Classification of mitral regurgitation in echocardiography based on deep learning methods

Introduction

Mitral regurgitation (MR) is the most common valvular heart disease; if remained untreated, it can lead to impaired cardiac function and, in severe cases, heart failure or death (1). The incidence of MR correlates with increased age, and considering a progressively aging society, the number of patients requiring intervention for MR continues to increase. Professor Carpentier, a pioneer of mitral valve repair techniques, proposed a standardized functional classification for MR that has significant implications for its diagnosis and treatment (2,3). Carpentier’s classification heavily relies on information gleaned from echocardiography. It stratifies MR based on the activity of the mitral valve leaflets, categorizing degenerative mitral regurgitation (DMR) (4) as type II, and designating atrial functional mitral regurgitation (AFMR) (4-8) and ventricular functional mitral regurgitation (VFMR) (4,9,10) as type I.

The accurate classification of MR is key for the development of effective treatment plans and is clinically imperative. Different treatment strategies may be considered, depending on the specific type of MR. For cases involving leaflet perforation or inadequate leaflet coaptation owing to excessive leaflet motion (e.g., mitral valve prolapse), interventions like the MitraClip procedure may be a viable option (2,3). Conversely, in situations where MR is primarily a result of valve dysfunction leading to restricted leaflet motion (e.g., rheumatic MR), valve replacement therapy may be considered (11). The selection of the most suitable treatment strategy depends on the precise diagnosis and classification of MR, which underscores the clinical significance of the Carpentier classification system (11).

Echocardiography is currently the most important imaging modality for diagnosing and assessing MR. It is widely used to determine the etiology of MR, qualitatively and quantitatively assess the severity of MR, and evaluate changes in cardiac hemodynamics (12-14). According to the 2020 updated American College of Cardiology Expert Consensus on Mitral Regurgitation Management (2) and the American Society of Echocardiography Guidelines on Echocardiography (12), the analysis of MR via echocardiography evidently requires a high degree of proficiency from cardiac sonographers and specialists. This analysis significantly relies on the expertise of skilled healthcare professionals. Therefore, there is a clinical need for a standardized, unified, and intelligent decision-support system to augment the precision and efficiency of MR diagnoses. This system has the potential to assist healthcare professionals in obtaining more precise and reliable measurements, ultimately enhancing patient care relative to MR.

Recently, the utilization of deep learning technology has improved the speed and efficiency of disease diagnosis and treatment (15), and has been increasingly applied to image and video analysis (16), becoming one of the most widely used and effective methods in automatic medical image analysis (17). Deep learning methods are applied in many medical imaging methods, such as assessing heart failure through cardiothoracic ratio in chest X-rays (18,19), automating the measurement of coronary artery stenosis in computed tomography (CT) images (20,21), and detecting chronic myocardial infarction in magnetic resonance imaging (MRI) (22). In the field of echocardiography, Deep learning methods have also made significant advancements and are primarily applied in tasks like the classification of echocardiogram views (23-25), assessment of echocardiogram video quality (26), segmentation of left atrium (LA) and left ventricle (LV) (27-31), detection of ventricular dysfunction (32,33), computation of ejection fraction (34,35), and classification of cardiovascular disease (36,37).

Some researchers have focused on segmenting and locating the mitral valve using echocardiographic images. Vafaeezadeh et al. (38) proposed the CarpNet model, which uses the Inception-ResNet-v2 architecture to classify normal mitral valve function (NM) and three types of MR (prolapsed, rheumatic, and ischemic) in the parasternal long-axis (PLAX) echocardiographic view. In addition, they utilized the ResNeXt50 architecture to classify NM valves and two types of MR (rheumatic and ischemic) in a four-chamber echocardiographic view, achieving precision rates of 69% and 80%, respectively. Moghaddasi et al. (39) proposed two new texture features to classify the severity of apical four-chamber (A4C)/apical two-chamber (A2C)/parasternal short-axis (PSAX) views using Support Vector Machine (SVM); the precision of NM, mild, moderate, and severe MR was 99.52%, 99.38%, 99.31%, and 99.59%, respectively. Edwards et al. (40) used a convolutional neural network Convolutional Neural Network (CNN) to detect MR in the PLAX view with a precision of 0.86. Wu et al. used the ResNet18 model to classify fibroelastic deficiency, Barlow’s disease, and NM in the A2C, A3C, and A4C views, achieving a precision of 0.95. Wu et al. (41) designed the ARUNet-MTL model to classify NM, DMR, and functional MR in A2C, A3C, and A4C views; the model segments the LA, LV, and mitral valves. It achieved a classification precision of 0.918 and segmentation Dice coefficients of 0.9438 for the LA, 0.9157 for the LV, and 0.7951 for the mitral valve.

Currently, research on the automatic classification of MR based on Carpentier’s functional classification remains ongoing, and there is significant room for improvement in terms of identifying the types of MR using echocardiography. This study was aimed at developing a four-classification model, encompassing NM and three types of MR (DMR, AFMR, and VFMR). According to the Carpentier functional classification, different types of MR exhibit distinct patterns of activity in the LA, LV, and mitral valve, all of which are reflected in echocardiographic videos. Consequently, the classification model should focus on the relevant regions and their characteristic activities in echocardiographic images. Based on the distinct activity patterns exhibited by the LA, LV, and mitral valve for different types of MR, and combined with the diagnostic methods of experienced medical professionals, we designed the Attention-guided Residual Multiscale Feature Fusion Network (ARMF-Net). The key highlights are summarized as follows: cascading of different residual modules, integration of multiple attention mechanisms, and fusion of auxiliary feature information from various modalities. The ARMF-Net guides the model to focus on relevant regions and their characteristic activities, thereby improving the diagnostic efficiency. We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-120/rc).

Methods

Dataset

The data utilized in this study were sourced from the Cardiac Ultrasonography Department at Zhongshan Hospital, affiliated with Fudan University. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Zhongshan Hospital (No. B2022-356R) and individual consent for this retrospective analysis was waived. Echocardiographic video data were collected using ultrasound scanners of various brands, such as GE, Philips, and Siemens. Disease diagnoses were further confirmed by experts with over 10 years of experience in the echocardiographic diagnosis of mitral valve diseases. The final dataset consisted of 266 participants: 59 with NM, 67 with DMR, 80 with VFMR, and 60 with AFMR. As shown in Table 1, 754 video clips were captured in the A2C, A3C, and A4C views. Each video clip was 2 seconds in duration and included one or more complete cardiac cycles.

MR classification network

Pre-processing

Before further analysis, all echocardiography videos underwent pre-processing. For each echocardiographic view, we applied our segmentation network (42) to each video frame to segment the LA and LV. The Dice coefficients of the segmentation network for the LA and LA were 0.935 and 0.915, respectively, and the segmentation results were examined by professionals. Subsequently, we calculated the changes in their respective areas as well as the area difference between the LA and LV. These parameters generate curves that depict the temporal variations in the left atrial area (C_LA), left ventricular area (C_LV), and the difference between the left ventricular and left atrial areas (C_AV). Based on the C_LA and C_LV, four key image frames were extracted for both the end-diastolic (maximum area of the LA or LV) and end-systolic phases (minimum area of the LA and LV). Twenty frames centered around the four selected key frames were chosen as inputs for the subsequent video network, and data augmentation operations were performed on the selected key frames, including random rotation and random horizontal flipping. All image frames were resized to 256×256 pixels, and their grayscale values were normalized to the range [0,1].

Overall structure

The network model constructed in this study can classify the four types of MR (DMR, AFMR, VFMR, and NM) in echocardiogram videos from three different views [A2C, apical three-chamber (A3C), and A4C].

The frame sequences of the input video contained complex spatial and temporal motion information. To effectively extract multiscale feature information, we employed a deep convolutional network consisting of multiple concatenated modules. Each module utilized a different number of 3D convolutions with residual connections, facilitating more efficient features and gradient transmission while mitigating issues such as gradient vanishing and explosions. As the features used for MR classification were derived from the LA, LV, and mitral valve regions across different echocardiogram views, excess attention from the network to the right heart chambers or other valves beyond the mitral valve during training could increase the risk of overfitting and classification errors, thereby weakening the generalization ability of the model and reducing its performance. In this regard, we incorporated multiple attention mechanisms, in which the attention maps obtained from both the channel and spatial dimensions were multiplied by the input feature maps, enabling adaptive feature optimization.

VFMR is characterized by LV dilation or weakened contraction, whereas the LA function remains normal. Conversely, AFMR involves LA enlargement and reduced contraction with normal LV function. Based on the activity characteristics of the LA and LV during the occurrence of these two types of MR, we focused on the C_LA, C_LV, and C_AV. After multiple one-dimensional convolutions, these features were concatenated, and position encoding was added before being input into the transformer encoder module, enabling the obtaining of auxiliary features from different modalities.

In the final stage of the model, the video and auxiliary feature information were fused with different weight values assigned to each component. This combined information was then processed through a fully connected layer. The overall structure of the model is shown in Figure 1, with annotations provided for each section.

Figure 1 Overall Architecture of the ARMF-Net. C_LA, C_LV, and C_AV are curves depicting the temporal variations in left atrial area, left ventricular area, and the difference between left ventricular and left atrial areas, respectively. AFMR, atrial functional mitral regurgitation; C_AV, curve that depicts the difference between the left ventricular and left atrial areas; C_LA, curve that depicts the temporal variations in the left atrial area; C_LV, curve that depicts the temporal variations in the left ventricular area; DMR, degenerative mitral regurgitation; NM, normal mitral valve function; VFMR, ventricular functional mitral regurgitation;

Video feature extraction

The video feature extraction component was composed of multiple cascaded modules, including Res-Conv3dBlock and ResAttn-Conv3dBlock, as illustrated in Figure 2.

Figure 2 Structure of different modules. (A) The structure of the Res-Conv3dBlock. (B) The structure of the ResAttn-Conv3dBlock. (C) The structure of channel attention, including dimensional changes. (D) The structure of spatial attention, including dimensional changes. ReLU, rectified linear unit.

The Res-Conv3dBlock module consists of two 1×1×1 3D convolutional layers, one 3×3×3 3D convolutional layer, batch normalization, and a rectified linear unit (ReLU) activation function that extracts video features without altering the input dimensionality.

The ResAttn-Conv3dBlock module follows the Res-Conv3dBlock and incorporates both channel and spatial attention mechanisms. The channel attention mechanism compresses the feature maps along the spatial dimension by utilizing the average and max pooling layers to aggregate the spatial information from the feature maps. This process reduces the spatial dimensionality of the input features and is followed by element-wise summation for merging. Channel attention focuses on the weight of each channel in the input feature map. The spatial attention mechanism compresses the channels by applying average and maximum pooling along the channel dimensions, highlighting the important content within the feature maps. Average pooling provides feedback for each pixel in the feature map, whereas max pooling offers only gradient feedback from the locations with the highest responses during backpropagation.

Auxiliary information feature extraction

The auxiliary information feature extraction section (Figure 1) primarily focused on the activity characteristics of the LA and LV during the occurrence of VFMR and AFMR. During the video pre-processing stage, three key curves were obtained from the echocardiographic video: C_LA, which reflects the diastolic and systolic functions of the LA; C_LV, which indicates the diastolic and systolic functions of the LV; and C_AV, which represents the difference in the area between the LA and LV, indicating any enlargement.

Each of these three input curves underwent two iterations of one-dimensional convolution (with a kernel size of three), followed by batch normalization and the ReLU activation function. The resulting outputs were then concatenated and fed into the transformer encoder module, which consists of eight attention heads and six encoder layers, ultimately producing an auxiliary information feature vector.

Modal fusion and classification

The video and auxiliary information feature vectors were fused using a weighted concatenation method. The fusion operation is defined as follows:

ffusion=concat(W1f1,W2f2)[1]

where W₁ and W₂ represent the different weight values, f₁ represents the video feature vector, and f₂ represents the auxiliary information feature vector.

The resulting fused feature vector was then passed through a fully connected layer, followed by a softmax layer, to perform the classification task (Figure 1). This architecture enables the model to leverage both the temporal dynamics captured in the video features and the relevant auxiliary information, thereby improving the precision and robustness of the classification of mitral valve regurgitation types.

Training strategies

The dataset was randomly split into training and validation set (85%), and test set (15%) based on patient IDs, the distribution of the divided data was shown in Table 2; efforts were made to ensure that the NM and the three types of MR maintained the same ratio in the split. We conducted a five-fold cross-validation on the training and validation set, with a training and validation ratio of 75% and 10%, respectively.

First, we trained the video feature extraction components independently. After completing the training, an auxiliary feature extraction network was incorporated and the parameters of the video feature extraction network were frozen. The training process lasted for 120 epochs, and the learning-rate adjustment utilized the ReduceLROnPlateau method with an initial value of 1e−2. The learning rate decreased when the loss function did not improve for 15 epochs. The weight decay was set to 5e−3, and the dropout rate was set to 0.2. The cross-entropy loss function was used. The weight values for multimodal fusion were set as W₁ =0.6, W₂ =0.4. During the training process, data augmentation techniques such as random rotation and random flipping were applied to the input image data to enhance the generalization capability of the model.

Evaluation metrics

In this study, precision (Eq. [2]), F1-score (Eq. [3]) were used as performance evaluation metrics. F1-score is used to measure the classifier’s overall performance.

Precision=TPTP+FP[2]

Where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively.

F1-score=2×Precision×RecallPrecision+Recall=TPTP+0.5×(FP+FN)[3]

Where Recall=TPTP+FN.

Results

In the paper, the training and testing sets were divided by participant level, typically consisting of 1–3 different perspectives of several echocardiography videos per participant. For ARMF-Net, each participant’s video is treated as a separate input. When exploring the performance of the model in MR classification, evaluation metrics were calculated at both the video and participant levels. At the participant level, the model adds up the predicted classification probabilities of multiple videos corresponding to each participant, and then selects the category with the highest total occurrence probability as the model’s classification result for that participant. Table 3 presents the results of five-fold cross validation. At the video level, ARMF-Net achieved an overall precision of 0.913±0.024 for the classification results of NM and three different types of MR on the test dataset, with the highest precision reaching 0.93; The highest precision at the participant level is 0.961. Figure 3A demonstrates the confusion matrices for the four classes, with the predicted labels on the horizontal axis and the true labels on the vertical axis. It is noted that misclassifications related to DMR and NM were rare, whereas the majority of prediction errors were concentrated within the AFMR and VFMR classes.

Figure 3 Classification results. (A) The confusion matrix of the classification results of ARMF-Net. (B) t-SNE visualization results. AFMR, atrial functional mitral regurgitation; ARMF-Net, Attention-guided Residual Multiscale Feature Fusion Network; DMR, degenerative mitral regurgitation; NM, normal mitral valve function; t-SNE, t-distributed stochastic neighbor embedding; VFMR, ventricular functional mitral regurgitation.

We analyzed 115 videos from 42 cases in an independent test set and found that in the VFMR prediction errors, the left ventricular field of view in the corresponding videos was very small, and there were significant noise artifacts that affected the model’s judgment; In the data with AFMR recognition errors, the left atrial region is incomplete, which affects the model’s judgment. In 42 cases, the model accurately detected MR in patients, but there were misjudgments between individual categories.

We concatenated the final output features of the model into a list, applied the k-means clustering method, and used t-distributed stochastic neighbor embedding (t-SNE) to visualize the distribution of different classes with the parameters set to n_clusters =4, and random_state =42. As shown in Figure 3B, our method effectively distinguished the different classes.

Discussion

Comparison and analysis

Comparison with other studies

In this section, we compare the classification results of the proposed method with those of other studies. Given the lack of research that matches the content of our study exactly, we recalculated the results for the same classification targets using confusion matrices provided in the original texts of other researchers to enable a meaningful comparison.

It can be seen from Table 4 that researches regarding echocardiography-based MR has predominantly focused on the NM and DMR, with different studies employing varying views.

Compared to other classification studies, our model not only performed well in classifying VFMR and AFMR, but also demonstrated a better performance in distinguishing NM and DMR. Our model performs classification across three different views, offering a greater applicability. It detects the three categories of MR and NM using 3D convolutional layers to extract features from the temporal dimension. We extracted multimodal data specific to the pathological features of AFMR and VFMR, and performed multimodal data fusion.

Comparison with existing video classification models

We compared the performance of several video classification networks, training and testing them on the same dataset. The training strategy was identical to that described in Training Strategies section; the results are presented in Table 5. Compared with these video classification networks, the ARMF-Net proposed in this study exhibited a more substantial improvement in classification precision, which was observed across different types of MR and the classification of NM. The model proposed in this study demonstrated a high precision, affirming the effectiveness of the algorithm.

The ARMF-Net exhibits a significant improvement in the classification precision for two specific types of MR (AFMR and VFMR). This enhancement was attributed to the auxiliary feature extraction network branch, which appears to capture and refine the abstract depth information related to the diastolic and systolic capabilities of the LA and LV, particularly over a complete cardiac cycle. Furthermore, for the classification of DMR and NM, our proposed model achieved a comparable or even better classification precision.

Ablation studies

This section discusses the effectiveness of our approach in improving the classification performance. First, we present the most basic multilayer 3D convolutional model and progressively validate the impact of the pre-processing process, residual modules, attention mechanisms, and fusion of different modality feature information on the performance, as shown in Table 6.

Effectiveness of the pre-processing

As indicated in the “Pre-processing” section, each frame of the echocardiogram video was segmented using an established segmentation network. Based on the area changes of the LA and LV, we selected 20 frames, including those at the end-diastole and end-systole of the LA and LV, as the input for the video network. We believe that the frames in echocardiograms corresponding to the end-diastole and end-systole of the LA and LV best reflect the motion features of these chambers, as well as the amplitudes of the mitral valve opening and closing.

To verify the effectiveness of the pre-processing procedure, we conducted ablation experiments on a basic multilayer 3D convolutional model using three sampling methods: random frame sampling, equidistant frame sampling, and pre-processing sampling, corresponding to experiment 1, experiment 2 and experiment 3, respectively, in Table 6. The results showed that the pre-processing sampling method achieved a better performance within the same model, with a precision improvement of 0.4–0.6.

Effectiveness of the residual module and attention mechanism

Building on pre-processing sampling, we designed our model based on a basic multilayer 3D convolutional model by concatenating residual modules that contain varying numbers of 3D convolutional layers and attention mechanisms. This design was used for feature extraction and classification of the input video. As shown in experiment 4 in Table 6, this approach of concatenating multiple modules achieved a better performance, with a precision improvement of 0.12.

This study also analyzed the effectiveness of attention mechanisms via network visualization. The gradient-weighted class activation mapping (Grad-CAM) visualization method was employed to generate heat maps indicating the regions of interest within the video images as perceived by the network. Figure 4 provides a visual comparison, with the upper and lower portions displaying the heat map results before and after integration of the attention mechanisms for the same echocardiographic video. With the integration of attention mechanisms, the network demonstrated an increased capacity to concentrate on key areas relevant to different types of MR. For instance, it focused on the mitral valve for DMR, the left ventricular area for VFMR, and the left atrial area for AFMR. This visualization underscores the efficacy of the attention mechanisms in guiding the attention of networks to critical regions.

Figure 4 The heatmap results using Grad-CAM, verifying the effectiveness of the attention module. (A-C) Results of model without attention: (A) DMR, A4C; (B) AFMR, A3C; (C) VFMR, A2C. (D-F) Results of model with attention: (D) DMR, A4C; (E) AFMR, A3C; (F) VFMR, A2C. A2C, apical two-chamber; A3C, apical three-chamber; A4C, apical four-chamber; AFMR, atrial functional mitral regurgitation; DMR, degenerative mitral regurgitation; Grad-CAM, gradient-weighted class activation mapping; NM, normal mitral valve function; VFMR, ventricular functional mitral regurgitation.

Effectiveness of the auxiliary information feature extraction and modal fusion

In our pre-processing, we obtained the changes in the LA (C_LA) and LV (C_LV) areas, as well as the changes in the difference between the LA and LV (C_AV) areas, and used these as auxiliary information inputs from different modalities.

Referring to the diagnostic process of echocardiography experts, when determining AFMR and VFMR, they observed changes in the diastole and systole of the LA or LV over a complete cardiac cycle, as well as whether the volume of the LA or LV was abnormal. These aspects are reflected in the C_LA, C_LV, and C_AV.

We froze the parameters of the video network and added an auxiliary information feature extraction network as well as a weighted feature fusion network. Ablation experiments were conducted, demonstrating that the overall precision improved by 0.11 (experiment 5 in Table 6). Notably, the classification precision for AFMR and VFMR improved by 0.1 and 0.03, respectively.

Generalization ability

To further verify the validity of our model for classifying MR, we collected a multicenter combined dataset consisting of 52 participants, including 23 cases of DMR, 20 cases of AFMR, 30 cases of FMR, and 22 cases of NM, with a total of 95 video clips. The network achieved a classification precision of 86.3%. The confusion matrix shown in Figure 5 indicates that our method demonstrates a certain degree of generalization ability.

Figure 5 The confusion matrices of the classification results on multi-center data. AFMR, atrial functional mitral regurgitation; DMR, degenerative mitral regurgitation; NM, normal mitral valve function; VFMR, ventricular functional mitral regurgitation.

Limitations and future scope

This study has several important limitations that warrant consideration as well as areas for future improvement.

Variability in data quality

The data were obtained from the echocardiography department of the hospital, and the quality of the ultrasound videos can vary depending on the skill levels of the different medical professionals. Variability in data quality can introduce challenges such as unclear image edges, increased noise, and ultrasound artifacts in videos, which may affect the performance of the model. In future studies, pre-processing steps should be implemented on echocardiographic videos to enhance their quality, ensure clarity, and minimize noise as well as ultrasound artifacts.

Device-specific data distribution

The distribution of echocardiographic videos collected from different devices was uneven, with certain devices containing significantly few videos. This imbalance in data distribution may potentially affect the ability of the model to be generalized across different imaging devices.

Limited sample size

The dataset used in this study had a relatively small sample size, with only A2C, A3C, and A4C views included. It did not cover a wider range of echocardiographic views, and not all participant s had complete videos of the three views. This limitation may affect the generalizability of the algorithm. Further data collection with a larger and more diverse sample size is required to validate the effectiveness of the model.

Lack of large-scale public datasets

Due to the lack of publicly available annotated MR datasets, we were unable to validate our model on large-scale datasets, limiting our ability to thoroughly assess the generalizability and robustness of the model across different patient populations and imaging conditions. In the future, we will continue to collect additional multicenter combined datasets to improve this aspect of the study.

Model optimization

The ARMF-Net model utilizes the initial segmentation information of the LA and LV for AFMR and VFMR, but does not separately add the activity information of the mitral valve. In future research, we will add more information and integrate it into the model.

These limitations highlight the need for further research and data collection efforts to refine and validate the performance of the algorithm, especially in real-world clinical settings, where data quality and diversity can considerably vary.

Conclusions

The ARMF-Net model proposed in this study serves as a robust framework for classifying echocardiographic videos containing NM and the three types of MR based on the diagnostic processes of professional physicians. It employs a series of techniques to enhance its classification performance, including video pre-processing, sampling, residual structures, multi-attention mechanism integration, auxiliary information feature extraction, and feature fusion. The model achieved an average classification precision of 0.93 for NM and the three types of MR based on echocardiographic videos. This achievement provides significant support for the automation and rapid screening of echocardiographic image analysis in clinical practice, thereby facilitating a more effective and accurate diagnosis and treatment of patients with MR.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the CLEAR reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-120/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-120/dss

Funding: This work was supported by the National Natural Science Foundation of China (grant No. 61801123), Shanghai Municipal Commission of Economy and Information Technology (grant No. GYQJ-2018-2-05), Shanghai Municipal Science and Technology Major Project (grant Nos. 2017SHZDZX01 and 16441907900), Zhongshan Hospital, Fudan University (grant No. 2022-010), Medical Engineering Fund of Fudan University (grant No. yg2021-38), and Youth Fund Program in Zhongshan Hospital, Fudan University (grant Nos. ZHUPEI2022-010 & N0.2023ZSQN38).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-120/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are 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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Zhongshan Hospital (No. B2022-356R) and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. DesJardin JT, Chikwe J, Hahn RT, Hung JW, Delling FN. Sex Differences and Similarities in Valvular Heart Disease. Circ Res 2022;130:455-73. [Crossref] [PubMed]
2. Bonow RO, O'Gara PT, Adams DH, Badhwar V, Bavaria JE, Elmariah S, Hung JW, Lindenfeld J, Morris AA, Satpathy R, Whisenant B, Woo YJ. 2020 Focused Update of the 2017 ACC Expert Consensus Decision Pathway on the Management of Mitral Regurgitation: A Report of the American College of Cardiology Solution Set Oversight Committee. J Am Coll Cardiol 2020;75:2236-70. Erratum in: J Am Coll Cardiol 2020;75:3000. [Crossref] [PubMed]
3. Wang S, Meng X, Luo Z, Pan X. Transapical Beating-Heart Mitral Valve Repair Using a Novel Artificial Chordae Implantation System. Ann Thorac Surg 2018;106:e265-7. [Crossref] [PubMed]
4. Abe Y, Takahashi Y, Shibata T. Functional mitral regurgitation, updated: ventricular or atrial? J Echocardiogr 2020;18:1-8. [Crossref] [PubMed]
5. Gertz ZM, Raina A, Saghy L, Zado ES, Callans DJ, Marchlinski FE, Keane MG, Silvestry FE. Evidence of atrial functional mitral regurgitation due to atrial fibrillation: reversal with arrhythmia control. J Am Coll Cardiol 2011;58:1474-81. [Crossref] [PubMed]
6. Ito K, Abe Y, Takahashi Y, Shimada Y, Fukumoto H, Matsumura Y, Naruko T, Shibata T, Yoshiyama M, Yoshikawa J. Mechanism of atrial functional mitral regurgitation in patients with atrial fibrillation: A study using three-dimensional transesophageal echocardiography. J Cardiol 2017;70:584-90. [Crossref] [PubMed]
7. Kagiyama N, Mondillo S, Yoshida K, Mandoli GE, Cameli M. Subtypes of Atrial Functional Mitral Regurgitation: Imaging Insights Into Their Mechanisms and Therapeutic Implications. JACC Cardiovasc Imaging 2020;13:820-35. [Crossref] [PubMed]
8. Ring L, Dutka DP, Wells FC, Fynn SP, Shapiro LM, Rana BS. Mechanisms of atrial mitral regurgitation: insights using 3D transoesophageal echo. Eur Heart J Cardiovasc Imaging 2014;15:500-8. [Crossref] [PubMed]
9. Robbins JD, Maniar PB, Cotts W, Parker MA, Bonow RO, Gheorghiade M. Prevalence and severity of mitral regurgitation in chronic systolic heart failure. Am J Cardiol 2003;91:360-2. [Crossref] [PubMed]
10. Trichon BH, Felker GM, Shaw LK, Cabell CH, O'Connor CM. Relation of frequency and severity of mitral regurgitation to survival among patients with left ventricular systolic dysfunction and heart failure. Am J Cardiol 2003;91:538-43. [Crossref] [PubMed]
11. El Sabbagh A, Reddy YNV, Nishimura RA. Mitral Valve Regurgitation in the Contemporary Era: Insights Into Diagnosis, Management, and Future Directions. JACC Cardiovasc Imaging 2018;11:628-43. [Crossref] [PubMed]
12. Zoghbi WA, Adams D, Bonow RO, Enriquez-Sarano M, Foster E, Grayburn PA, Hahn RT, Han Y, Hung J, Lang RM, Little SH, Shah DJ, Shernan S, Thavendiranathan P, Thomas JD, Weissman NJ. Recommendations for Noninvasive Evaluation of Native Valvular Regurgitation: A Report from the American Society of Echocardiography Developed in Collaboration with the Society for Cardiovascular Magnetic Resonance. J Am Soc Echocardiogr 2017;30:303-71. [Crossref] [PubMed]
13. Stone GW, Vahanian AS, Adams DH, Abraham WT, Borer JS, Bax JJ, Schofer J, Cutlip DE, Krucoff MW, Blackstone EH, Généreux P, Mack MJ, Siegel RJ, Grayburn PA, Enriquez-Sarano M, Lancellotti P, Filippatos G, Kappetein APMitral Valve Academic Research Consortium (MVARC). Clinical Trial Design Principles and Endpoint Definitions for Transcatheter Mitral Valve Repair and Replacement: Part 1: Clinical Trial Design Principles: A Consensus Document From the Mitral Valve Academic Research Consortium. J Am Coll Cardiol 2015;66:278-307. [Crossref] [PubMed]
14. O'Gara PT, Grayburn PA, Badhwar V, Afonso LC, Carroll JD, Elmariah S, Kithcart AP, Nishimura RA, Ryan TJ, Schwartz A, Stevenson LW. 2017 ACC Expert Consensus Decision Pathway on the Management of Mitral Regurgitation: A Report of the American College of Cardiology Task Force on Expert Consensus Decision Pathways. J Am Coll Cardiol 2017;70:2421-49. Erratum in: J Am Coll Cardiol 2018;71:587. [Crossref] [PubMed]
15. Kufel J, Bargieł-Łączek K, Kocot S, Koźlik M, Bartnikowska W, Janik M, Czogalik Ł, Dudek P, Magiera M, Lis A, Paszkiewicz I, Nawrat Z, Cebula M, Gruszczyńska K. What Is Machine Learning, Artificial Neural Networks and Deep Learning?—Examples of Practical Applications in Medicine. Diagnostics 2023;13:2582. [Crossref] [PubMed]
16. Abada R, Abubakar AM, Bilal MT. An Overview on Deep Leaning Application of Big Data. Mesopotamian Journal of Big Data 2022;2022:31-5.
17. Fourcade A, Khonsari RH. Deep learning in medical image analysis: A third eye for doctors. J Stomatol Oral Maxillofac Surg 2019;120:279-88. [Crossref] [PubMed]
18. Kufel J, Czogalik Ł, Bielówka M, Magiera M, Mitręga A, Dudek P, Bargieł-Łączek K, Stencel M, Bartnikowska W, Mielcarska S, Modlińska S, Nawrat Z, Cebula M, Gruszczyńska K. Measurement of Cardiothoracic Ratio on Chest X-rays Using Artificial Intelligence-A Systematic Review and Meta-Analysis. J Clin Med 2024;13:4659. [Crossref] [PubMed]
19. Kufel J, Paszkiewicz I, Kocot S, Lis A, Dudek P, Czogalik Ł, Janik M, Bargieł-Łączek K, Bartnikowska W, Koźlik M, Cebula M, Gruszczyńska K, Nawrat Z. Deep Learning in Cardiothoracic Ratio Calculation and Cardiomegaly Detection. J Clin Med 2024;13:4180. [Crossref] [PubMed]
20. Lin A, Manral N, McElhinney P, Killekar A, Matsumoto H, Kwiecinski J, et al. Deep learning-enabled coronary CT angiography for plaque and stenosis quantification and cardiac risk prediction: an international multicentre study. Lancet Digit Health 2022;4:e256-65. [Crossref] [PubMed]
21. Zreik M, Lessmann N, van Hamersvelt RW, Wolterink JM, Voskuil M, Viergever MA, Leiner T, Išgum I. Deep learning analysis of the myocardium in coronary CT angiography for identification of patients with functionally significant coronary artery stenosis. Med Image Anal 2018;44:72-85. [Crossref] [PubMed]
22. Zhang N, Yang G, Gao Z, Xu C, Zhang Y, Shi R, Keegan J, Xu L, Zhang H, Fan Z, Firmin D. Deep Learning for Diagnosis of Chronic Myocardial Infarction on Nonenhanced Cardiac Cine MRI. Radiology 2019;291:606-17. [Crossref] [PubMed]
23. Gao X, Li W, Loomes M, Wang L. A fused deep learning architecture for viewpoint classification of echocardiography. Information Fusion 2017;36:103-13.
24. Khamis H, Zurakhov G, Azar V, Raz A, Friedman Z, Adam D. Automatic apical view classification of echocardiograms using a discriminative learning dictionary. Med Image Anal 2017;36:15-21. [Crossref] [PubMed]
25. Madani A, Arnaout R, Mofrad M, Arnaout R. Fast and accurate view classification of echocardiograms using deep learning. NPJ Digit Med 2018;1:6. [Crossref] [PubMed]
26. Ge R, Yang G, Chen Y, Luo L, Feng C, Ma H, Ren J, Li S. K-Net: Integrate Left Ventricle Segmentation and Direct Quantification of Paired Echo Sequence. IEEE Trans Med Imaging 2020;39:1690-702. [Crossref] [PubMed]
27. Kim T, Hedayat M, Vaitkus VV, Belohlavek M, Krishnamurthy V, Borazjani I. Automatic segmentation of the left ventricle in echocardiographic images using convolutional neural networks. Quant Imaging Med Surg 2021;11:1763-81. [Crossref] [PubMed]
28. Wang Y, Zhang Y, Wen Z, Tian B, Kao E, Liu X, Xuan W, Ordovas K, Saloner D, Liu J. Deep learning based fully automatic segmentation of the left ventricular endocardium and epicardium from cardiac cine MRI. Quant Imaging Med Surg 2021;11:1600-12. [Crossref] [PubMed]
29. Ali Y, Janabi-Sharifi F, Beheshti S. Echocardiographic image segmentation using deep Res-U network. Biomedical Signal Processing and Control 2021;64:102248.
30. Aly AH, Khandelwal P, Aly AH, Kawashima T, Mori K, Saito Y, Hung J, Gorman JH 3rd, Pouch AM, Gorman RC, Yushkevich PA. Fully Automated 3D Segmentation and Diffeomorphic Medial Modeling of the Left Ventricle Mitral Valve Complex in Ischemic Mitral Regurgitation. Med Image Anal 2022;80:102513. [Crossref] [PubMed]
31. Awasthi N, Vermeer L, Fixsen LS, Lopata RGP, Pluim JPW. LVNet: Lightweight Model for Left Ventricle Segmentation for Short Axis Views in Echocardiographic Imaging. IEEE Trans Ultrason Ferroelectr Freq Control 2022;69:2115-28. [Crossref] [PubMed]
32. Attia IZ, Tseng AS, Benavente ED, Medina-Inojosa JR, Clark TG, Malyutina S, Kapa S, Schirmer H, Kudryavtsev AV, Noseworthy PA, Carter RE, Ryabikov A, Perel P, Friedman PA, Leon DA, Lopez-Jimenez F. External validation of a deep learning electrocardiogram algorithm to detect ventricular dysfunction. Int J Cardiol 2021;329:130-5. [Crossref] [PubMed]
33. Cai Q, Lin M, Zhang M, Qin Y, Meng Y, Wang J, Leng C, Zhu W, Li J, You J, Lu X. Automated echocardiographic diastolic function grading: A hybrid multi-task deep learning and machine learning approach. Int J Cardiol 2024;416:132504. [Crossref] [PubMed]
34. Smistad E, Ostvik A, Salte IM, Melichova D, Nguyen TM, Haugaa K, Brunvand H, Edvardsen T, Leclerc S, Bernard O, Grenne B, Lovstakken L. Real-Time Automatic Ejection Fraction and Foreshortening Detection Using Deep Learning. IEEE Trans Ultrason Ferroelectr Freq Control 2020;67:2595-604. [Crossref] [PubMed]
35. Li X, Liao L, Wu K, Meng AT, Jiang Y, Zhu Y, Cui C, Xu X, Shi B, Fei H. An automatic and real-time echocardiography quality scoring system based on deep learning to improve reproducible assessment of left ventricular ejection fraction. Quant Imaging Med Surg 2025;15:770-85. [Crossref] [PubMed]
36. Muraki R, Teramoto A, Sugimoto K, Sugimoto K, Yamada A, Watanabe E. Automated detection scheme for acute myocardial infarction using convolutional neural network and long short-term memory. PLoS One 2022;17:e0264002. [Crossref] [PubMed]
37. Kusunose K, Abe T, Haga A, Fukuda D, Yamada H, Harada M, Sata M. A Deep Learning Approach for Assessment of Regional Wall Motion Abnormality From Echocardiographic Images. JACC Cardiovasc Imaging 2020;13:374-81. [Crossref] [PubMed]
38. Vafaeezadeh M, Behnam H, Hosseinsabet A, Gifani P. CarpNet: Transformer for mitral valve disease classification in echocardiographic videos. Int J Imaging Syst Tech 2023;33:1505-14.
39. Moghaddasi H, Nourian S. Automatic assessment of mitral regurgitation severity based on extensive textural features on 2D echocardiography videos. Comput Biol Med 2016;73:47-55. [Crossref] [PubMed]
40. Edwards LA, Feng F, Iqbal M, Fu Y, Sanyahumbi A, Hao S, McElhinney DB, Ling XB, Sable C, Luo J. Machine Learning for Pediatric Echocardiographic Mitral Regurgitation Detection. J Am Soc Echocardiogr 2023;36:96-104.e4. [Crossref] [PubMed]
41. Wu J, Ge Z, Huang H, Wang H, Li N, Hu C, Pan C, Wu X. Attention-guided model for mitral regurgitation analysis based on multi-task learning. Biomedical Signal Processing and Control 2025;101:107169.
42. Huang H, Ge Z, Wang H, Wu J, Hu C, Li N, Wu X, Pan C. Segmentation of Echocardiography Based on Deep Learning Model. Electronics. 2022;11:1714.
43. Vafaeezadeh M, Behnam H, Hosseinsabet A, Gifani P. Automatic morphological classification of mitral valve diseases in echocardiographic images based on explainable deep learning methods. Int J Comput Assist Radiol Surg 2022;17:413-25. [Crossref] [PubMed]
44. Wu Z, Ge Z, Ge Z, Xing Y, Zhao W, Dong L, Wang Y, Kong D, Hu C, Liang Y, Chen H, Xue W, Pan C, Ni D, Shu X. Feasibility validation of automatic diagnosis of mitral valve prolapse from multi-view echocardiographic sequences based on deep neural network. Eur Heart J Imaging Methods Pract 2024;2:qyae086. [Crossref] [PubMed]
45. Liu Z, Ning J, Cao Y, Wei Y, Zhang Z, Lin S, Hu H. Video Swin Transformer. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); New Orleans, LA, USA; 18-24 June 2022. IEEE; 2022.
46. Bertasius G, Wang H, Torresani L. Is Space-Time Attention All You Need for Video Understanding? Proceedings of the 38th International Conference on Machine Learning, PMLR. 2021;139:813-24.
47. Li K, Wang Y, Gao P, Song G, Liu Y, Li H, et al. UniFormer: Unified Transformer for Efficient Spatiotemporal Representation Learning. Proceedings of the 38th International Conference on Machine Learning, PMLR. 2021.
48. Carreira J, Zisserman A. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE: 2017:6299-308.
49. Liu Z, Mao H, Wu CY, Feichtenhofer C, Darrell T, Xie S. A ConvNet for the 2020s. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); New Orleans, LA, USA; 18-24 June 2022. IEEE; 2022.