DCFFNet: A New Dual-channel Cross-Feature Fusion Net for Evaluating the Degree of Aortic Valve Calcification Based on Echocardiographic Images

Abstract

Aortic valve calcification is a common cause of stenosis. Echocardiography, although being the preferred and most prevailing diagnostic technique for aortic valve diseases, lacks effective methods for accurately rating the degree of aortic valve calcification. In this study, a dual-channel cross-feature fusion neural network was developed to classify the degree of aortic valve calcification. The dual-channel input is designed to accept both end-systolic and end-diastolic ultrasound images of a patient, thereby maximizing the retention of vital information from both phases of the cardiac cycle. Feature extraction scale was also dynamically adjusted using the squeeze-and-excitation module. To better integrate multilevel and multiscale information, a dual-branch feature fusion module with cross-feature fusion and multiscale feature extraction was designed, thereby enabling the network to merge global and local feature information. Moreover, to address the specific noise characteristics of ultrasound images and low valve occupancy in the aortic short-axis view, a unified preprocessing algorithm was developed. A total of 420 volunteers were internally selected and classified based on computed tomography scan calcification scores (140 cases per category: healthy, nonsevere, and severe).Each patient contributed 2–4 cardiac cycles, resulting in a final effective dataset of 1092 samples. The classification model achieved an accuracy, precision, F1 score, and recall of 96.79%, 98.59%, 97.97%, and 97.22%, respectively. The artificial intelligence-assisted diagnosis system proposed in this study exhibits high precision in evaluating the degree of aortic valve calcification, positioning echocardiographic examination as a promising alternative in routine aortic valve calcification analysis and screening.

Introduction

Aortic stenosis (AS) has become a significant cause of cardiovascular morbidity and mortality, affecting more than 12.6 million adults worldwide with an estimated 102,700 deaths annually [1, 2]. Due to aging or congenital factors, the aortic valve leaflets experience degenerative changes and form aortic valve calcification (AVC), the predominant source of AS [3–5]. Early detection of the extent of valve calcification and implementing the necessary preventive or therapeutic measures may delay the progression of aortic valve disease to some extent [6, 7]. Therefore, accurate grading of the extent and degree of valve calcification helps in evaluating the course and prognosis of patients and supports physicians in developing an evidence-based treatment plan. The two most common diagnostic modalities for AVC are echocardiography (Echo) and computed tomography (CT) [8, 9]. On the one hand, CT examination is highly accurate and reproducible and considered the gold standard in scoring the degree of AVC. However, due to its associated radiation exposure, it is not routinely performed [10, 11]. On the other hand, although ultrasound remains the routine examination method of AS due to its simplicity, accessibility, and radiation safety [12, 13], there are no established tools or software systems in echocardiographic techniques to quantitatively evaluate the extent of leaflet calcification, which is highly expertise-dependent and cumbersome in practice [14], reporting a 25% rate of interphysical confliction, which significantly impacts clinical decision-making [15, 16]. Current ACC/AHA and ESC/EACTS guidelines recommend semi-quantitative echocardiographic grading of leaflet calcification (assigning each leaflet a score from 0 to 3 based on echogenicity and mobility), which has demonstrated moderate correlation with hemodynamic severity and clinical outcomes [17, 18], while reserving CT-derived Agatston scoring—defining severe AVC as ≥ 1200 AU in women and ≥ 2000 AU in men—for cases with inconclusive echo findings or low-flow low-gradient AS [19]; echo-derived methods to calculate AVC density (calcium area divided by annulus surface area) have shown good agreement with CT calcium scores in research settings [20] but lack standardized clinical software and suffer from interobserver variability [21]. Philippe Unger [22] et al. found no significant difference between patients with high AVC scores and those with concordant severe AS (CSev-AS) in clinical trials. In fact, AVC is a strong independent predictor of mortality in patients with AS [23], and predicts faster AS progression [24]. Therefore, there exists an urgent need for a method for the accurate characterization of AVC based on ultrasound images. Previous studies have used neural networks to train on patient echocardiograms to assist doctors in diagnosing whether a patient has rheumatic heart disease (RHD) [25, 26]. Yazan et al. used patients’ ECG signals to diagnose their valve disease based on artificial intelligence (AI) [27]. Furthermore, researchers have used neural networks to analyze 2D and Doppler echocardiograms of AS with different degrees of lesions [28, 29]; for example, Ma et al. developed an AI-enhanced framework for comprehensive assessment of the AS continuum directly from cine-loop echoes, achieving AUC > 0.90 for differentiating mild, moderate, and severe AS [30]. However, these studies did not focus on analyzing the degree of AVC [31]. To address severe AS detection specifically, Smith et al. proposed an AI algorithm that automatically identifies severe AS from standard B-mode echo videos with sensitivity = 95% and specificity = 92% in a 1200-patient cohort [32]. Some researchers have also applied deep learning algorithms to analyze nonenhanced electrocardiogram-gated cardiac CT data to determine the severity of AVC [33]; however, CT is not routinely used in clinical practice. In the field of valvular heart disease, the focus of AI has been on the echocardiographic evaluation and phenotyping of patient populations to identify high-risk groups [34]. In another study, deep learning networks were used to classify AVC in a single ultrasound image of a patient, but it did not integrate ultrasound images from multiple states of the patient for a comprehensive analysis [35]. More recently, Elvas et al. demonstrated a two-stage CNN pipeline that automatically segments the aortic valve region and detects calcific deposits in B-mode echoes with precision = 92%, paving the way for automatic calcium quantification directly from echo images [36].

This study propose a new AI-based assisted diagnostic algorithm for evaluating the degree of AVC. To address echocardiography-specific challenges, such as low contrast, presence of speckle noise, and a small region of interest (ROI) area, the study further optimized the process by prefiltering image noise with a homomorphic filter and segmenting the ROI with a YOLOv8 detector before inputting into the neural network. The deep layers of the convolutional neural network can extract detailed low-level information from the original image, which serves potential clues for detecting echocardiographic problems [37, 38]. To better integrate useful information from patients’ ultrasound images, this study designs a new dual-channel cross-feature fusion neural network, aimed at more accurately quantifying valve calcification. In this method, the study designed two input channels for the neural network to combine image features from both the end-systolic and end-diastolic phases of the patient. The squeeze-and-excitation module (SEM) includes a multihead attention mechanism to adaptively adjust the reduction and dropout ratio parameters according to the training outcomes of the network. A dual-feature fusion module is integrated into the model architecture, which performs multichannel image feature fusion by feature vector splicing and introducing feature scale maps. The fusion result is used as the input for the next stage of feature extraction. Finally, a combined feature module (CFM) is added before classification to merge the outputs of the dual channels. Using multiscale feature fusion, the receptive field of the neural network expands, allowing for the extraction of rich image features. Additionally, an Optimal Transport-based Feature Alignment Fusion (OT-FAF) module aligns the feature distributions via Wasserstein distance minimization, mitigating imaging variations and further enhancing the discriminative power of the fused features, ultimately classifying the degree of valve calcification. As of the most recent update, no studies have reported the combination of both end-systolic and end-diastolic phases for classifying the degree of AVC.

Methods

Data

The inclusion criteria for this study were as follows: (1) volunteers with varying degrees of AVC and (2) age > 50 years. The exclusion criterion was volunteers who showed unclear 2D ultrasound images. This study strictly followed the inclusion and exclusion criteria and internally collected short-axis aortic echocardiographic images from 160 patients with severe valve calcification, 160 patients with nonsevere valve calcification, and 170 volunteers with healthy valves who were treated at Qilu Hospital of Shandong University from January 2021 to November 2024. All images were acquired and analyzed using Philips ultrasound machines. Among all patients, 420 were randomly chosen to train the DCFFNet neural network, with 2–4 cardiac cycles collected from each patient, and the remaining 70 patients were used to train the YOLOv8 model. Moreover, all patients underwent a CT scan for acquiring the reference calcification volume measure and checking the valve morphology as well. A calcification volume >$600{\text{mm}}^3$ was considered severe AVC, whereas volumes ranging from $200$ to $600{\text{mm}}^3$ were classified as nonsevere calcification. Based on the criteria, this study externally collected data from volunteers with normal valves (N = 72) and nonsevere AVC (N = 67), and patients with severe AVC (N = 69) were used as an external validation set.

Overall Process of the Experiment

Figure 1 depicts the technical workflow of this study. The proposed assisted diagnostic system comprises two components, viz., image preprocessing and DCFFNet.

Fig. 1.

a Flow chart of the study. Homomorphic filtering was used to filter the ultrasound image, and the target region was clipped using the YOLOv8 model. Finally, the DCFFNet neural network was used to classify patients with valve calcification. b Data collection flowchart. Among these, the images in the CT report correspond to the examination results of the ultrasound images shown in the figure

Image Prepossessing

Medical classification tasks often face challenges of limited dataset size and small ROIs to provide decision support. A specially designed data preprocessing pipeline to address these challenges can exert a significant impact on the final results [39].

Homomorphic Filter

Echocardiography is affected by an intrinsic artifact known as “speckle.” Speckle noise present in ultrasound images behaves as a multiplicative noise [40]. A logarithmic transform of the original image can convert the multiplicative noise into an additive component, which can then be treated with conventional frequency-domain filtering techniques.

This study used the homomorphic filtering algorithm for ultrasound noise reduction. The homomorphic filtering algorithm is a frequency-domain filtering technique for images based on the illumination-reflectance model. The illumination-reflection model assumes that a digital image $\text{f}(\text{x},\text{y})$ is composed of an illumination function $\text{i}(\text{x},\text{y})$ and a reflection function $\text{r}\left(\text{x},,,\text{y}\right)$ [41]. In this context, $i(x,y)$ describes the distribution of the illumination intensity on the object, which is independent of the object itself, with$0<i\left(x,,,y\right)<\infty$. The term $r(x,y)$ represents information regarding the surface reflectance of the object, including its details, with$0<r\left(x,,,y\right)<1$. Because illumination changes are relatively small, they can be considered low-frequency components of the image, whereas the reflectance function represents the rapidly changing high-frequency components that vary with the content of the object [42]. Taking logarithms of $\text{f}(\text{x},\text{y})$ turns it into an additive relationship in the spatial domain. A fast Fourier transform is performed on both sides of $\text{ln}f(x,y)$ to transform the image from the spatial domain to the frequency domain.

For ultrasound images, the low-frequency components primarily represent the background or homogeneous tissue structure, whereas the high-frequency components predominantly represent the boundaries of the cardiac tissue or calcified regions [40, 43]. Hence, filters should be designed to improve high-frequency components to highlight calcified areas.

Equation (1) filter is used for frequency-domain filtering to effectively improve the high-frequency components:

$$\begin{array}{c}Z=\left(r_h,-,r_l\right)\times \left(1,-,e^{-c\times \frac{D^2}{d_0^2}}\right)+r_l\#\left(1\right)\end{array}$$

where $D$ is the distance from the center of the frequency domain. $d_0$ is the cutoff frequency, which determines the frequency elements to be boosted or suppressed. $r_l\text{and}rh$ control the filter shape and the degree of enhancement of the high and low frequencies. $\text{c}$ is the sharpness of the slopes of the control function. The Fourier inverse transform is applied to the filtered image, converting it from the frequency domain back to the spatial domain to obtain the final filtered image.

YOLOv8

In ultrasound images, the diagnosis of AVC primarily depends on the short-axis view of the aorta, which comprises only one-tenth of the original image area. Therefore, it is crucial to accurately extract this region.

YOLOv8 (You Only Look Once version 8) [44, 45] is an advanced object detection model that builds upon and improves the previous versions of the YOLO series. The model uses a deeper and wider network architecture, which improves its ability to capture intricate details in complex scenes. By incorporating residual connections, YOLOv8 improves the richness and robustness of feature representation. YOLOv8 also uses a Feature Pyramid Network (FPN) [46], enabling it to extract information from feature maps of different scales, thereby improving its ability to detect objects of different sizes. The integration of self-attention mechanisms further increases the model’s ability to capture long-range dependencies, resulting in improved detection accuracy. The abovementioned methods provide a more robust foundation for subsequent analysis and diagnosis.

Main Architecture of DCFFNet

In clinical practice, both end-systolic and end-diastolic ultrasound phases play a vital role in diagnosing the degree of AVC. Therefore, the study decided to implement a dual-channel input design to leverage the features of both phases. Cross-fusion is performed after feature extraction, and the results are used as inputs for subsequent feature extraction, thus enabling the mining of complex integration rules [47]. The structure of DCFFNet, illustrated in Fig. 2, includes a Pointwise Conv [48] that increases the number of channels in the feature map, thereby enabling the model to capture more complex features.

Fig. 2.

Overall model structure. Our proposed DCFFNet. The model adopts dual input channels, adding SEM, DFFM, CFM, and other modules, and finally realizes the three classifications of patients with aortic valve disease. AvgPool is an adaptive average pooling layer with an output size of 1 × 1

The feature extraction module in the network consists of an SEM and a convolution block. The activation function used in the network backbone is Hard Swish [49], which eliminates the need for computing exponential functions, thereby significantly reducing computational complexity. Despite being piecewise linear, this function maintains a smooth transition near zero, facilitating a more effective learning of complex patterns and features.

Squeeze-and-Excitation Module

The squeeze-and-excitation (SE) [50–52] block is a module designed to improve the feature representation capability of convolutional neural networks. It improves the network’s ability to represent features by adaptively recalibrating channel-wise dependencies. The traditional SE block operates in two major stages, squeeze and excitation.

Nevertheless, the traditional SE block has limitations in dynamically adjusting its hyperparameters based on the input data characteristics. This study proposes an SEM structure, depicted in Fig. 3, which combines the SE block with a multihead self-attention mechanism. This combination allows for the adaptive adjustment of reduction and dropout probabilities, thereby increasing the flexibility and performance of the module.

Fig. 3.

SEM structure drawing. The architecture of the squeeze-and-excitation module (SEM). A self-attention mechanism is added to the module, and the reduction and dropout probabilities are dynamically adjusted

Squeeze stage: Global average pooling is applied to the input feature map $\text{X}\in R^{H\times W\times C}$ to produce a global channel descriptor $\text{z}\in R^C$. Excitation stage: The channel descriptor is recalibrated through two fully connected layers and activation functions to generate the channel weights $\text{s}\in R^C$.

The attention of the feature map (Fig. 3, self-attention zoomed in) is computed as a dot product between three inputs, viz., the query (Q), the key (K), and the value of the input with feature dimension d (V) [53].

The scaling factor is processed through the multihead attention mechanism to obtain the attention output A as follows:

$$\begin{array}{c}A=MultiHead\left(\text{Q},,,\text{K},,,\text{V}\right)=Concat\left({\text{head}}_1,,,\cdots ,,,{\text{head}}_h\right)W^O\#\left(2\right)\end{array}$$

Based on the mean of the attention output $\text{mean}(\text{A})$, the reduction and dropout probability are dynamically adjusted as follows:

$$\begin{array}{c}Reduction=\text{max}\left(2,,,\left[\text{mean},\left(\text{A}\right),\times ,10\right]\right)\#\left(3\right)\end{array}$$

$$\begin{array}{c}Dropoutprobability=\text{max}\left(0.1,,,\text{min},\left(0.5,,,\text{mean},\left(\text{A}\right)\right)\right)\#\left(4\right)\end{array}$$

The recalibrated scaling factor is used to rescale the input feature map as follows:

$$\begin{array}{c}{\widehat{X}}_c=X_c\bullet s_c\#\left(5\right)\end{array}$$

The self-attention mechanism captures the global dependencies between the input features, thereby improving the accuracy and robustness of feature recalibration. By introducing the adaptive attention mechanism, the improved SEM significantly improves the flexibility and capability of the model in handling different input features. The module adapts to various input characteristics by dynamically adjusting the reduction and dropout probabilities based on the attention output, thereby improving the effectiveness of feature recalibration and overall performance.

Dual-Feature Fusion Module

Feature fusion [54] is an important technique in deep learning models that combines information from various feature sources to improve the representation and generalization capabilities of the model. To better utilize the effective features from different states of patients’ ultrasound images, the study propose a dual feature fusion module (DFFM) structure, as illustrated in Fig. 4. This module fuses the features of ultrasound images in different patient states and then introduces a channel attention module, which limits the attention weight between 0 and 1 using the sigmoid function to generate the feature scale map.

Fig. 4.

DFFM structure drawing. The architecture of the dual feature fusion module (DFFM)

Ultrasound images from the two states of the patient are concatenated along the channel dimension as follows:

$$\begin{array}{c}{F}_{\text{combined}}=concat\left(x_1,,,x_2\right)\#\left(6\right)\end{array}$$

By introducing the feature scale map, the proposed DFFM can better capture the complex relationship between different features and significantly improve the effectiveness of feature fusion. To avoid the gradient explosion phenomenon during feature extraction, a residual learning block [55] is introduced to add the feature map before fusing it with the feature scale map. This approach helps improve the overall performance of the model by preventing the gradient explosion phenomenon during feature extraction.

Combined Feature Module

In our proposed medical image classification framework, after multilayer feature extraction and feature fusion using SEM and DFFM, the feature vector is obtained through an average pooling layer and then concatenated within the CFM, as illustrated in Fig. 5. These feature vectors are subsequently passed through three convolutional layers with kernel sizes of 3, 5, and 7, respectively, to perform multiscale feature extraction. The resulting feature maps are reduced to a $1\times 1$ dimension via adaptive mean pooling, and the multiscale features of each image are concatenated to form a fused multifeature map. Multiscale feature fusion (MSFF) [56] integrates features from different scales to capture the multi-scale information inherent in the input data, thereby enhancing the richness of the feature representation. This process is mathematically described as:

$$\begin{array}{c}{f}_{\mathit{ms}}=Concat\left({\mathit{Conv}}_s,\left(x\right),,,{\mathit{Conv}}_m,\left(x\right),,,{\mathit{Conv}}_l,\left(x\right)\right)\#\left(7\right)\end{array}$$

where $f_{\mathit{ms}}$ is the feature vector after multiscale feature fusion. The kernel size of ${\mathit{Conv}}_{\text{small}}\left(x\right),$${\mathit{Conv}}_{\text{medium}}\left(x\right)$ and ${\mathit{Conv}}_{\text{large}}\left(x\right)$ are 3,5 and 7, respectively.

Fig. 5.

Architecture of the combined feature module (CFM). The CFM accepts the feature vectors output from the previous modules and performs feature extraction using convolutional blocks of different sizes to increase the sensory field of the model and extract different image features. AvgPool is an adaptive average pooling layer with an output size of 1 × 1

However, due to variations in imaging conditions, anatomical structures, and pathological states, images acquired from the same patient may yield feature distributions with noticeable discrepancies. Such misalignment in the extracted features could adversely affect the subsequent classification task. To address this issue, this study propose an Optimal Transport-based Feature Alignment Fusion (OT-FAF) module `[57], which leverages optimal transport theory to align the feature distributions of the two images prior to fusion, thereby reducing inter-image distribution discrepancies and enhancing the discriminative capability of the fused features.

Specifically, assume that after deep feature extraction, the two images yield feature sets:

$$\begin{array}{c}{f}_1={\left\{f_1^i\right\}}_{i=1}^{1}and{f}_2={\left\{f_2^j\right\}}_{j=1}^2\#\left(8\right)\end{array}$$

where $N_1$ and $N_2$ denote the number of features in each set. This study employ the Wasserstein distance to quantify the discrepancy between these two feature distributions, defined as:

$$\begin{array}{c}W\left(f_1,,,f_2\right)=\underset{\mathrm{{\rm Y}}\in \mathrm{\Gamma }\left(f_1,,,f_2\right)}{\text{min}}{\sum }_{i=1}^1{\sum }_{j=1}^2{\mathrm{{\rm Y}}}_{\mathit{ij}}c\left(f_1^i,,,f_2^j\right)\#\left(9\right)\end{array}$$

where ${\mathrm{{\rm Y}}}_{\mathit{ij}}$ represents the optimal transport plan satisfying $\mathrm{{\rm Y}}\in \mathrm{\Gamma }\left(f_1,,,f_2\right)$ and $c\left(f_1^i,,,f_2^j\right)$ is the cost function, typically defined as the Euclidean distance between features.

During training, the Wasserstein distance is incorporated as a regularization term in the overall loss function:

$$\begin{array}{c}L={\mathcal{L}}_{\mathit{cls}}+\lambda W\left(f_1,,,f_2\right)\#\left(10\right)\end{array}$$

where ${\mathcal{L}}_{\mathit{cls}}$ is the classification loss and $\lambda$ is a balancing parameter. By minimizing this combined loss, the model is guided to align the feature distributions from the two images of the same patient, ensuring a more consistent input for subsequent concatenation and multiscale fusion, which ultimately enhances the overall classification performance.

This approach not only provides a rigorous mathematical framework for aligning disparate feature distributions but also empirically mitigates the adverse effects caused by variations in imaging conditions and anatomical differences. Similar methodologies have demonstrated promising results in fields such as image generation and style transfer [58], and their application in medical image classification further underscores the potential for cross-modal data fusion.

Data Dimensionality Reduction and Visualization

In the experimental design, to better understand the impact of each module in DCFFNet on the classification results, this study used t-distributed stochastic neighbor embedding (t-SNE) [59] to downscale the feature vectors and visualize the analysis. t-SNE is a probabilistic nonlinear downscaling technique that maps high-dimensional data into a low-dimensional space simultaneously preserving the local structure between data points to the maximum possible extent. First, this study calculated the similarity in the high-dimensional space. For each data point $x_i$, this study calculated its conditional probability $p_{j|i}$ with other data points $x_j$, which denotes the probability that the data point $x_j$ is a neighbor of $x_i$ in the high-dimensional space as follows:

$$\begin{array}{c}{\text{p}}_{\text{j}|\text{i}}=\frac{\text{exp}\left(-,\frac{{\Vert {\text{x}}_{\text{i}}-{\text{x}}_{\text{j}}\Vert }^2}{2{\mathrm{\sigma }}_{\text{i}}^2}\right)}{{\sum }_{\text{k}\ne \text{i}}\text{exp}\left(-,\frac{{\Vert {\text{x}}_{\text{i}}-{\text{x}}_{\text{k}}\Vert }^2}{2{\mathrm{\sigma }}_{\text{i}}^2}\right)}\#\left(11\right)\end{array}$$

where ${\mathrm{\sigma }}_i$ is the bandwidth parameter associated with $x_i$. In the low-dimensional space, for each pair of data points $y_i$ and $y_j$, the study calculated their corresponding similarity $q_{\mathit{ij}}$ as follows:

$$\begin{array}{c}{\text{q}}_{\text{ij}}=\frac{{\left(1,+,{\Vert {\text{y}}_{\text{i}}-{\text{y}}_{\text{j}}\Vert }^2\right)}^{-1}}{{\sum }_{\text{k}\ne \text{l}}{\left(1,+,{\Vert {\text{y}}_{\text{k}}-{\text{y}}_{\text{l}}\Vert }^2\right)}^{-1}}\#\left(12\right)\end{array}$$

The mapping from the high-dimensional space to the low-dimensional space is achieved by minimizing the Kullback–Leibler (KL) scatter of the distribution of data points in the high-dimensional and low-dimensional spaces as follows:

$$KL(P||Q)=\sum _i\sum _j{p}_{\mathit{ij}}\mathit{log}\frac{p_{\mathit{ij}}}{q_{\mathit{ij}}}(13)$$

where $p_{\mathit{ij}}=\frac{p_{i|j}+p_{j|i}}{2N}$, and N is the total number of data points.

Results

The demographics of the subjects are presented in Table 1. No significant difference was observed in demographics between patients with severe and nonsevere valve calcification.

Table 1.

Demographics

	Total	Severe	Nonsevere	P
Gender	70.7% (male)	71.4% (male)	69.8% (male)	0.423
Age	72.3 ± 6.3	72.6 ± 6.6	72.1 ± 5.9	0.29
Height	[162, 173]	166.1 ± 7.9	167.6 ± 7.3	0.695
Weight	[60, 70.5]	66 ± 9.4	67.2 ± 11.3	0.541
HR	79.3 ± 14	81.9 ± 15.4	[65.85]	0.422
SBP	125.8 ± 20.7	123.7 ± 18.1	128.6 ± 23.8	0.067
DBP	[63,76.8]	70.7 ± 9.5	68.8 ± 12.6	0.286

These data indicate the demographics of patients, in which the percentage of gender is the proportion of male patients. Data conforming to normal distribution are represented by mean ± SD, and other data are represented by [Q2, Q3], where HR is the heart rate, SBP is the systolic blood pressure, and DBP is the diastolic blood pressure. There are no data with significant differences in baseline characteristics

Statistical Analysis

Data are expressed as mean ± SEM or [Q1, Q3]. A single linear regression analysis was conducted on the baseline characteristics of the study subjects. Simple logistic regression was performed on the ultrasound diagnostic variable parameters. Multiple logistic regression was applied to evaluate the diagnostic scores combining multiple ultrasound variables. Receiver operating characteristic (ROC) curves were plotted, and the area under the curve (AUC) was calculated to compare the diagnostic performance across different methods. All statistical analyses were conducted using Python, GraphPad Prism (version 5.2.1), and SPSS 29. Demographic data were analyzed using GraphPad Prism (version 5.2.1) and SPSS 29, and model evaluation was performed in Python. Statistical significance was defined as P < 0.05.

Preprocessing Results

The Result of Homomorphic Filtering

The homomorphic filter can effectively suppress the speckle multiplicative noise in ultrasound images and improve the image details.

The Result of YOLOv8

In ultrasound images, the diagnostically significant ROI for patients occupies only approximately one-tenth of the overall image. Therefore, accurately and completely cropping the ROI can effectively mitigate the impact of irrelevant areas and image artifacts on the diagnosis, rendering the diagnostic results of the model more consistent with the actual condition of the patients’ valve calcification. In this study, 1757 ultrasound images were collected and split into a training set and validation set in a 7:3 ratio, with annotations provided by physicians with more than 10 years of clinical experience.

The optimal model achieved an accuracy of 0.88 and a validation loss of 0.67. It also used the optimal model to crop the ultrasound images for subsequent experiments (N = 2010). After review by physicians, 1840 images met the criteria for subsequent experiments, resulting in a cropping success rate of 92%.

Model Evaluation

After image preprocessing, 2184 ultrasound images were used for model training and evaluation. This dataset included 1092 ultrasound images from both end-systolic and end-diastolic phases. The images were categorized into the following three groups: severe valve calcification (N = 360), nonsevere valve calcification (N = 365), and healthy valves (N = 367). The dataset was split into training and validation sets in an 8:2 ratio.

ROC curves were generated to demonstrate the accuracy of the classification scoring model developed in this study (Fig. 6). This model aimed to distinguish between healthy adults and patients with AVC and further distinguish between patients with nonsevere and severe AVC. Figure 6 also compares the classification results of our model with those obtained using conventional ultrasonography and hemodynamic parameters in patients with nonsevere versus severe AVC. To distinguish between healthy adults and patients with AVC, as well as between patients with nonsevere and severe AVC, our model achieved AUC values of 0.99 and 0.94, respectively, with statistical significance (P < 0.01). A review of ultrasound and hemodynamic diagnostic factors revealed AUC values of 0.67, 0.65, and 0.65 for aortic valve area, mean pressure gradient across the aortic valve, and peak flow velocity, respectively, in classifying patients with nonsevere versus severe AVC (P < 0.05). Multiple logistic regression analysis of these three parameters yielded an AUC of 0.75. Moreover, as depicted in Fig. 6, two deep learning classification models were used, ResNet [55, 60] and Vision Transformer (Vit) [61], for classification training on ultrasound images from both end-systolic and end-diastolic phases. Remarkably, the AUC values indicate that our model is highly accurate in diagnosing AVC. Furthermore, our model demonstrated greater accuracy in classifying patients with AVC than other models using ultrasound parameters.

Fig. 6.

ROC curve. a Classification of severe aortic valve calcification and nonsevere aortic valve calcification by ultrasound diagnostic parameters (echo1: aortic valve area; echo2: mean pressure gradient across the aortic valve; echo3: peak flow velocity). b Classification of healthy adults versus patients with aortic valve calcification using different models (model1: Resnet50 targets end-systole; model2: Vit targets end-diastole; model3: Vit targets end-systole; model4: Resnet50 targets end-diastole). c Classification of severe aortic valve calcification and nonsevere aortic valve calcification using different models. d ROC curves on the external validation set. Healthy or not: the model determines whether the patient’s valve is healthy; severe or not: the model determines whether the patient’s valve calcification is severe

To test the performance of the ROC curves of the different diagnostic methods, this study performed a Delong test on our model, ultrasonic diagnostic parameters, and three ultrasonic diagnostic parameters, and the results are presented in.

In Table 2, the ROC curve of our model was significantly different from those of the two deep learning classification models, and the performance significantly improved. Figure 7 depicts the classification confusion matrix for classification between our model and two deep learning classification models on the verification set.

Table 2.

Model performance

	Precision	Accuracy	Recall	F1-Score
Resnet (diastole)	74.08	74.61	73.93	73.93
Resnet (systole)	77.99	77.99	77.99	77.81
Vit (diastole)	67.82	68.37	67.67	67.59
Vit (systole)	75.88	76.17	76.02	75.89
Echo	72.41	73.29	57.89	85.14
DLFFNet	82.5	82.43	91.25	82.5
Our model (diastole)	95.65	95.4	95.41	95.46
Our model (systole)	93.48	93.5	93.49	93.58
No preprocessing	92.47	92.61	93.69	92.63
No dynamic adjustment	96.68	96.39	96.73	96.34
Our model	98.59	96.79	97.22	97.97

Our model (diastole) and (systole) are single-input modes for diastolic and systolic images, respectively. No preprocessing is based on the original model without the image preprocessing process. No dynamic adjustment is based on the original model without the dynamic adjustment function in the SEM

Fig. 7.

Confusion matrix. a Resnet50 targets end-systole. b Vit targets end-diastole; model3. c Resnet50 targets end-diastole. d Vit targets end-systole. e Our model. f our model on the external validation set. (S: severe aortic valve calcification. M: nonsevere aortic valve calcification. N: aortic valve healthy adults)

Correlation Analysis

As shown in Fig. 8a, correlation analysis was performed between ultrasonic diagnostic parameters and other ultrasonic parameters and the calcification volume derived from CT scans. The results demonstrated that the correlation coefficient between each parameter and the CT calcification volume was not high. Figure 8b illustrates a Pearson correlation of AVA versus the CT calcification volume fraction, where $\text{r}=-0.277(P<0.01)$. AVA correlated with the degree of valve calcification; however, the correlation was not significant. To further validate the findings, the results based on CT calcification volume (serving as the control group) were compared with those obtained from the model, and an odds ratio (OR) of 0.944 was calculated. This OR suggests that our model aligns closely with the CT calcification volume, indicating a strong diagnostic agreement.

Fig. 8.

Correlation analysis. a Correlation analysis of ultrasound diagnostic parameters, other ultrasound parameters, and CT calcification volume (AVA: aortic valve opening area; mean PG: mean trans valve differential pressure; Vmax: peak blood flow velocity; AVAD: aortic ring diameter; AAO: ascending aorta diameter; AOD: sinus diameter; STJ: sinus canal junction diameter; CT: CT calcified volume). b AVA vs. CT pearson correlation test plot

Comparisons with State-of-the-Art

Moreover, to further establish the accuracy of our model, it was compared with conventional ultrasound diagnostic parameters and contemporary state-of-the-art classification models. The results, as shown in Table 2, demonstrated that our model outperformed previous models. To further verify the significance of the advantages of the proposed model in this study, a Delong Test was conducted as shown in Table 3.The advantages of the DCFFNet model were statistically significant.

Table 3.

Delong test for different ROC curves

	Resnet (diastole)	Resnet (systole)	Vit (diastole)	Vit (systole)	Echo	DLFFNet	Our model (diastole)	Our model (systole)	No preprocessing	No dynamic adjustment	Ours
P	*	**	**	**	***	*	*	*	*	*	/

P < 0.05 indicates that our model is significantly different from ultrasonic parameters

Ablation Study

Ablation studies have been widely used in the field of neuroscience to tackle complex systems such as the extensively investigated Drosophila central nervous system, the human brain, and artificial neural networks [62]. Therefore, this study conducted ablation experiments to explore the necessity of each step in the model. As illustrated in Table 2, in the single-input channel mode, the model maintained a high level of accuracy when using either end-diastolic or end-systolic images; however, its overall performance was significantly lower than that of the dual-input channel model. Furthermore, to further verify that the absence of different modules has a significant impact on the model, this study conducted DeLong tests. As shown in Table 3, after removing each module, the ROC curves exhibit significant differences compared to those of the original model. Furthermore, the experiment indicated that the image preprocessing step is vital for model performance, and the dynamic adjustment function in SEM provides a degree of performance improvement. In conclusion, these key steps are essential in the model design process and significantly improve the overall performance of the model.

External Data Validation

To validate the generalization capability of the DCFFNet neural network, an external validation set was established using data collected from patients with healthy valves (N = 72), patients with nonsevere AVC (N = 67), and patients with severe AVC (N = 69). During model validation, our model achieved a precision, recall, F1 score, and accuracy of 0.907, 0.88, 0.889, and 0.905, respectively. As depicted in Figs. 6d and 7f, the ROC curves and confusion matrix similarly demonstrate the excellent performance of our model, with the diagnosis of valve calcification showing high accuracy, as evidenced by the AUC values all of which were > 0.90.

Model Visualization

As illustrated in Fig. 9, t-SNE dimensionality reduction and visualization were performed on the feature vectors generated by each module. It found that the feature vectors from individual channels during end-systolic and end-diastolic phases are more dispersed in the two-dimensional plane, exerting a less pronounced clustering effect; the feature vectors from the DFM begin exhibiting a clustering tendency, although some sample points still overlap; and the clustering effect of the feature vectors significantly improves after CFM processing, with different categories of sample points forming a clearer cluster. Although there were sporadic sample misclassification errors, the overall classification accuracy is high, and the boundaries between categories are clear. The experimental results showed that the feature vector processing in the DCFFNet model is hierarchical, progressively improving the clustering effects and confirming the effectiveness of the model in feature extraction.

Fig. 9.

Visualization of dimension-reduced features on the dataset. Dimension reduction is performed by t-SNE

Discussion

AS is a chronic, progressive disease, accounting for a significant proportion of cardiovascular-related morbidity and mortality [63]. The pathogenesis of AS encompasses an initiation phase characterized by mechanical injury, lipid deposition, and a localized inflammatory response. This is followed by a propagation phase, during which progressive valve fibrosis and calcification further exacerbate the aortic valve stenosis [4, 5]. Calcification is vital to the development of AS, and echocardiographic studies should describe the pattern and burden of calcification. Frequently, in clinical work, patients often show inconsistencies between mean transvalvular pressure difference, peak flow velocity, aortic valve area, and dimensionless index [64]; accordingly, it is not possible to determine whether a patient has severe stenosis by depending solely on conventional ultrasound parameters, and an accurate evaluation of the degree of calcification serves as a critical adjunct. Traditional ultrasound diagnostic modalities inadequately evaluate the extent of AVC. Although CT examination can precisely evaluate the degree of AVC, it is not a standard screening technique for aortic valve disease.

This study introduce an AVC classification model that performs two tasks: image preprocessing and classification of the degree of AVC. The model aims to evaluate the degree of AVC based on 2D ultrasound images. By extracting the diagnostic information from the patient’s end-systolic and end-diastolic echocardiographic images, the degree of AVC was comprehensively evaluated, which not only improves the clinical diagnostic efficiency of physicians but also increases the utilization of patients’ echocardiographic images. As demonstrated in Fig. 9a, a comparison with existing related studies revealed that our model exceeded others in accuracy (96.79%), precision (98.59%), F1 score (97.97%), and recall (97.22%). In the intramodel comparison, the model with cross-feature fusion also outperformed the model with a single input. Our model is more efficient in extracting key information from patient images, which is promising to be an important basis for clinical diagnosis by physicians. In addition, Model training and validation were conducted in an Ubuntu 22.04.5 environment using identical hardware (Intel® Core™ i9-10920X CPU; NVIDIA GeForce RTX 3090 GPU), with training time per fold under 25 min; software inference was performed in a Windows 11 environment on an Intel Core i7-13,700 CPU, achieving a runtime below 4 s, fully demonstrating clinical applicability.

The existing evaluation methods for AVC are primarily based on CT scans. Pawade T et al. proposed the use of noncontrast CT scans for AVC scoring (CT-AVA) [65, 66]. Furthermore, Cartlidge T RG et al. utilized contrast-enhanced CT to evaluate both calcified and noncalcified lesions of the aortic valve [67]. Nevertheless, CT is not a routine diagnostic technique for aortic valve diseases. In studies based on echocardiography, Krishna H et al. proposed the use of the Us2.ai AI platform to measure aortic valve area, peak aortic valve velocity, and mean pressure gradient across the valve in patients. Their results demonstrated a similarity coefficient > 0.9 with the measurements taken by physicians [29]. Playford et al. used an AI system to analyze patient echocardiograms without the need for LVOT data, aiding in the evaluation of the severity of AS. Both the sensitivity and specificity for diagnosing AS were $91.4\text{\%}$ [28]. Ahmadi et al. used a transformer-based approach to perform spatiotemporal analysis on two-dimensional echocardiography data from patients. This method effectively and efficiently integrates both anatomical characteristics and the dynamic motion of the aortic valve to classify the severity of AS [26]. However, neither echocardiographic parameters nor the degree of AS alone can accurately reflect the degree of valve calcification in patients. Frequently, the hemodynamic changes of aortic valve-related disease and the parameters measured by ultrasound, as well as the clinical symptoms, often do not coincide. Recently, Tang L et al. developed a novel dynamic local feature fusion network for automatic AVC recognition using echocardiography [35]. They used a U-Net neural network to segment the highly echogenic regions within the aortic valve orifice (dice = 0.7488). Subsequently, a classification model was used to categorize patients (micro-AUC = 92.39; macro-AUC = 92.25). Their study focused on classifying valve calcification based on a single two-dimensional ultrasound image, achieving a high classification accuracy. Nevertheless, their study did not combine the ultrasound images of different states of the patients for a comprehensive analysis. Echocardiographic images of patients taken at different periods of time have abundant diagnostic information. Furthermore, their research did not use CT calcification volume and perform correlation tests on the study subjects, which would be beneficial because CT results are considered the gold standard in AVC scoring. The model demonstrated excellent performance in normal individuals, patients with nonsevere AVC, and patients with severe AVC, with AUC values of 0.99 and 0.94 (P < 0.01). Although, in the external validation set, the AUC and accuracy values of our model decreased slightly, they remained > 0.9 overall. Therefore, the classification results provided by our model serve as a valuable reference for clinical diagnosis by physicians. In the future, physicians could incorporate our method into their clinical practice to more accurately track the progression of calcification in patients with early-stage AS. This tracking is vital because it would enable more timely and targeted interventions, thereby potentially delaying the onset of severe symptoms and reducing the need for emergency surgery. By integrating calcification data with echocardiographic findings, this approach provides a comprehensive strategy for patient management, ensuring that intervention decisions are based on a fuller understanding of the patient’s condition. In the next step, the model was encapsulated into software that allows physicians to select the appropriate end-systolic and end-diastolic images and perform a one-button operation to output the patient’s AVC extent. Incorporation of this approach into the routine follow-up of patients with early-stage AS could provide a more precise and dynamic method to monitor disease progression, thus significantly improving clinical outcomes.

Ultrasound is a popular tool for patient examination due to its convenience and low radiation exposure. Valve evaluation must be performed by experienced and senior physicians, and the reliability of the evaluation depends on the experience and expertise of the operator. Ultrasound-related diagnostic parameters are often used as important criteria by physicians to diagnose the degree of AVC. However, due to variations in cardiac function, volume, and other factors in patients, as demonstrated in Fig. 6a, the use of ultrasound parameters to diagnose the degree of AVC is not reliable; the correlation coefficients between the degree of AVC and each of the diagnostic parameters of ultrasound, as depicted in Fig. 8, are similarly low. In particular, in some patient groups with complex lesions, traditional evaluation methods can no longer fulfill the needs of clinical diagnosis and treatment. Therefore, accurate qualitative analysis of the degree of AVC using echocardiography based on AI has become an urgent challenge in the field of cardiology.

With the exception of patients with severe congenital abnormalities of the aortic valve, most patients with AS have varying degrees of calcification, which can occur early. The aim of our study was also to precisely detect early calcification and enable timely prevention and management of patients with AS that causes hemodynamic abnormalities. Future study will focus on the degree of calcification in patients with different degrees of AS and potentially incorporate methodologies such as biofluid dynamics to more comprehensively evaluate stenotic lesions and provide more reference information for clinical treatment.

Although our AVC classification model can accurately evaluate the degree of AVC and effectively aid physicians in diagnosing the patient’s condition, the following limitations should be acknowledged in our study: (1) the sample size was limited, and the training data lacked data support from more centers and were limited to image analysis from one ultrasound device, (2) the selection of end-systolic and end-diastolic ultrasound images still requires physician involvement in their selection and does not allow for fully automated detection, and (3) in the future, we will promote deep integration of our system with hospital PACS platforms, achieving real-time synchronized storage of imaging data and automatic annotation of results, thereby enhancing the continuity and efficiency of clinical workflows.

Conclusion

This study present a pioneering classification model designed for evaluating AVC, demonstrating its efficacy in accurately evaluating the degree of calcification using two-dimensional echocardiography. This innovative approach equips physicians with an increased understanding of AVC levels, facilitating the creation of precise treatment plans in tandem with the use of ultrasound diagnostic parameters. Using this method, healthcare professionals are empowered to make informed decisions, thereby improving the overall management of patients with AVC simultaneously providing clinicians with an additional tool for the accurate diagnosis of AS.

Funding

The work received partial funding from the Shandong Provincial Natural Science Foundation General Project (ZR2021MH126, ZR2021MF037, ZR2023MH121) and the National Natural Science Foundation of China (82270457, 82470499).

Data Availability

The datasets processed and analyzed during the current study are not publicly available due to institutional policies and patient privacy regulations but are available from the corresponding author upon reasonable request.

Declarations

Ethics Approval

This study adhered to the principles outlined in the Declaration of Helsinki and was approved by the Ethics Review Committee of Qilu Hospital of Shandong University, with project number KYLL-202312–025.