Enhanced Atrial Fibrosis Detection Using 2D Echocardiogram Images with an Advanced Deep Learning Framework and Weighted Multi-Feature Fusion

Research on AF detection using artificial intelligence has evolved from traditional machine learning toward deep learning–based frameworks. Existing studies can broadly be categorized into three groups: CNN-based approaches, RNN-based approaches, and hybrid models [13].

(1). CNN-based models

CNNs have been widely applied to extract spatial or spectral features from ECG and echocardiographic signals.

Pourbabaee et al. [1] proposed a deep convolutional neural network (DCNN) for AF detection from ECG time-series images, showing improvements over traditional classifiers. Cao et al. [2] developed Multi-Scale Decomposition Enhanced Residual CNNs (MSResNet) and FDResNets [14], combined via transfer learning, to improve F1-score and accuracy.

Nurmaini et al. [3] integrated Discrete Wavelet Transform (DWT) with 1D-CNNs for multiclass AF classification, demonstrating better generalization and early diagnostic capability. Chandra et al. [15] employed wavelet and Fourier transforms to generate 2D spectrograms fed into a DCNN, reporting enhanced AF detection reliability. Schwab et al. [7] applied a CNN-based framework validated on external datasets, demonstrating improved prediction of high-risk AF cases compared to classical models. Although CNNs can extract powerful features, they often require large balanced datasets, are sensitive to noise, and may lack interpretability in clinical contexts.

(2). RNN-based models

RNNs and their variants (e.g., LSTMs) are particularly suited for sequential biomedical signals.

Faust et al. [4] applied LSTMs to heart rate signals for AF detection, showing robustness to noisy and incomplete data. Andersen et al. [6] combined CNN and RNN layers for end-to-end AF classification from RR intervals, achieving strong performance even on unseen datasets. RNNs are effective for temporal dependencies but face challenges such as vanishing gradients, high computational cost, and limited scalability to large echocardiographic datasets [16].

(3). Hybrid models

Hybrid approaches combine CNNs with RNNs or other optimization strategies to address data imbalance and enhance generalization.

Petmezas et al. [5] proposed a CNN-LSTM model trained with focal loss to mitigate data imbalance, achieving higher sensitivity and specificity. While hybrids improve classification, they often increase architectural complexity [17], require fine-tuned hyperparameters, and lack efficient segmentation pipelines for echocardiographic imaging [18].

2.1 Problem statement

Atrial fibrosis detection methods have several limitations, including subjective interpretation, inability to detect early stages of fibrosis, variability in image quality [19], high computational cost, and inability to handle large datasets, etc. It may also face overfitting and underfitting issues. The features and challenges of existing atrial fibrosis detection techniques are given in Table 1 and the traditional model's experimental gap is given.

• In the existing techniques, high computational requirements and long processing times [20] hinder real-time diagnosis and clinical relevance. These issues can be overcome by using an optimization algorithm that minimizes computational requirements making the model more accurate, efficient, and clinically applicable.
• Existing atrial fibrosis detection methods [21] may suffer from limited data quality, availability, and high dimensionality. These problems can be overcome by using feature extraction methods, which extract relevant information from the data and reduce the data dimensionality.
• Existing atrial fibrosis diagnostic models may lack transparency, making it challenging to understand detection decisions. This problem can be overcome by combining two or more other algorithms to solve the identical issue, which improves computational efficiency and accuracy [22].

Table 1. Designed HC-TMU-based segmented images

Original Image	Ground Truth Image	UNet [23]	UNet 3+ [24]	Proposed HC-TMU
1.png	1.png	1.png	1.png	1.png
2.jpg	2.png	2.png	2.png	2.png
3.jpg	3.png	3.png	3.png	3.png
4.jpg	4.png	4.png	4.png	4.png
5.jpg	5.png	5.png	5.png	5.png

4.1 Echocardiogram 2D image feature extraction

Feature extraction is the major phase in this designed framework, the main intent behind this feature extraction is to extract the significant features from the accumulated 2D ECG images $C l_y^{E C G}$ that promote better detection outcomes. During feature extraction, three sets of features are extracted from the data. A detailed description of each Feature Set is given below.

Feature Set 1: The first set of features is attained by concatenating three diverse features including fuzzy entropy vector, wavelet packet energy vectors, and hierarchical theory vectors are extracted from the given input data $C l_y^{E C G}$, and the obtained vectors are combined to form the Feature Set 1.

Fuzzy entropy vector [29]: Fuzzy entropy (FE) is employed to evaluate the time series data's irregularity and complexity and the execution procedure of FE is discussed below.

Let us consider, a time series data $Z=\{z(i): 1 \leq i \leq B\}$, in which time series data's length is denoted as B. Later, the mean value $z_0(i)$ is evaluated via Eq. (1):

$z_0(i)=\frac{1}{n} \sum_{k=0}^{n-1} z(i+k)$                (1)

Here, the embedding dimension is denoted as n, and the n-dimension vector $Z_i^n(i=1,2, \ldots, B-n)$ is modified as in Eq. (2):

$Z_i^n=\{z(i), z(i+1, \ldots, z(i+n-1))\}-z_0(i)$               (2)

The fuzzy function $\mu\left(s_{i k}^n, b, e\right)$ is described in Eq. (3).

$\mu\left(s_{i k}^n, b, e\right)=\exp \left(-\frac{\left(s_{i k}^n\right)^b}{e}\right)$               (3)

Here, the function of exponential is denoted as $exp\ (\cdot)$, and the boundary gradient and width are signified as $b$ and $e$. Later, the degree of similarity $S_{i k}^n$ among $Z_i^n$ and $Z_j^n$ is described in Eq. (4):

$S_{i k}^n(b, e)=\mu\left(s_{i k}^n, b, e\right)$               (4)

Finally, the FE time series $\{z(i): 1 \leq i \leq B\}$ is evaluated via Eq. (5):

$F E(Z, n, b, e)=\lim _{B \rightarrow \infty}\left(\ln \Phi^n(b, e)-\ln \Phi^{n+1}(b, e)\right)$               (5)

Here, the natural logarithm operation is represented as $ln (\cdot)$. If $B$ is finite, then the FE time series $F E(Z, n, b, e)$ is represented as in Eq. (6):

$F E(Z, n, b, e, B)=\ln \Phi^n(b, e)-\ln \Phi^{n+1}(b, e)$               (6)

The extracted fuzzy entropy vector is expressed as $F E_x^{V t}$.

Wavelet packet energy vectors [30]: It is defined as the energy in diverse frequency bands that are evaluated from the wavelet packet decomposition outcome. The wavelet packet decomposition energy is estimated by Eq. (7) and the overall energy of the signal is derived by Eq. (8):

$W_i=\int_{-\infty}^{+\infty} z_k^i(t) d t$                (7)

$W_{\text {Tol }}=\sum_{i=1}^{2^k} W_i$              (8)

Here, the energy within each sub-band is indicated as $W_i$. Further, the value of normalized energy that is equivalent to each wavelet packet’s energy is defined by Eq. (9):

$Q_l=\frac{W_i}{W_{T o l}}$               (9)

Here, each sub-band’s probability distribution is expressed as $Q_l$. The extracted wavelet packet energy vectors are represented as $W p_v^{s t}$.

Hierarchical theory vectors [31]: Hierarchical theory is described as the procedure of arranging or expressing the features within the given input in the hierarchical format. During hierarchical feature extraction, it extracts multiple features from the input to enhance the efficiency of the model in employing global context data. During feature extraction, each feature extracted from the image is represented in the vector format that helps to monitor the connection among the diverse features. Moreover, employing hierarchical theory in feature extraction helps in determining complicated patterns or connections within the features and accelerates the performance of the model during segmentation. The extracted hierarchical theory vectors are indicated as $H f_w^{s c}$.

Feature concatenation: In order to attain the final Feature Set 1. The obtained fuzzy entropy vector $F E_x^{V t}$, wavelet packet energy vectors $W p_v^{s t}$, and hierarchical theory vectors $H f_w^{s c}$ are concatenated together and are denoted as $F s 1_a^{\text {Con }}$ and it is expressed in Eq. (10):

$F E_x^{V t}+W p_v^{s t}+H f_w^{s c}=F s 1_a^{c o n}$                (10)

Here, the first Feature Set is indicated as $F s 1_a^{\text {con }}$.

Feature Set 2: The second Feature Set is obtained by employing R-ViT as a tool to extract the features from the input data $C l_y^{E C G}$. The R-ViT [32] is a variant of classical ViT that comprises a regional-to-local attention module that helps to understand both global and local characteristics from the input data. The R-ViT model comprises two tokenization procedures that transfer input data into local and regional tokens, which assist in extracting significant features from the image. Moreover, this tokenization procedure is considered a convolution function with diverse path sizes. In R-ViT, the path size of the local and regional tokens is 4² and 28². Moreover, by employing the regional-to-local attention module the R-ViT processes both the tokens. Additionally, down-sampling procedures were carried out to separate the token’s spatial resolution. At last, the features are obtained by taking an average value among the regional tokens. Similarly, the fine-grained location data are also obtained from the local token. The features extracted by employing R-ViT are represented as $F s 2_b^{R-V i T}$.

Feature Set 3: The third set of features is extracted by employing the CNN [33]. CNN is employed to extract the deep features from the given input information. The key layers of the CNN model that are responsible for feature extraction are the convolution and max pooling layers. Each layer embedded in the CNN model has its own tasks to perform.

Convolutional layer: This layer comprises filters and feature maps. The filers within the layer are viewed as the neuron and the outcome from the filter utilized by the prior level is referred to as the feature maps.

Max pooling layer: The key function of this layer is to carry out the down-sampling on the feature maps with an aim to minimize the issues associated with overfitting. The deep features extracted from the input images by utilizing CNN are denoted as $F s 3_c^{C N N}$.

The pictorial view to describe the feature extraction phase is represented in Figure 3.

3.png

Figure 3. Echocardiogram 2D image feature extraction

4.2 Proposed HDF-RBMO-based weighted feature fusion process

Followed by feature extraction, weighted feature fusion is the next phase that assists in enhancing detection performance by optimizing weight along with feature fusion. At this phase, the significant features from each extracted features set $F s 1_a^{C o n}, F s 2_b^{R-V i T}$ and $F s 3_c^{C N N}$ are chosen optimally by utilizing the designed HDF-RBMO. During feature selection, the feature $O F 1_t^{\text {Con }}$ from Feature Set 1 is selected from the range $[1,10]$, the feature $O F 2_u^{\mathrm{R}-\mathrm{ViT}}$ from Feature Set 2 is chosen from the limit $[11,20]$ and the feature $O F 3_r^{\mathrm{CNN}}$ from Feature Set 3 is taken from the limit [21,30]. Additionally, the optimized weights $W t_t^{o p}, W t_u^{o p}$ and $W t_r^{o p}$ that lie within the range [0.01,0.99] are multiple with the corresponding feature, which results in weighted Feature Sets as expressed in the following equations:

$W f_t^{f 1}=O F 1_t^{C o n} * W t_t^{o p}$              (11)

$W f_u^{f 2}=O F 2_u^{R-V T} * W t_u^{o p}$                (12)

$W f_r^{f 3}=O F 3_r^{C N N} * W t_r^{o p}$                  (13)

Later, the resultant weighted Feature Sets are fused together as shown in Eq. (14) and the final fused weighted feature $F u_g^{F e a}$ is taken for further processing:

$F u_g^{F e a}=W t_t^{f 1}+W t_u^{f 2}+W t_r^{f 3}$              (14)

The key objective behind weighted feature fusion is to increase the relief score as well as the correlation coefficient and it is numerically derived in Eq. (15):

$O b j_1=\underset{\left\{\substack{O F 1_t^{C o n}, O F 2_u^{R-V i T}, O F 3_r^{C N N}, W t_t^{o p}, W t_u^{o p}, W t_r^{o p}}\right\}}{\operatorname{argmin}}\left(\frac{1}{R F_{s c}+\operatorname{Cor}_{c o f}}\right)$              (15)

Relief score $R F_{s c}$ is described as the method to evaluate the features with multiple classes from the given input ECG images and they are capable of evaluating how the features within the images distinguish its instances among multiple identical classes. Moreover, it determines the weight T of the feature D, i.e., T[D] by employing Eq. (16):

$T[D]=Q\left(D_{S B} \mid S^{n i}\right)-Q\left(D_{S B} \mid F^{n i}\right)$              (16)

From the above equation, the livelihood of the diverse values of the feature D on different instances is represented as Q, different values of the feature D are indicated as $D_{S B}$, the closest instance from a different class is denoted as $S^{n i}$, and the closest instance from a similar class is represented as $F^{n i}$.

Correlation coefficient $\operatorname{Cor}_{\text {cof }}$ is defined as the statistical metrics that evaluate the robustness of the linear connection among two variables as well as it is the normalized evaluation of the covariance among the variables and their values lying between the limit [-1,1] and it is numerically defined in Eq. (17):

$C o r_{c o f}=\frac{n \sum S G-\sum S \sum G}{\left(n \sum S^2-\left(\sum S\right)^2\right) \cdot\left(n \sum G^2-\left(\sum G\right)^2\right)}$            (17)

The total count of data points within the given data is n and the summation of the product of $S^{\text {th }}$ and $G^{\text {th }}$ value for every data point is represented as $\sum S G$. The diagrammatic representation of HDF-RBMO-based weighted feature fusion is provided in Figure 4.

4.png

Figure 4. Developed HDF-RBMO-based weighted feature fusion process

4.3 Proposed HDF-RBMO

In this designed framework, a hybrid optimization approach is developed by integrating the functions of classical DFA and RBMO and named HDF-RBMO. The prime function of the developed HDF-RBMO is to select the optimal features within the extracted Feature Sets and to optimize the weight for multiplying with optimal features to enhance the relief score and correlation coefficient, which helps to improve the detection performance. In this hybrid strategy, DFA [27] is employed by considering its ability to manage huge-scale optimization issues, which makes it more applicable in complicated situations. Moreover, they offer better performance in various tuning issues like multi-model sand non-linear issues, which results in robust performance. However, the DFA is highly complicated to execute and learn. The efficiency of DFA is vulnerable to parameter optimization, selecting the most significant parameter results in suboptimal outcomes. Additionally, in certain cases, the DFA faces difficulties in converging effectively for the optimal outcome, particularly in dynamic surroundings that also result in sub-optimal outcomes. Therefore, the RBMO [24] strategy is integrated with the DFA to address these shortcomings since they present effective performance in complicated regions and avoid local minimum as well as offer better convergence than other classical tuning approaches. In this developed HDF-RBMO, a random value r within the range [0,1] is upgraded according to the developed concept as shown in Eq. (18).

$r=\frac{C_{\text {fit }}}{\left(W_{\text {fit }}+M_{\text {fit }}+B_{\text {fit }}\right)}$               (18)

In the above equation, the best, worst, mean, and current fitness values are indicated as $B_{f i t}, W_{f i t}, M_{f i t}$ and $C_{f i t}$. The pseudocode of the proposed HDF-RBMO is presented in Algorithm 1.

6.1 Simulation setup

The introduced AF detection framework was built and trained on the Python platform. The setup for this detection framework was established by considering the population counts as 10, the highest count of iteration as 50, and the chromosome length as 33. Additionally, the performance analysis was carried out by analyzing the outcome of the designed mode with a set of classical optimization strategies including Reptile Search Algorithm (RSA), Mud Ring Algorithm (MRA), Dark Forest Algorithm (DFA) and RBMO and the segmentation performance is analyzed with DCNN, MSResNet, LSTM, and RNN.

6.2 Evaluation measures

The effectiveness of the designed AF detection framework over diverse optimization strategies and segmentation approaches is demonstrated based on the following measures.

(a) Accuracy of the approach is estimated by Eq. (27).

$A y=\frac{Q^T+M^T}{Q^T+M^T+Q^F+M^F}$              (27)

(b) The dice coefficient of the model is evaluated using Eq. (28).

$D f=2 \times \frac{\left|M t_c^{i g} \cap D i_t^{i g}\right|}{\left|M t_c^{i g}\right|+\left|D i_t^{i g}\right|}$               (28)

(c) Jaccard is determined by employing Eq. (29).

$J_{c f}=\frac{\left|M t_c^{i g} \cap D i_t^{i g}\right|}{\left|M t_c^{i g}\right|+\left|D i_t^{i g}\right|-\left|M t_c^{i g} \cap D i_t^{i g}\right|}$                 (29)

(d) Sensitivity of the model is derived using Eq. (30).

Sensitivity $=\frac{Q^T}{Q^T+M^F}$             (30)

(e) Specificity is evaluated by Eq. (31).

Specificity $=\frac{M^T}{M^T+Q^F}$              (31)

Here, the false positive, false negative, true positive, and true negative are represented as $Q^F, M^F, Q^T$, and $M^T$. The ground truth and segmented image are described as $M t_c^{i g}$ and $D i_t^{i g}$.

6.3 Designed HC-TMU-based segmented images

The segmented images obtained by the developed HC-TMU model are presented in Table 2.

Table 2. Overall performance validation with classical techniques

6.4 Detection performance estimation of the designed model

The introduced HC-TMU's segmentation performance is monitored by considering epoch count and the obtained graphs are presented in Figure 7. This evaluation estimates how effectively this designed segmentation model determines the infected area with the given input data. This evaluation helps to determine the quality of resources employed by the segmentation models during the implementation phase. It is necessary to analyze that the introduced framework has the capacity to process the input data more effectively even in the presence of unwanted noise or artifacts. Moreover, this evaluation is performed with the standard framework to monitor the performance in terms of accuracy, Jaccard, and dice coefficient by varying the epoch count from 50–250. From Figure 7(a), the accuracy of the introduced HC-TMU framework is 14.45%, 11.76%, 9.19%, and 5.55% more advanced than DCNN, MSResNet, LSTM, and RNN, when considering the epoch count as 50. Hence, the segmentation efficiency of the introduced model is superior to other standard frameworks.

(a)

(b)

(c)

(d)

(e)

(f)

(g)

Figure 7. Detection performance evaluation of the proposed framework with traditional approaches with respect to (a) Accuracy, (b) Dice-coefficient, (c) Jaccaed, (d) MSE, (e) PSNR, (f) Sensitivity, (g) Specificity

6.5 Convergence analysis for the developed model

The convergence graph attained by the developed HDF-RBMO model is shown in Figure 8. This evaluation is essential for evaluating the convergence of the employed tuning strategy throughout the iteration. From the given convergence graph, the straight line describes that the best fitness value identified by the tuning strategy is stable and reveals that the optimization technique is stuck within the local minimum. According to the obtained graph, the convergence of the proposed HDF-RBMO model has the ability to offer more advanced performance, when compared to the traditional techniques.

8.png

Figure 8. Convergence analysis for the proposed framework

6.6 Overall detection performance validation with classical models

The overall segmentation performance of the designed model is presented in Table 3. The segmentation performance is determined by considering the mean, best, median, and worst values. By varying the statistical measures, a constant threshold value is handled during the segmentation procedure. Moreover, the developed model’s flexibility is improved which is necessary for determining the most significant features from the given input. From the Table, the dice coefficient of the suggested framework, when analyzing the mean performance is 14.73%, 13.42%, 11.91%, and 9.21% more effective than other classical techniques such as DCNN, MSResNet, LSTM, and RNN, and provided improved performance throughout AF segmentation.

6.7 Statistical evaluation for the suggested framework

Statistical analysis is essential to identify the effectiveness of the developed HDF-RBMO technique based on weighted feature fusion. The statistical analysis of the designed technique with standard approaches is given in Table 3. From Table 3, the best performance of suggested HDF-RBMO is 20.58%, 23.40%, 22.09%, and 26.11% better than the traditional techniques such as RSA, MRA, DFA, and RBMO. Thus, the analysis outcome revealed that the effectiveness of the designed optimization model is more effective than other standard techniques.

Table 3. Statistical evaluation for the developed framework