Deep learning-based myocardial infarction detection using echocardiography

Abstract

Myocardial infarction (MI) is a life-threatening condition caused by reduced oxygen supply to the heart muscle due to blockage of the coronary arteries. Delayed or missed diagnosis increases the risk of mortality and heart failure, making early detection critical. Echocardiography is a non-invasive imaging technique that uses real-time ultrasound to examine the heart’s structure and function, including the evaluation of coronary artery disease and detection of regional wall motion abnormalities linked to MI. However, current approaches encounter limitations such as susceptibility to noise, restricted motion analysis capabilities, and significant computational demands. Hence, this work proposes EfficientNet_SpinalNet (Efficient_SpinalNet) for MI detection using echocardiography video. The proposed model utilizes the Hamad Medical Corporation Heart Hospital & Qatar University (HMC-QC) dataset and the Cardiac Acquisitions for Multi-structure Ultrasound Segmentation (CAMUS) dataset. The process begins by retrieving a video sample from the dataset, which is then decomposed into individual image frames. A hybrid filtering approach, integrating both median and Gaussian techniques, is first applied to the frames to suppress noise and enhance image quality. Following this preprocessing step, the Left Ventricle (LV) wall is fully extracted using the Fuzzy Local Information C-Means Clustering (FLICM) method, which further enables precise identification of the endocardial border. Following segmentation, displacement metrics and area variation curves are computed and passed into the feature extraction module. The extracted features, along with the displacement and area data, are then utilized for MI detection using the Efficient_SpinalNet architecture, which synergistically merges EfficientNet-B3-attn-2 with SpinalNet for enhanced diagnostic accuracy. Moreover, the experimental result reveals that the Efficient_SpinalNet functioned efficiently, attaining accuracy, sensitivity, and specificity values of 91.19%, 92.01%, and 92.435%. These results indicate that Efficient_SpinalNet is a reliable and efficient approach for real-time MI detection, offering potential improvements in clinical decision-making and patient outcomes.

Introduction

Cardiovascular diseases (CVDs) rank among the top causes of death globally. CVDs refer to cardiovascular and blood vessel illnesses that include angina, arrhythmia, and MI. Moreover, MI is the most dangerous and prevalent CVD, with symptoms such as sudden onset and fast alterations that endanger patients' lives [1]. A significant number of deaths worldwide are caused by myocardial infarction (MI). About 4 million individuals in the United States alone experience cardiac pain yearly [2]. MI is a life-threatening heart disorder resulting from blockage in the coronary arteries, which limits blood flow to cardiac tissue. Infarction is defined as tissue death due to inadequate blood supply [3]. Myocardial necrosis is irreversible and can result in permanent cardiac damage [4]. MI symptoms include extreme chest discomfort, breathing difficulties, and rapid heartbeats, which necessitate immediate therapy, such as endoscopic myocardial recanalization or bypass grafting of the coronary arteries. Precise assessment of the global and regional motion of the Left Ventricle (LV) of the heart is required for the early diagnosis of MI [5]. Early indications of MI are typically reflected in variations of the ventricular wall motion. As the condition progresses, the abnormalities are classified as hypokinesia, dyskinesia, and akinesia. In modern imaging methods, these irregularities are difficult to detect because the final interpretation is highly subjective [1]. Such conditions can significantly affect the patient’s cardiac function and overall health adversely [6].

Cardiac imaging tools have evolved over the past few years and produce extensive and precise information, such as real-time images of the heart valves and chambers. These recordings allow cardiologists to detect Regional Wall Motion Abnormalities (RWMA) in the LV, which are associated with MI [7]. Conventional heart attack diagnosis schemes rely primarily on electrocardiography (ECG) analysis, which measures the electrical activity of the heart [8]. Although ECG can help in diagnosing MI, the symptoms are often slight in the early stage, and the findings may overlap with myocardial ischemia. Echocardiography, on the other hand, can be used to assess the heart functions and detect RWMA [1]. Echocardiography is the most widely utilized tool for diagnosing MI due to its non-invasive nature, high safety and performance, low cost and easy availability [9]. Echocardiography functions on the principles of ultrasound imaging, capturing images of cardiac structures from multiple perspectives by varying the probe orientation. Echocardiography devices can offer a detailed view of the chambers of the heart and their walls, thus assisting in identifying the early signs of MI through RWMA [1]. Meanwhile, the accuracy of the detection is influenced by the sonographer's experience [5]. As a result, there is a critical demand for an automatic, more robust, and precise diagnostic method to address this problem [10]. This motivates scientists and cardiac specialists to create novel techniques that seek to identify and evaluate cardiac problems based on this data [7].

In the past few years, Machine Learning (ML) has become recognized as an efficient and precise technology that aids in the resolution of complex medical tasks such as MI [10]. Moreover, Artificial Intelligence (AI) methods [11], like Deep Learning (DL), have demonstrated high processing power. DL has been frequently employed for video dataset-based multi-classification. It can perform a broad range of analyses, including visual quality evaluation, view categorization, boundary separation, and disease detection. Recurrent Neural Networks (RNNs) are effective for forecasting and categorizing sequential information such as waveform signals, video, and natural language. Furthermore, the DL method, like Long Short Term Memory (LSTM), is widely used for multiple classifications of video images in massive video datasets with excellent precision [5]. Other DL methods, like Convolutional Neural Network (CNN), learn the features without the need for mining the unique characteristics. The precision of image categorization schemes is determined by the fundamental concepts of Artificial Intelligence (AI). The detection efficiency is primarily enhanced by various deep Convolutional networks (ConvNet). This is done by extracting the improved features and identifying the images using the inbuilt Convolutional networks [12]. The Bidirectional Long Short-Term Memory (B-LSTM), Multilead CNN (MLCNN), DL-based Least Square Support Vector Machine (DL-LSSVM), and other techniques are also utilized for MI identification [13].

Literature review

Kiranyaz, S., et al. [5] developed the active polynomial model for MI detection. It provided the clinicians with increased viewing capabilities of echo images via the color-coded segments. This model showed a high recall while reducing the number of parameters. Nonetheless, it did not depend on real-time data and failed to assess each section of the heart independently. Mohan, H.M., et al. [12]. devised a Faster Region-based Convolution Neural Network (RCNN) for real-time MI detection. It was a sophisticated non-invasive algorithm for identifying MI. This model needed minimal computing cost and resulted in the least amount of loss. However, it did not utilize advanced data augmentation techniques to boost the performance. Meanwhile, Degerli, A., et al. [1] introduced the Support Vector Machine (SVM) based early detection model for detecting MI. The process was lightweight, power-efficient, and energy-efficient. This approach required the largest collection of data for MI detection of both males and females of various ages to attain better performance. Mohan, H.M., et al. [14] devised a Deep Neural Network (DNN)-based real-time MI detection model. This approach was dependable and did not necessitate a large amount of memory. Nonetheless, this strategy demanded greater processing power.

Hamila, O., et al. [7]. developed the 3-Dimensional CNN (3D CNN) for video segmentation and MI detection. The model followed a pipeline structure, using a 2D CNN in the preprocessing stage to segment the LV chamber from apical four-chamber (A4C) data. This approach was dependable and did not necessitate more memory, but it required a lot of computer power. Lin, X., et al. [15] developed the LSTM-Unet for detecting the abnormalities in regional wall motion and MI. This model calculated the dimensions of the heart chambers and the LV ejection fraction automatically. Furthermore, it was efficient and required minimal training time. Nonetheless, it encountered the black box issue and failed to identify the degree of movement variations. Uchiyama, R., et al. [4] devised the End-to-end CNN model for MI detection. This model learned waveform properties of an unprocessed raw ECG image obtained from 12-lead ECG signals. This method did not necessitate any complicated preprocessing and it was very generalizable. However, it encountered a class imbalance issue and generated misclassified outcomes. Li, W., et al. [10] devised the Single-Lead ECG data -Generative Adversarial Networks (SLC-GAN) for MI detection. Here, issues such as insufficient data and class imbalance were effectively mitigated using the SLC-GAN approach. However, it failed to construct a more robust generative model that was capable of dealing with diverse forms of medical information. Bulbule, S. and Soma, S., et al. [16] introduced the parallel SqueezeNet (PSq-Net) for the detection of MI from echocardiography videos. By leveraging a parallel architecture of multiple SqueezeNet branches, the model effectively captured diverse spatial features from echocardiographic images, enhancing its sensitivity to subtle myocardial abnormalities. However, the parallel design of the model introduced a redundancy in feature extraction, potentially leading to increased training time and complexity in model optimization. Mingxi Han, et al. [17] introduced MSRC-TransBLSTM, a deep learning-based hierarchical hybrid model for automatic myocardial infarction (MI) detection. This model employed a hierarchical modeling strategy that combined spatial and temporal features for enhancing both local and global representations. This model effectively captured complex ECG signal patterns. However, a lack of validation on large-scale datasets restricts the model’s overall generalization capability. Chen, Y., et al. [18] proposed a multi-domain feature fusion convolutional neural network (MFF–CNN) that integrated time-domain, frequency-domain, and time–frequency domain features of ECG signals for automatic MI detection and localization. The model delivered robust outcomes, demonstrating the benefit of multi-domain feature fusion for improving generalization in complex medical diagnostic signals. However, using only Lead II data for training and testing could limit its ability to represent frequency and time–frequency interactions across all 12 leads.

Problem statement and objective

Despite significant advancements in automated MI detection, current approaches face several limitations. Most methods struggle to capture fine-grained temporal and spatial variations in cardiac motion, which are essential for accurate diagnosis. Many existing approaches are computationally intensive, reducing their suitability for deployment in real-time clinical settings. Furthermore, challenges such as noisy echocardiography data, limited or imbalanced datasets, and variability in patient conditions reduce the robustness and generalizability of current techniques. DL–based approaches often act as black boxes, limiting clinical interpretability and failing to quantify important functional variations in the heart. To address these limitations, this study aims to design an efficient, accurate, and robust framework for myocardial infarction detection using echocardiography.

Contributions

This study focuses on the development of Efficient_SpinalNet for detecting MI through echocardiography. The key contributions of this paper are outlined as follows,

• Proposed Efficient_SpinalNet for MI detection using echocardiography: This study focuses on identifying MI through echocardiography. The methodology involves segmenting the complete LV wall and its endocardial boundary using the FLICM algorithm. From these segmented regions, displacement profiles and area variation curves are derived. These outputs, along with extracted features, are then utilized for MI classification using the proposed Efficient_SpinalNet architecture, a hybrid model that integrates EfficientNet-B3 with an attention-enhanced mechanism and SpinalNet to improve diagnostic precision.

Proposed Efficient_SpinalNet for MI detection using echocardiography

This study aims to identify myocardial infarction (MI) using the proposed Efficient_SpinalNet architecture, whose overall framework is illustrated in Fig. 1. The workflow starts with acquiring input video data from a selected dataset. After that, video frame extraction is carried out to convert a video into a number of image frames, and the image frame is fed to the preprocessing module. Preprocessing is performed using a combination of a median filter [19] and a Gaussian filter [20], which effectively reduces speckle noise while preserving important structural boundaries, thereby improving overall image clarity. Next, the LV wall is segmented using the FLICM [21] algorithm, which is well-suited for ultrasound data due to its ability to incorporate local spatial information and produce smooth, accurate segmentations. The same algorithm is subsequently applied to segment the endocardial boundary, enabling precise identification of the inner LV contour. The segmented results are used to derive displacement and area variation curves, which help characterize ventricular motion patterns associated with MI. From endocardial boundary segmentation, feature extraction is performed, where the features, like LOOP [22], LBP [23], and statistical features [24], like mean, variance, standard deviation, kurtosis, and energy, are extracted to provide complementary texture and shape characteristics. The extracted features, along with the displacement and segmented area curves, are subsequently input into the MI detection module. Finally, MI detection is executed using the proposed hybrid deep learning Efficient-SpinalNet, which is obtained by the combination of EfficientNet-B3-attn-2 [25] and SpinalNet [26]. This hybrid network enhances feature representation while maintaining computational efficiency, thereby enabling accurate and robust MI classification.

Fig. 1.

Framework of efficient_spinalnet for MI detection

Video acquisition

An input video is gained from a database, which is formulated as,

1

where, symbolizes the video, and n is the video count.

Video frame extraction

The video frame extraction process involves splitting the acquired input video into a sequence of frames, and is described as,

2

where, denotes the frames per second.  signifies the number of frames, and the frame is taken from the is depicted as .

Pre-processing

The extracted frame is subjected to a pre-processing step, where unwanted things are removed. In this stage, image enhancement is performed using a hybrid filtering strategy that integrates both median and Gaussian filters. The formulation of this preprocessing technique is presented as follows:

3

In this formulation, F denotes the result obtained after preprocessing, corresponds to the output generated by the Gaussian filter, ϕ refers to a fixed scalar value used in the computation, and represents the outcome of the median filtering process.

Gaussian filter

Gaussian filter [20] is a filter that includes the impulse response of a Gaussian function, and it is minimizes the rise and fall time of the input. The Gaussian filter is a better integration of suppression of higher frequencies while reducing the spatial spread. The main advantages of this filter is effective in smoothing out noise and blurring, and is mathematically expressed as,

4

where, indicates the size of the filter, and indicates variance.

Median filter

The median filter is a nonlinear, statistics-based method used for noise reduction in images. It maintains important image details by replacing each pixel’s intensity with the median value of its neighboring pixels. This method reduces image noise and helps sharpen blurred edges. It works by arranging the pixel values within a local window in order and replacing the target pixel with the median (middle) value. A mathematical expression of the median filter is given below,

5

where, indicates the gray level of a pixel in the input , and R signifies the sub-image window.

Segmentation of the entire Left Ventricle (LV)

The LV is segmented by FLICM using the pre-processed output F. An output of the LV segmentation is indicated as, , and the functioning of FLICM is elaborated below.

FLICM

The robust framework based on Fuzzy C-Means (FCM) is introduced in the FLICM technique [21]. The method incorporates both local spatial information and gray-level fuzzy local similarity within its objective function, which is defined as,

6

where, N denotes the neighboring window, g indicates the cluster prototype, b indicates the fuzzyfication parameter, signifies the neighboring pixel, and represents the fuzzy factor. To ensure that reaches a significant local minimum, two essential conditions must be satisfied, each derived from the parameters and . These conditions are demonstrated as follows:

7

8

The FLICM technique is performed based on the subsequent steps,

Stage 1: Initially, set the g, b, and stopping condition α.

Stage 2: The fuzzy division matrix is initiated randomly.

Stage 3: Set the loop counter b is equal to zero.

Stage 4: Estimate the cluster prototypes based on the expression (7).

Stage 5: Compute the value of membership utilizing an expression (8).

Stage 6: Stop if , else set i=i+1 go to stage 4.

Once the algorithm converges, a defuzzification step is applied to transform the fuzzy partition matrix into a crisp segmentation. During this process, each pixel is assigned to the class with the highest membership value.

9

where, denotes the pixel-based class. This technique is used to achieve crisp segmented images, where the output of FLICM is represented as and is processed robustly, and is simple for computation.

Endocardial boundary segmentation

Following the segmentation of the LV, the next phase involves isolating the endocardial boundary to facilitate myocardial infarction (MI) detection. This step focuses on capturing the boundary's positional shifts, which are indicative of MI-related motion abnormalities. The segmented LV output serves as the foundation for this boundary extraction process. In this segmentation, FLICM is also applied, and the output generated is described by .

Find the displacement curve

After segmentation, the output is utilized to find the displacement curve [1], where boundary segment displacements are estimated using an echocardiography is expressed as,

10

where, y,k both indicate pixel coordinates, t denote the current frame, and tr represent the reference frame. Then, displacement segmentation of every frame is determined as,

11

where, indicates the displacement curve, and the total features are denoted as N. From the endocardial boundary of the LV wall, every segment extracts assured motion features. In addition, a similar process is applied to the segmentation of center points. where is one point on every segmentation to retrieve its motion via the overall frames. By determining the centroid points of six segmented regions, the full LV wall can be reconstructed. The segment area is denoted as , which designates the number of pixels enclosed in a segment.

Feature extraction

The feature extraction model acquires the segmented output , where a diverse set of descriptors is derived. These include texture-based patterns like LOOP and LBP, which effectively capture local intensity variations and micro-texture patterns associated with myocardial wall motion abnormalities, along with statistical measures such as mean, variance, standard deviation, kurtosis, and energy, which describe the global intensity distribution and underlying structural characteristics of the region of interest. These features are robust to noise and well-suited for handling heterogeneous echocardiography data. The feature extraction is utilized to decrease the dimension of the data for faster and more effective performance, and the output is denoted as C.

LOOP

LOOP [22] is presented as the non-linear amalgamation of a Local Binary Pattern (LBP) and Local Directional Pattern (LDP), which conquer the limitations of both the methods while conserving their strengths, LOOP is illustrated below,

12

where,

13

 

where, signifies the center pixel intensity, designates the neighborhood pixel, symbolizes the intensity of pixel, represents the LOOP-based feature extraction, and denotes the rank of magnitude. Thus, LOOP encodes a rotation of invariance into a major formulation. As a result, it reduces computational complexity while improving classification accuracy.

LBP

The LBP [23]operates based on neighboring pixels, and it gives an 8-bit code to every pixel depending on neighboring pixels around a central. Consider the pixel at , and the mathematical form of LBP is expressed as,

14

where, q indicates the neighbors of a central pixel.

15

After extracting LOOP and LBP, and concatenating both the texture features, we get _.

Statistical features

The output of texture features is forwarded to the extraction of the statistical features. The statistical measurement is based on the first order of pixels, which does not analyze the neighbouring pixels and is estimated directly from an image’s histogram. First-order statistical measurements, denoted by G include standard deviation, mean, energy, variance, and kurtosis, and these are elaborated below.

Mean

The mean represents the central tendency of a data distribution and is computed from a specific image region, and is formulated as,

16

where, the possibility of happening is denoted as , u indicates grey level in the image, indicates the mean, and the overall grey level count a is denoted as O.

Variance

Variance measures the extent to which pixel intensities deviate from the mean gray-level value in an image, and is formulated as,

17

where, indicates the variance.

Standard deviation

Standard deviation serves as a statistical measure that reflects how data points are distributed in relation to the mean. A smaller value indicates that the data is tightly clustered around the average, suggesting greater uniformity, while a larger value points to wider variation, implying increased diversity within the dataset. In image analysis, this metric helps assess the degree of texture consistency or variation, and is mathematically represented as follows:

18

where, indicates the variance.

Kurtosis

It is known as the fourth normalized moment and describes how a distribution’s shape differs in terms of sharpness or flatness relative to a Gaussian distribution,

19

where, reveals the kurtosis.

Energy

Energy is computed by the summation of the square region, and is illustrated as,

20

Here, refers to a pixel pair based on the Gray Level Co-occurrence Matrix (GLCM) N, and represents the computed energy value.

The overall statistical feature is denoted as,

21

where, indicates mean, indicates the variance, represents the standard deviation, symbolizes kurtosis, and indicates energy.

MI detection using hybrid deep learning

The MI detection is done by Efficient_SpinalNet and is developed using the EfficientNet-B3-Attn-2 [25] and SpinalNet [26]. The work evaluates EfficientNet-B3-attn-2, an enhanced EfficientNet-B3 model that incorporates an attention module to improve feature learning. The proposed hybrid architecture is composed of three distinct components: EfficientNet-B3-attn-2, the MI processing layer, and SpinalNet. Initially, the displacement curve, denoted as is input into the EfficientNet-B3-attn-2 module. The resulting output, labeled , along with statistical features represented as G are then passed to the MI layer. Within this layer, a fusion mechanism integrates both inputs, and the combined representation is processed through a regression-based modeling approach to facilitate myocardial infarction detection. Here, the Fractional Concept (FC) [27] is used for improving the regression modelling. The FC has been used to boost the computational operation of the various techniques and reduce the information loss. The output of the MI layer and the segment area curve is fed to the SpinalNet [26]. After completing the SpinalNet operation, the final output of the Efficient_SpinalNet is , which represents the MI detected output. Figure 2 represents the structure of the Efficient_SpinalNet.

Fig. 2.

Structure of the proposed Efficient_SpinalNet

EfficientNet-B3-attn-2

The displacement curve output is subjected to the EfficientNet-B3 with attention. Here, the EfficientNet-B3-attn-2 model [25] is based on the pre-trained EfficientNet-B3 CNN enhanced with an attention mechanism. In the EfficientNet-B3-attn-2 model, some convolutional layers are presented for processing. The convolution layer is incorporated with other kinds of layers, named as, normalization layer, pooling layer, and activation layer, that assist the method's approximation of non-linear functions.

The initial convolution layer retrieves abstract features from the displacement curve. In the consequent layer of a CNN, every neuron is associated with a similar region of a previous layer. The first convolution layer extracts high-level features from the displacement curve. In subsequent CNN layers, each neuron connects to a localized region of the preceding layer, forming a feature representation that can be viewed as a three-dimensional curve describing the receptive field.

During CNN execution, the extracted features are converted into a vector using two operations: flattening, which stacks feature maps into a single vector, and pooling, which aggregates features through operations such as averaging and summation.

Global Average Pooling (GAP) performs simple averaging across features with uniform weighting, without emphasizing any particular receptive region of the input curve. The performance of GAP is deliberated in the given expression,

22

where, is indicated as the input layer of GAP operation, denotes the output of GAP operation, signifies the features included in the layer , and Q × Q represents the number of feature vectors, and the attention process is carried out using a weighted averaging formulation given below,

23

Commonly, the attention scheme provides the neural network of the ability to concentrate on a subset of a feature vector by providing its various weights. The attention layer is expressed as,

24

25

Here, y is the input vector, T corresponds to the feature map, and w denotes the attention vector ranging between 0 and 1. The attention network is defined with parameter θ, and Θ indicates element-wise multiplication. The model output is represented as , while Fig. 3 shows the structure of EfficientNet.

Fig. 3.

Structure of EffecientNet-B3-Attn-2

MI layer

The outputs from EfficientNet-B3-attn-2, denoted as and the statistical feature set, labelled G are directed to the MI Layer. Within this layer, a fusion mechanism integrates both data streams, and the combined representation is processed using a regression-based approach to facilitate myocardial infarction detection. To enhance the regression model, the FC [27] method is utilized. The output of the fusion layer is denoted as that is forwarded to the SpinalNet model. The operation of the MI layer is expressed as,

26

where, indicates a feature vector, n denotes a number of vectors, U and is output from the interval.

To minimize the information loss, apply FC [27],

27

Then, the equation became,

28

where, χ represents the constant, and describes the output of the MI layer.

SpinalNet

The MI layer output and segmentation curve are forwarded to the SpinalNet [26]. SpinalNet is a specialized type of Artificial Neural Network (ANN) based on the structural and functional characteristics of the human spinal cord. It processes input data in a sequential, layered fashion, simulating the progressive transmission of signals across spinal segments. The architecture evaluates both localized and aggregated responses. Each segment generates a distinct local output, while an expanded version of the input contributes to a broader global representation. The network's parameters are optimized using training data to adjust its internal weights. In its conventional designThe architecture of SpinalNet comprises three main functional components: an input layer for segmented data, an intermediate processing unit, and an output layer responsible for final prediction. In the neural network (NN), the initial layer does not include an intermediate split; however, subsequent intermediate splits contain two elements: the output of the previous intermediary split and the current output of the standard input segment. These intermediate outputs are weighted and passed to the output split of each layer, where they are aggregated. The overall two inputs exist in every hidden layer of the intermediate piece. The output of the neutron in a hidden layer is expressed as,

29

where the activation function is denoted as ρ, r is the number of inputs, represents the weight, and indicates the threshold value, and Fig. 4 depicts the SpinalNet. The SpinalNet output is denoted as , which is the final output of MI detection.

Fig. 4.

SpinalNet structure

Results

The efficiency of Efficient_SpinalNet for MI detection using echocardiography is examined in this section. Moreover, this section provides an elaboration of experimental settings, experimental results, and a description of the dataset, performance validation, and comparative discussion.

Experimental set-up

The implementation of the devised Efficient_SpinalNet for MI detection using echocardiography is executed on the Python TensorFlow platform using the HMC-QC dataset [28] and CAMUS dataset [29]. A summary of the experimental parameters is provided in Table 1.

Table 1.

Experimental parameters

Parameters	Values
Efficient_SpinalNet
Epochs	100
Batch Size	64
Optimizer	Adam
Learning Rate	0.001
Loss Function	Binary Crossentropy
Kernel Size	3 × 3
Activation	ReLU
Dropout	0.2
Weight Initialization	He Normal
FLICM
Number of clusters (C)	3
Fuzziness exponent (m)	2
Convergence threshold (ε)	1e-5
Maximum iterations	100
Neighborhood window size (Nᵢ)	3 × 3
Distance metric	Euclidean
Weighting factor (α)	0.5

Dataset description

Dataset 1

The echocardiography videos are acquired from the HMC-QC dataset [28], which is created by the collaboration between Hamad Medical Corporation (HMC), Tampere University, and Qatar University. The dataset comprises 2D echocardiography recordings in A4C and A2C views. These recordings were acquired using ultrasound devices from different vendors, specifically Philips and GE Vivid (GE-Health-USA) systems. The temporal resolution of the recordings is 25 fps, while the spatial resolution ranges from 422 × 636 to 768 × 1024 pixels.

Dataset 2

The CAMUS dataset [29] consists of clinical echocardiographic examinations from 500 patients 500 patients, acquired at the University Hospital of St Etienne (France). To ensure clinical realism, no prerequisite processing or data selection is applied to the acquisitions, which are optimized for left ventricle ejection fraction measurement. The dataset contains challenging cases characterized by difficult tracing, significant variability in acquisition settings, and instances of incomplete wall visibility in the images. It is organized into two subsets: a training set including 450 patients with manual references based on the evaluation of a clinical expert, and a testing set comprising 50 patients. All raw input images are provided in.mhd file format. Table 2 presents a detailed overview of the datasets used in this study, including HMC-QC and CAMUS, along with their key characteristics and experimental configuration.

Table 2.

Dataset details and experimental configuration

Parameter	HMC-QC Dataset	CAMUS Dataset
Number of Patients	162	500
Number of Videos	322 echocardiography clips	500 echocardiography clips
Class Distribution	MI: 93, non-MI: 69	MI: 250, non-MI: 250
Total Frames	25,760	30,000
MI Frames	14,880	15,000
Non-MI Frames	10,880	15,000
Frame Rate	25 fps	25 fps
Echo Views Used	Apical 2-Chamber (A2C), Apical 4-Chamber (A4C)	Apical 2-Chamber (A2C), Apical 4-Chamber (A4C)
Data Split	80% Train/10% Val/10% Test	80% Train/10% Val/10% Test
Task Level	Frame-level classification	Frame-level classification

Experimental results

Image results of the proposed Efficient_SpinalNet

Figure 5 indicates the experimental result of Efficient_SpinalNet for MI detection using echocardiography. Figure 5a depicts the input frame image, Fig. 5b and c correspond to the outputs obtained using the median filter and Gaussian filter. The LV segmented image and its endocardial boundary are illustrated in Fig. 5d. Figure 5e and f present the LOOP image and the LBP image, respectively.

Fig. 5.

Experimental result of Efficient_SpinalNet for MI detection using echocardiography. (a) input frame image, (b) median filter image, (c) Gaussian filter image, (d) LV segmented image, (e) endocardial boundary image, (f) LOOP image, and (g) LBP image

Segmentation results of different approaches

Figure 6 presents the visual comparison of LV segmentation outputs produced by six state-of-the-art models, namely SAUN, U-Net, DeepLabV3 +, ReU-Net, AS-Unet + +, and FLICM. Each subfigure shows the segmented LV mask and endocardial boundary image. Figure 6 presents the segmentation results of multiple methods. The SAUN scheme is shown in Fig. 6a, followed by the UNet model in Fig. 6b, DeepLab V3 + in Fig. 6c, and ReU-Net in Fig. 6d. The results of AS-UNet + + and the FLICM scheme are illustrated in Fig. 6e and f, respectively.

Fig. 6.

Visual comparison of LV segmentation results, (a) SAUN, (b) UNet, (c) DeepLab V3 +, (d) ReU-Net, (e) AS-Unet + + and (f) FLICM

Evaluation measures

The effectiveness of Efficient_SpinalNet, the various performance measures like accuracy, sensitivity, and specificity, are measured here.

Accuracy

Accuracy refers to the ratio of correctly detected outputs to the total number of outputs produced by Efficient_SpinalNet, and is formulated by,

30

where, øverrightarrow A indicates accuracy, true positive signifies as , true negative represents as , false positive symbolizes as , and false negative is represented by .

Sensitivity

This metric represents the proportion of correctly identified positive instances with respect to the total number of actual positive cases in the dataset. Its formulation is given as follows:

31

where, denotes the sensitivity.

Specificity

Specificity represents the ratio of correctly identified negative instances to the total number of actual negative samples in the dataset. This evaluation metric is computed using the following expression:

32

where, indicates the specificity.

Performance evaluation concerning epochs

The performance of the proposed Efficient_SpinalNet model for myocardial infarction detection across multiple training epochs is illustrated in Fig. 7. Specifically, Fig. 7a illustrates the validation of the devised scheme with respect to the accuracy achieved at different stages of training. When trained on 90% of the data, the recorded accuracy values are 79.98% at epoch 20, 81.65% at epoch 40, 83.21% at epoch 60, 84.97% at epoch 80, and achieved high accuracy values of 90.12% at epoch 100. Figure 7b reveals the validation of Efficient_SpinalNet utilizing sensitivity, and here, when training data is 90%, the sensitivity of Efficient_SpinalNet is 81.55% for epoch 20, 83.06% for epoch 40, 84.95% for epoch 60, 87.66% for epoch 80, and 91.59% for epoch 100. The specificity of Efficient_SpinalNet is shown in Fig. 7c. From this evaluation, the Efficient_SpinalNet attained the specificity of epoch 20 is 81.97%, epoch 40 is 83.54%, epoch 60 is 85.08%, epoch 80 is 88.46%, and epoch 100 is 92.24% for the training data 90%.

Fig. 7.

Performance validation of Efficient_SpinalNet with (a) accuracy, (b) sensitivity, and (c) specificity based on training data

Comparative technique

To evaluate the MI detection effectiveness of Efficient_SpinalNet, Efficient_SpinalNet is compared with existing schemes, like the Active Polynomial model [5], PSq-Net [16], Convolutional Bidirectional-Long Short-Term Memory-Attention (Conv-BiLSTM-Attn) [30], Faster RCNN [12], Social Ski-Spider optimization algorithm-Deep Residual Network (SSS-DRN) [31], MFF–CNN [18], SVM [1], and DNN [14].

Comparative validation

The efficiency of Efficient_SpinalNet is analyzed based on the varying K-fold and training data, which are elaborated below,

Model validation using Dataset 1

Comparative assessment based on varying training data

The evaluation of Efficient_SpinalNet for MI detection using training data across different performance metrics is presented in Fig. 8. Specifically, Fig. 8a shows the accuracy analysis of the proposed approach. With 90% of the training data, the model achieves an accuracy of 90.47%. Moreover, the traditional schemes, like Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN, attained an accuracy of 79.97%, 80.319%, 81.118%, 81.66%, 81.895%, 82.951%, 83.02%, and 86.65%. The high accuracy results achieved by the proposed Efficient_SpinalNet model reduce the likelihood of misdiagnosis and enhances the reliability of automated clinical decision-support systems. Figure 8b illustrates the evaluation of Efficient_SpinalNet in terms of sensitivity, where the sensitivity attained by the number of methods used for MI detection is 81.43% by Active Polynomials, 82.545% by PSq-Net, 82.713% by Conv-BiLSTM-Attn, 83.05% by Faster RCNN, 83.878% by SSS-DRN, 83.878% by MFF–CNN, 84.97% by SVM, 86.87% by DNN, and 92.01% by Efficient_SpinalNet for training data is 90%. The improved sensitivity performance of the proposed scheme demonstrates the model’s strong capability in detecting subtle ventricular wall motion abnormalities and ensures that fewer MI cases remain undetected, which is vital for timely intervention. Figure 8c illustrates the specificity calculated by the various MI detection techniques. The specificity of 81.970%, 82.633%, 82.966%, 83.53%, 84.488%, 84.912%, 85.01%, and 87.97% is obtained for 90% of training data by the existing methods Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN, where the specificity attained by Efficient_SpinalNet is 90.63%. The high specificity obtained by the proposed scheme indicates a reduction in false-positive cases, which minimizes unnecessary follow-up diagnostic procedures and reduces patient anxiety and clinical workload.

Fig. 8.

Comparative validation of Efficient_SpinalNet with Dataset 1 (a) accuracy, (b) sensitivity, (c) specificity in terms of altering training data

Comparative validation based on K-fold

The examination is done to determine the assessment of Efficient_SpinalNet, and the analysis of MI detection using K-fold with various metrics is revealed in Fig. 9. Figure 9a portrays the accuracy noted by the number of schemes utilized for MI detection. With k-value as 9, the accuracy of 81.64%, 82.518%, 82.768%, 83.11%, 83.896%, 84.051%, 84.98%, and 86.86% is measured by Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN, whereas the accuracy obtained by the Efficient_SpinalNet is 91.19%. The high accuracy of the proposed scheme demonstrates that the model maintains reliable classification performance across different K-fold splits, reducing diagnostic uncertainty and ensuring consistent MI detection in diverse patient scenarios. Figure 9b illustrates the sensitivity analysis of various methods, where Efficient_SpinalNet achieves a sensitivity of 91.43% for K-fold = 9. Moreover, the sensitivity attained by the traditional technique, such as Active Polynomials is 79.85%, PSq-Net is 80.646%, Conv-BiLSTM-Attn is 80.976%, Faster RCNN is 81.540%, SSS-DRN is 81.846%, MFF–CNN is 82.565%, SVM is 83.06%, and DNN is 86.43%. The superior sensitivity achieved by the proposed method highlights its effectiveness in correctly identifying true MI cases, which is critical for minimizing missed diagnoses and enabling timely medical intervention. Figure 9c describes the assessment utilizing specificity of MI detection methods, where the specificity attained by various methods used for MI detection is 81.97% by Active Polynomials, 82.676% by PSq-Net, 83.087% by Conv-BiLSTM-Attn, 83.53% by Faster RCNN, 83.707% by SSS-DRN, 83.854% by MFF–CNN, 85.04% by SVM, 88.76% by DNN, and 92.34% by Efficient_SpinalNet for a K-fold value of 9. The high specificity of the devised model confirms the model’s ability to accurately recognize non-MI cases, thereby reducing false alarms, avoiding unnecessary follow-up procedures, and improving resource utilization in clinical settings.

Fig. 9.

Comparative validation of Efficient_SpinalNet with dataset 1, (a) accuracy, (b) sensitivity, (c) specificity in terms of altering K-fold

Model validation using Dataset 2

Concerning training data

The performance evaluation of Efficient_SpinalNet for MI detection on Dataset 2 using 90% training data across multiple metrics is shown in Fig. 10. Figure 10a depicts the accuracy results of the proposed model, where an accuracy of 90.02% is achieved. In comparison, the existing methods such as Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN obtain accuracy values of 80.602%, 80.676%, 80.754%, 82.254%, 82.382%, 82.697%, 83.704%, and 86.162%, respectively. A high accuracy value indicates precise overall classification performance, ensuring reliable identification of both MI and non-MI cases. The sensitivity performance of Efficient_SpinalNet is depicted in Fig. 10b. For 90% training data, the sensitivity values achieved by Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN are 81.807%, 82.687%, 82.889%, 83.356%, 83.664%, 84.747%, 85.794%, and 87.571%, respectively, while Efficient_SpinalNet attains a sensitivity of 91.901%. A high sensitivity value reflects the model’s effectiveness in correctly identifying positive MI cases, which is crucial for ensuring that affected patients are accurately detected. Figure 10c presents the specificity analysis of different MI detection techniques. For 90% training data, the specificity values of Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN are 81.804%, 82.975%, 83.801%, 84.021%, 84.296%, 84.525%, 85.878%, 87.588%, respectively, whereas Efficient_SpinalNet achieves a specificity of 90.704%. A high specificity value indicates the ability of the model to correctly identify non-MI cases, thereby reducing false-positive outcomes and supporting accurate clinical assessment.

Fig. 10.

Comparative validation of Efficient_SpinalNet with Dataset 2, (a) accuracy, (b) sensitivity, (c) specificity using training data

Concerning K-fold Cross validation

The performance of Efficient_SpinalNet for MI detection is evaluated using K-fold cross-validation, and the results across various metrics are presented in Fig. 11. Figure 11a shows the accuracy achieved by different schemes. With the K-fold value set to 9, the accuracy obtained by Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN is 81.279%, 82.205%, 82.624%, 83.355%, 83.554%, 84.374%, 85.684%, and 87.389%, respectively, while Efficient_SpinalNet attains an accuracy of 91.011%. A high accuracy value signifies consistent and reliable overall classification performance across different folds. Figure 11b presents the sensitivity analysis of various MI detection methods. Efficient_SpinalNet achieves a sensitivity of 91.004% for K-fold = 9. In contrast, the sensitivity values of Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN are 80.378%, 80.664%, 80.917%, 81.200%, 81.655%, 82.535%, 83.535%, and 86.955%, respectively. A high sensitivity value indicates the model’s capability to correctly identify MI cases, ensuring that positive cases are effectively detected. Figure 11c depicts the specificity obtained by different MI detection techniques. For K-fold = 9, the specificity values of Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN are 81.925%, 82.716%, 83.346%, 83.389%, 84.089%, 84.453%, 85.544%, and 89.662%, respectively, whereas Efficient_SpinalNet achieves a specificity of 92.435%. A high specificity value reflects the ability of the model to correctly identify non-MI cases, thereby reducing false-positive outcomes and supporting accurate clinical decision-making.

Fig. 11.

Comparative validation of Efficient_SpinalNet with dataset 2, (a) accuracy, (b) sensitivity, (c) specificity using K-fold

Segmentation analysis

Figure 12 showcases the evaluation of the segmentation FLICM algorithm based on training data, highlighting its performance through key metrics such as segmentation accuracy, Dice Score coefficient, Jaccard Coefficient, and Structural Similarity Index (SSIM). Here, the developed FLICM is compared with models, namely stack attention U-Net (SAUN) [16], U-Net [32], DeepLabV3 + [33], ReU-Net [34], and AS-Unet + + [35]. The segmentation accuracy results of the proposed segmentation approach is illustrated in Fig. 12a. When training data is 90%, then the segmentation accuracy gained by FLICM is 90.86%, while the classical segmentation approaches obtained segmentation accuracy of 83.43% for SAUN, 86.64% for U-Net, 87.68% for DeepLabV3 +, 88.55% for ReU-Net, and 89.43% for AS-Unet + +. These results emphasize the superior performance of FLICM in leveraging extensive training data for enhanced precision. Figure 12b demonstrates segmentation performance evaluated through the Dice score coefficient. When the training data is considered as 90%, the FLICM algorithm attained a Dice score value of 0.895, which is minimum among other comparative models. Specifically, SAUN achieved 0.836, U-Net registered 0.858, DeepLabV3 + obtained 0.861, ReU-Net achieved 0.871, and AS-Unet + + achieved 0.883. These figures highlight the enhanced precision of FLICM in delineating structures compared to its counterparts. The Jaccard coefficient graph is provided in Fig. 12c. Here, the Jaccard coefficient values of 0.823 is obtained by SAUN, 0.847 for U-Net, 0.859 for DeepLabV3 +, 0.869 for ReU-Net, 0.876 for AS-Unet + +, and 0.893 for FLICM for 90% training data. Figure 12d shows the SSIM analysis results. With 90% training data, the SAUN, U-Net, DeepLabV3 +, ReU-Net, AS-Unet + +, and FLICM models achieved SSIM values of 0.846, 0.880, 0.883, 0.902, 0.913, and 0.926. The segmentation analysis demonstrated that the devised scheme achieved high-performance results, accurately delineating the left ventricular wall and endocardial boundaries.

Fig. 12.

Segmentation Analysis concerning Training Data, (a) Segmentation accuracy, (b) dice score coefficient, (c) Jaccard Coefficient, and (d) SSIM

Ablation study

Figure 13 displays the ablation study of the proposed work on both datasets. Figure 13a shows the analysis using Dataset 1. Using 90% of the training data, the proposed Efficient_SpinalNet demonstrated a accuracy of 90.907%, whereas, the Efficient_SpinalNet without preprocessing, Efficient_SpinalNet without segmentation of the entire left ventricle wall, Efficient_SpinalNet without segmentation of the endocardial boundary, Efficient_SpinalNet without feature extraction, EfficientNet, and SpinalNet attained the accuracy of 85.876%, 86.591%, 87.584%, 88.251%, 89.354%, and 89.762% concerning dataset 1. It shows the proposed model maintained a significant edge, highlighting its robustness and precision in medical image analysis. The ablation experiment concerning dataset 2 is shows in Fig. 13b For Dataset 2, with the training data of 90%, Efficient_SpinalNet without Preprocessing, Efficient_SpinalNet without Segmentation of entire left ventricle Wall, Efficient_SpinalNet without Segmentation of endocardial boundary, Efficient_SpinalNet without Feature Extraction, EfficientNet, and SpinalNet attained the accuracy of 84.731%, 85.785%, 86.641%, 87.357%, 88.506%, and 89.435%, whereas the proposed Efficient_SpinalNet accomplished a high accuracy of 90.137%. The results clearly demonstrate the contribution of each component in the proposed Efficient_SpinalNet model toward improved MI detection performance.

Fig. 13.

Ablation study of Efficient_SpinalNet concerning training data, (a) Dataset 1 and (b) Dataset 2

Statistical test

To determine whether the proposed approach offers statistically significant improvements over existing models, paired t-tests are employed. The evaluation is based on three key performance indicators accuracy, sensitivity, and specificity derived from experiments using k-value 9. For each metric, the method’s fold-wise results are systematically compared against those of the baseline techniques. A p-value threshold of 0.05 is used to establish statistical relevance. The outcomes of these comparative tests are detailed in Table 3.

Table 3.

Paired T-test

Proposed Method	Comparative Method	T-statistic			P-Value
		ACC	SEN	SPE	ACC	SEN	SPE
Proposed Efficient_SpinalNet	Active Polynomials	2.626	2.575	2.769	0.029	0.032	0.038
	PSq-Net	2.336	2.158	2.600	0.029	0.028	0.031
	Conv-BiLSTM-Attn	2.218	2.087	2.575	0.028	0.028	0.029
	Faster RCNN	2.040	1.921	2.310	0.028	0.026	0.026
	SSS-DRN	1.876	1.735	2.199	0.028	0.026	0.022
	MFF–CNN	1.707	1.698	2.014	0.027	0.024	0.015
	SVM	1.705	1.598	1.679	0.011	0.014	0.014
	DNN	1.246	1.340	1.265	0.009	0.010	0.011

As presented in Table 3, the evaluation reveals that the proposed approach demonstrates superior performance over all benchmark models namely, Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN across the key metrics of accuracy, sensitivity, and specificity. Each comparative analysis yields statistically significant improvements, confirmed by p-values below the 0.05 threshold. The proposed approach yields notably higher accuracy scores when compared to several benchmark models, including Active Polynomials (t = 2.626, p = 0.029), PSq-Net (t = 2.336, p = 0.029), Conv-BiLSTM-Attn (t = 2.218, p = 0.028), Faster RCNN (t = 2.040, p = 0.028), SSS-DRN (t = 1.876, p = 0.028), MFF–CNN (t = 1.707, p = 0.027), SVM (t = 1.705, p = 0.011), and DNN (t = 1.246, p = 0.009). Comparable statistically significant gains are also evident in sensitivity and specificity metrics. These consistent improvements underscore the reliability and enhance predictive capability of the proposed model over conventional classification techniques.

Computational time

Computational or execution time indicates how long an algorithm or model takes to complete a designated task from initiation to completion. Table 4 provides a comparative overview of the time efficiency of various techniques applied for MI detection. The computational time analysis of Efficient_SpinalNet and the existing MI detection methods for both datasets is presented in terms of execution efficiency. For Dataset 1, the computational time recorded by Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN is 0.7582 s, 0.6984 s, 0.6429 s, 0.6111 s, 0.5766 s, 0.5582 s, 0.5120 s, and 0.4784 s, respectively, whereas Efficient_SpinalNet requires 0.4274 s. For Dataset 2, the computational time observed for Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN is 0.7934 s, 0.7312 s, 0.6605 s, 0.6244 s, 0.5871 s, 0.5787 s, 0.5316 s, and 0.5044 s, respectively, while Efficient_SpinalNet achieves a computational time of 0.4654 s. The analysis reveals that the proposed Efficient_SpinalNet achieves the minimum computational time compared to all existing methods across both datasets, requiring 0.4274 s for Dataset 1 and 0.4654 s for Dataset 2. This reduced execution time indicates faster processing and improved computational efficiency, making the model more suitable for real-time MI detection and practical clinical deployment.

Table 4.

Computational time

Dataset/Methods	Active Polynomials	PSq-Net	Conv-BiLSTM-Attn	Faster RCNN	SSS-DRN	MFF–CNN	SVM	DNN	Proposed Efficient_SpinalNet
Computational Time (sec)
Dataset 1	0.7582	0.6984	0.6429	0.6111	0.5766	0.5582	0.5120	0.4784	0.4274
Dataset 2	0.7934	0.7312	0.6605	0.6244	0.5871	0.5787	0.5316	0.5044	0.4654

Discussion

The effectiveness of Efficient_SpinalNet for MI detection is determined by considering its performance with existing methods. The results noted from an experiment by schemes utilizing altered training data and K-fold data are illustrated in Table 5. The Efficient_SpinalNet achieved a superior process with a greater accuracy of 91.19%, sensitivity of 92.01%, and specificity of 92.435%. In addition, the accuracy achieved by Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN is 81.64%, 82.518%, 82.768%, 83.11%, 83.896%, 84.051%, 84.98%, and 86.86%, which also attained the sensitivity of 81.43%, 82.545%, 82.713%, 83.05%, 83.876%, 83.878%, 84.97%, and 86.87%. Also, the techniques received a specificity of 81.925% by Active Polynomials, 82.716% by PSq-Net, 83.346% by Conv-BiLSTM-Attn, 83.389% by Faster RCNN, 84.089% by SSS-DRN, 84.453% by MFF–CNN, 85.544% by SVM, and 89.662% by DNN. Efficient_SpinalNet performs better than other conventional methods, such as Active Polynomials, PSq-Net, Conv-BiLSTM-Attn, Faster RCNN, SSS-DRN, MFF–CNN, SVM, and DNN, during MI detection by attaining superior accuracy, sensitivity, and specificity. Specifically, while traditional models primarily rely on raw image inputs or limited feature representations, the proposed method enhances feature quality by incorporating both spatial and temporal information, namely, segmented area curves and displacement curves extracted from precisely segmented LV regions using FLICM. Additionally, the fusion of EfficientNet-B3-attn-2, known for its strong feature extraction capabilities, with SpinalNet, recognized for its efficient classification accuracy, leads to a more robust and accurate detection model. This integrated design enables Efficient_SpinalNet to achieve the highest accuracy and the lowest computational time. From a clinical perspective, the proposed Efficient_SpinalNet model can serve as a CAD tool to assist cardiologists in improving the speed, consistency, and reliability of myocardial infarction detection from echocardiography videos. By automatically analyzing left ventricular motion and structural changes, the system has the potential to reduce observer variability and support early clinical decision-making in routine cardiac assessment.

Table 5.

Comparative discussion

Variations	Metrics	Active Polynomials	PSq-Net	Conv-BiLSTM-Attn	Faster RCNN	SSS-DRN	MFF–CNN	SVM	DNN	Proposed Efficient_SpinalNet
Dataset 1
Training data (90%)	Accuracy (%)	79.970	80.319	81.118	81.660	81.895	82.951	83.020	86.650	90.470
	Sensitivity (%)	81.430	82.545	82.713	83.050	83.876	83.878	84.970	86.870	92.010
	Specificity (%)	81.970	82.633	82.966	83.530	84.488	84.912	85.010	87.970	90.630
K-fold (9)	Accuracy (%)	81.640	82.518	82.768	83.110	83.896	84.051	84.980	86.860	91.190
	Sensitivity (%)	79.850	80.646	80.976	81.540	81.846	82.565	83.060	86.430	91.430
	Specificity (%)	81.970	82.676	83.087	83.530	83.707	83.854	85.040	88.760	92.340
Dataset 2
Training data (90%)	Accuracy (%)	80.602	80.676	80.754	82.254	82.382	82.697	83.704	86.162	90.020
	Sensitivity (%)	81.807	82.687	82.889	83.356	83.664	84.747	85.794	87.571	91.901
	Specificity (%)	81.804	82.975	83.801	84.021	84.296	84.525	85.878	87.588	90.704
K-fold (9)	Accuracy (%)	81.279	82.205	82.624	83.355	83.554	84.374	85.684	87.389	91.011
	Sensitivity (%)	80.378	80.664	80.917	81.200	81.655	82.535	83.535	86.955	91.004
	Specificity (%)	81.925	82.716	83.346	83.389	84.089	84.453	85.544	89.662	92.435

Conclusion

Early detection of MI, often driven by lifestyle-related risk factors, is crucial as delayed diagnosis can lead to chronic heart failure or sudden cardiac death. Here, Efficient_SpinalNet is introduced for MI detection using echocardiography. The process begins by selecting a video sample from the dataset. This input undergoes enhancement through a combination of median and Gaussian filtering techniques to reduce noise and improve image clarity. The refined frames are then passed into the segmentation module, where the complete LV wall is delineated using the FLICM algorithm. Subsequently, the same algorithm is applied to isolate the endocardial boundary within the segmented LV region. From this, displacement profiles and area-based segmentation curves are derived. Upon completion of boundary extraction, the workflow advances to the feature extraction stage. Then, MI detection is done by Efficient_SpinalNet, which is designed by a combination of EfficientNet-B3-attn-2 and SpinalNet. The results show that superior performance is achieved by Efficient_SpinalNet with a maximum of 91.19% accuracy, 92.01% sensitivity, and 92.435% specificity. The evaluation is conducted on publicly available datasets, which may not fully represent the complete diversity of real-world clinical scenarios. In addition, variations in echocardiography data quality, patient-specific cardiac motion patterns, and acquisition conditions can still influence model performance. While cross-dataset testing improves robustness, further validation on large-scale multi-center clinical data is required to confirm consistent performance across different hospital environments, imaging devices, and operator settings. Moreover, the current framework is computationally efficient but may still require optimization for deployment in low-resource or real-time embedded system. In future work, the model can be extended toward real-time implementation and integration into ultrasound imaging systems or portable point-of-care devices. However, further validation on large-scale multi-center clinical datasets is required before routine clinical deployment to ensure robustness and generalizability across different clinical settings, imaging conditions, and patient variability.

Authors’ contributions

All authors have made substantial contributions to conception and design, revising the manuscript, and the final approval of the version to be published. Also, all authors agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Funding

This research did not receive any specific funding.

Data availability

The data that support the findings of this study are openly available in the HMC-QU dataset at https://www.kaggle.com/datasets/aysendegerli/hmcqu-dataset, reference number [28] and the CAMUS dataset at https://www.creatis.insa-lyon.fr/Challenge/camus/databases.html, reference number [29].

Declarations

Ethics approval and consent to participate

Not applicable.

Competing interests

The authors declare no competing interests.