Congenital heart disease (CHD) is the most common congenital anomaly, affecting approximately 0.8% of the general population [1]. Despite significant advancements in diagnosis and treatment, CHD remains a common cause of infant mortality in the first year of life, accounting for approximately 30% of deaths due to congenital malformations [2, 3]. Early diagnosis and timely intervention have been shown to improve postnatal outcomes [4, 5]. Therefore, distinguishing normal fetal hearts from CHD is important. Fetal heart assessment is the primary method for detecting fetal CHD [6]. However, ultrasonographers still face significant challenges in obtaining high-standard, high-quality fetal heart images as required by guidelines due to ultrasound imaging artifacts, speckle noise, changes in fetal position and scanning angle, blurred image boundaries, and variations in the quality of images [7]. Although routine mid-pregnancy fetal heart ultrasound screening using the five heart views recommended by guidelines can detect approximately 90% of complex CHD [8, 9], in practice, the detection rate of CHD is only 30%–50%, with a sensitivity of approximately 40%–50% [10, 11], and largely depends on the personal experience of the ultrasonographer [12, 13]. Additionally, there are significant differences in the identification of normal fetal hearts and detection of abnormalities in health care systems amongst various regions [8, 14].

In recent years, a computer-aided approach that assists fetal operators in automatically identifying and interpreting the anatomical structures of the fetal heart has gained significant interest to address these challenges [15, 16]. Artificial intelligence (AI) methods, exemplified by deep learning (DL) [17], have found extensive applications in the field of medical image analysis, including tasks such as image classification, recognition, segmentation, registration, and computer-aided diagnosis. There is a computer-based method for fetal echocardiography. DL has found its most significant application in fetal ultrasound for pre-diagnostic purposes, such as detecting standard planes [18], classifying and identifying CHD [19, 20, 21], and evaluating the development of the fetal heart [22]. Studies have used AI in fetal echocardiography to improve the diagnostic accuracy of fetal CHD [23] and have shown that some AI models equal the performance level of experts [24, 25]. Recently, Arnaout et al. [26] retrospectively collected 107,823 ultrasound images from 1326 mid-pregnancy fetal screening cardiac videos and trained a neural network ensemble using the five views recommended in guidelines, including the three-vessel trachea view, three-vessel view, left ventricular outflow tract view, right ventricular outflow tract view, and abdominal view, to identify normal fetal hearts and CHD. The findings demonstrated that on an internal test dataset of 4108 fetal examinations, the model exhibited an area under the curve of 0.99 in distinguishing normal heart conditions from CHD. Its sensitivity reached 95%, while the specificity was 96%. Notably, the negative predictive value attained was 100%, accurately identifying cases without CHD. Gong et al. [20] proposed a new DGACNN model for the identification of fetal CHD by training 2655 normal hearts and 541 CHD cases. The model demonstrated good performance in identifying CHD, with an accuracy of 84%, surpassing cardiac experts. Additionally, the model addressed the problem of insufficient training datasets to train a stable model. Deep learning, the most advanced type of machine learning, has been applied to adult echocardiography and has been shown to outperform clinicians in judging images when the images are too small or have poor resolution.

Currently, there are several clinical studies on AI for the diagnosis of CHD. However, these studies are mainly single-center studies with small sample sizes and are limited to clinical research, with no available relevant evidence-based medicine. Therefore, this study aims to retrieve published relevant literature through systematic review and meta-analysis methods, scientifically synthesize and analyze the data, and obtain comprehensive and reliable evidence-based medicine evidence to comprehensively evaluate the accuracy of AI in diagnosing CHD.

With the development of AI, various network models have been applied to CHD image recognition in the field of deep learning, among which convolutional neural networks (CNN) [33] are representative of deep learning structure models. CNNs can autonomously extract features from raw image data and learn complex feature information, demonstrating good performance in medical image recognition and classification [26]. The models included in this study were all deep learning models, and except for one study that did not specify the model, the rest were all CNN architectures. In recent years, Xu et al. [34] developed a DW-NET cascaded convolutional neural network for segmenting fetal echocardiographic four-chamber view images, which can correctly locate different structures and accurately delineate the boundaries of anatomical structures, effectively extracting image indicators to assist in early prenatal diagnosis. The study tested the DW-NET on a dataset of 895 fetal four-chamber view echocardiographic images and showed that it had better segmentation performance compared to other mainstream image segmentation methods. Nurmaini et al. [30] compared four CNN architectures, DenseNet121, DenseNet201, ResNet50, and ResNet101, and selected the optimal CNN architecture as DenseNet201. The DenseNet 201 architecture classified seven types of CHD, including ventricular septal defect, atrial septal defect, atrioventricular septal defect, Ebstein’s anomaly, tetralogy of Fallot, transposition of the great arteries, and hypoplastic left heart syndrome, along with normal controls. The model achieved a sensitivity, specificity, and accuracy of 100%, 100%, and 100% within patients, and a sensitivity, specificity, and accuracy of 99%, 97%, and 98% between patients, respectively.

This study included a total of 7 studies and extracted 19 sets of data, with a pooled sensitivity of 0.93 and specificity of 0.93, indicating that AI models have a low rate of misdiagnosis and missed diagnosis in diagnosing CHD. The area under the SROC curve was 0.98, suggesting high accuracy of AI models in diagnosing CHD. In this study, the positive likelihood ratio was 13.0, indicating that when an AI model diagnoses CHD, the probability of diagnosing CHD is high. The negative likelihood ratio was 0.08, indicating that when an AI model diagnoses a normal heart, the probability of CHD is low. The results of this study are consistent with the study by Taksøe-Vester et al. [32], who used logistic regression and backward feature selection in subjects diagnosed with CoA after birth (n = 73) and healthy controls (n = 7300). The AUC of the ROC curve generated by the predictive model was 0.96, with a specificity of 88.9% and sensitivity of 90.4%.

In this study, subgroup analyses were conducted based on region and model. The subgroup analysis by Asia, Europe, and the United States showed that AI models had high diagnostic performance for CHD in different populations, with no significant differences. Most deep learning-based methods are implemented on pioneering backbone networks, the two most notable are ResNet and DenseNet, as these two architectures have a simple design strategy and good performance [35]. Layers in conventional CNN architectures are progressively linked. In contrast, the ResNet architecture employs shortcut connections, bypassing a minimum of two layers. Conversely, the DenseNet architecture offers connections originating from all feature maps in the preceding layer. This implies that all feature maps are propagated to the following layers and linked to the newly produced feature maps [30, 35]. Therefore, this study selected ResNet, DenseNet, and other models for subgroup analysis. The results showed that the area under the ROC curve for ResNet and DenseNet models in diagnosing CHD was 0.98, while the area under the ROC curve for other models was 0.92, indicating that the diagnostic performance of ResNet and DenseNet models was superior to other models.

AI application can overcome the problem of operator experience, and improve physician work flow in the analysis of fetal echocardiographic images. AI has been proven to be more reproducible and consistent than human performance [31, 36]. However, the application of AI in fetal echocardiography is still in its infancy. The database format in articles uses grayscale images or selected frames from standard videos in.avi or.mov format. Each frame needs to be labeled and unified in pixels to create a standardized data set. Currently, deep learning models used for echocardiographic diagnosis are only used to predict two-dimensional plane images. However, the information from two-dimensional planes is limited and cannot fully display the lesions. Deep learning models trained on three-dimensional ultrasound data, with ultrasound dynamic videos or spatiotemporal volumetric data of multi-view lesions, can potentially improve the diagnostic accuracy of the model while fully displaying the lesions. In addition, developing deep learning models based on multimodal ultrasound, including two-dimensional grayscale ultrasound, Doppler ultrasound, and contrast-enhanced ultrasound, can provide complementary ultrasound information and also improve the diagnostic accuracy of deep learning models [37]. Supervised learning of fetal echocardiography is most widely used in the ultrasound field, and the model training process requires big data. However, unlike photography, electrocardiogram, or chest X-ray [38, 39, 40], each ultrasound examination includes thousands of image frames. Therefore, designing a model that can handle a large number of non-independent images from datasets with relatively few individuals is an important challenge to overcome. Unsupervised learning uses unlabeled data and is used as an exploratory method. When faced with the classification of complex small-sample ultrasound data, it significantly improves the accuracy of image recognition and may become a new focus for AI application in fetal echocardiography.

This study is the latest and the first meta-analysis to investigate the value of AI-assisted diagnosis of CHD, further confirming that AI has high accuracy in diagnosing CHD, consistent with previous studies. This provides the latest evidence-based data for the clinical application of AI-assisted diagnosis for CHD. The subgroup analysis showed that the diagnostic performance of ResNet and DenseNet models was superior to other models.

This study has several limitations. First, we only retrieved literature from the English language, and important data published in non-English languages may have been missed. Due to data limitations, we were unable to conduct subgroup analyses for each model and only selected two currently more focused models and other models for subgroup analysis. We were also unable to assess their performance with larger sample sizes and greater confidence. The studies included in this study were all retrospective studies, as few prospective studies have incorporated the diagnostic accuracy of AI into clinical work. Finally, the included studies only analyzed a single congenital heart disease, or only distinguished normal heart disease from normal heart disease, and few studies distinguished fetal congenital heart disease from normal heart disease and classified specific congenital heart disease. Going forward, we plan to expand the classification of AI models to cover the entire CHD spectrum. Additional studies are needed to examine how the results of the current study apply to different fetal conditions.

AI has significant diagnostic value in assisting the diagnosis of CHD. ResNet and DenseNet models having the best diagnostic efficacy. In view of the small number of studies included in this meta-analysis, and the fact that they were retrospective, further multicenter studies with larger sample sizes are needed to confirm the diagnostic efficacy of AI-assisted diagnosis of CHD. With advances in AI technology, the diagnosis of fetal diseases will become increasingly accurate. We look forward to testing and improving integrated learning models in larger populations, making the expertise of fetal cardiologists available to patients in all regions, obtaining homogeneous medical resources, and applying similar techniques to other diagnoses in medical imaging.

The data used to support the findings of this study are included in the article.

The authors confirm their contribution to the paper: study conception and design: JL, YG; data collection: YG; analysis and interpretation of results: JL, YG, LY; draft manuscript: JL, YG, LY. All authors contributed to editorial changes in the manuscript. All authors reviewed the results and approved the final version of the manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

Not applicable.

This research received no external funding.

The authors declare no conflict of interest.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/RCM28060.