Artificial intelligence based automatic classification, annotation, and measurement of the fetal heart using HeartAssist

Artificial intelligence based automatic classification, annotation, and measurement of the fetal heart using HeartAssist | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Artificial intelligence based automatic classification, annotation, and measurement of the fetal heart using HeartAssist Rina Kim, Mi-Young Lee, Yoo Jin Lee, Hye-Sung Won, Jinki Park, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5718511/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 16 Apr, 2025 Read the published version in Scientific Reports → Version 1 posted 13 You are reading this latest preprint version Abstract Objectives: This study evaluated the feasibility of HeartAssist, a novel automated tool designed for classification of fetal cardiac views, annotation of cardiac structures, and measurement of cardiac parameters. Unlike previous AI tools that primarily focused on classification, HeartAssist integrates classification, annotation and measurement capabilities, enabling a more comprehensive fetal cardiac assessment. Methods: Cardiac images from fetuses (gestational ages 20–40 weeks) were collected at Asan Medical Center between January 2016 and October 2018. HeartAssist was developed using convolutional neural networks to classify 10 cardiac views, annotate 26 structures, and measure 43 parameters. One expert performed manual classifications, annotations, and measurements, which were then compared to HeartAssist outputs to assess feasibility. Results: A total of 65,324 images from 2,985 fetuses were analyzed. HeartAssist achieved 99.4% classification accuracy, with recall, precision, and F1-score of 0.93, 0.95, and 0.94, respectively. Annotation accuracy was 98.4%, while the automatic measurement success rate was 97.6%, with an error rate of 7.62% and caliper similarity of 0.613. Conclusions: HeartAssist is a reliable tool for fetal cardiac screening, demonstrating high accuracy in classifying cardiac views and annotating structures, with comparable outcomes in measuring cardiac parameters. This tool could enhance prenatal detection of congenital heart disease and improve perinatal outcomes. Health sciences/Cardiology/Cardiovascular biology/Heart development Health sciences/Health care/Medical imaging/Ultrasonography/Echocardiography artificial intelligence classification congenital heart disease fetal cardiac measurement fetal echocardiography image interpretation computer-assisted Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Congenital heart disease (CHD) is the most common congenital anomaly and a leading cause of infantile morbidity and mortality worldwide [1]. An accurate prenatal diagnosis of CHD can improve neonatal outcomes and enhance parental counseling. Fetal echocardiography is the main method for diagnosing CHD, particularly in women with increased risks for fetal cardiac anomalies [2]. Despite several guidelines recommending essential scanning views and cardiac parameters for fetal echocardiography [3-6], the prenatal detection rate of CHD remains variable [7, 8]. The following factors contribute to this challenge: 1) fetal heart is small with indistinct anatomical appearance; 2) fetus moves with relative movement of the probe; 3) fetal heartbeats are fast; 4) contrast artifacts make imaging difficult; and 5) some examiners lack expertise in fetal echocardiography. These factors may result in extended time in the acquisition of clear images and missed prenatal diagnosis. Deep learning has become increasingly popular in the field of fetal imaging [9]. With advances in artificial intelligence (AI) technology, automated techniques employing convolutional neural networks (CNNs) have been introduced in fetal echocardiography [10, 11]. However, many studies on AI applications in fetal echocardiography have predominantly focused on view classification, with limited data available on their capability to measure cardiac structures [12-16]. In addition, there is a study on the automatic classification of eight fetal cardiac views using an AI system [16]. While this study encompasses the largest number of cardiac views reported to date, if multiple guidelines are integrated, more than 10 cardiac views could potentially be confirmed by fetal echocardiography [1, 5, 6, 17]. To address these limitations, we collaborated with Samsung Medison to develop an automated tool designed for classifying fetal cardiac views, annotating cardiac structures, and measuring cardiac parameters. This study aimed to evaluate the feasibility of this new system, HeartAssist (Samsung Medison Co., Ltd, Seoul, Korea). Methods 1. Study population This study prospectively and retrospectively analyzed the data of fetuses collected at Asan Medical Center, Seoul, Korea, from January 2016 to October 2018. The inclusion criteria were: 1) fetuses from both singleton and multiple pregnancies confirmed by prenatal ultrasonography; 2) required to undergo detailed ultrasonography, including echocardiography by several specialists in maternal fetal medicine: and 3) between 20 and 40 weeks of gestational age (GA); and 4) regardless of the presence of fetal anomalies. 2. Data acquisition All prenatal ultrasonographic evaluations were conducted using WS80A or HERA W10 (Samsung Medison Co., Ltd, Seoul, Korea) with associated transabdominal probes. This study was approved by the Institutional Review Board of Asan Medical Center (approval number: 2017-0688). Informed consent was obtained from the prospectively recruited subjects and was waived for the retrospectively recruited subjects as the study used pre-collected, anonymized data without direct patient interaction by the Institutional Review Board of Asan Medical Center. The waiver was granted in accordance with the Institutional Review Board of Asan Medical Center, the Declaration of Helsinki, and ethical guidelines. 2.1. 10 cardiac views and 43 parameters The recording data included 10 classified views; transverse abdominal view, four-chamber view (4CV), left ventricular outflow tract view (LVOTV), right ventricular outflow tract view (RVOTV), three-vessel view (3VV), three-vessel trachea view (3VTV), three-vessel view with pulmonary artery branches (3VV PA), aortic arch view (AAV), ductal arch view, and bicaval view. Among the 10 views, seven views (4CV, LVOTV, RVOTV, 3VV, 3VTV, 3VV PA, and AAV) were used to measure 43 parameters that were classified into two categories (Table 1). The measurement data (35) included cardiac axis, thoracic area, cardiac area, thoracic circumference, cardiac circumference, thoracic transverse diameter, cardiac transverse diameter, left atrium (LA) width, right atrium (RA) width, LA area, RA area, left ventricle (LV) width, right ventricle (RV) width, LV length, RV length, LV area, RV area, mitral valve (MV) annulus diameter, and tricuspid valve (TV) annulus diameter on 4CV; aortic valve (AV) annulus diameter and ascending aorta diameter on LVOTV; pulmonary valve (PV) annulus diameter on RVOTV; PA diameter, aorta diameter, superior vena cava (SVC) diameter, and thymus diameter on 3VV; aortic isthmus diameter and ductus arteriosus (DA) diameter on 3VTV; main pulmonary artery (MPA) diameter and right pulmonary artery (RPA) diameter on 3VV PA; and ascending aorta diameter, proximal transverse aortic arch diameter, distal transverse aortic arch diameter, aortic isthmus diameter, and descending aorta diameter on AAV. The calculation data (8) included cardiothoracic area ratio (CTAR), cardiothoracic circumference ratio (CTCR), cardiothoracic ratio (CTR), left to right ventricular width ratio [LV/RV (W)], left to right ventricular length ratio [LV/RV (L)], left to right ventricular area ratio [LV/RV (A)], aortic valve to aorta ratio (AV/Aorta), aortic to pulmonary valve ratio (AV/PV). 2.2. The measurement methods of the cardiac parameters The cardiac parameters were measured based on previously reported studies [4, 5, 18-30]. All cardiac parameters were measured from the inner to inner edge. 2.2.1. Parameters in 4CV The following criteria were used to measure the parameters in 4CV. ① The cardiac axis refers to the angle formed by the line dividing the thorax in half and the extension line of the interventricular septum of the heart. ② The area and circumference of the heart were measured with maximal distention of the heart at end-diastole (just after the atrioventricular valves close) along the outer border of the heart. ③ The cardiac transverse diameter was measured at the level of the atrioventricular valves at end-diastole. ④ The thoracic area, thoracic circumference, and thoracic transverse diameter were measured at the outer border from rib to rib at the same aforementioned cardiac cycle. ⑤ The CTAR and CTCR were calculated as the ratio of the area and circumference of the heart and thorax, respectively, and CTR was calculated as the ratio of the transverse diameter of the heart and thorax. ⑥ The atrial diameters and areas were measured when both atria showed maximal expansion at end-systole (just before the atrioventricular valves opened). ⑦ The LA and RA widths were measured in the largest transverse diameter. ⑧ The LA and RA areas were measured, excluding neither the pulmonary veins nor the atrioventricular valves. ⑨ The ventricular diameters and areas were measured when both ventricles showed maximal expansion at end-diastole. ⑩ The LV and RV widths were measured below the MV and TV, respectively. ⑪ The LV and RV lengths were measured from the basal to the apex of the ventricles, including the MV and TV. ⑫ The LV and RV areas were measured including the endocardium and the moderator band inside the RV cavity. ⑬ LV/RV (W), LV/RV (L), and LV/RV (A) were calculated using the respective result values from the LV and RV. ⑭ The TV and MV annulus diameters were measured between the hinge points of the leaflet attachment with maximal expansion when both atrioventricular valves opened at early-diastole. 2.2.2. Parameters in the outflow tract views The following criteria were used to measure the parameters in the outflow tract views. ① The AV and PV annuli diameters were measured perpendicular to the great vessels at the level of the valves at mid-systole (when the AV and PV valves opened) in LVOTV and RVOTV, respectively. ② The ascending aorta diameter was measured at maximal size above the AV annulus. ③ AV/aorta and AV/PV were obtained by calculating the ratio of each value. 2.2.3. Parameters in the 3VV The following criteria were used to measure the parameters in the 3VV. ① The diameters of PA, aorta, and SVC were measured at the widest diameter perpendicular to each vessel. ② The thymus diameter was measured at the largest transverse diameter. 2.2.4. Parameters in the 3VTV The following criteria were used to measure the parameters in the 3VTV. ① The diameters of the aortic isthmus and DA were measured at the largest diameter just before its junction into the descending aorta. ② The MPA and RPA diameters were measured at the widest diameter in 3VV PA. ③ The RPA diameter was measured just after bifurcation of the MPA. 2.2.5. Parameters in AAV The aortic diameters were measured at five levels: ascending, proximal transverse, distal transverse, isthmus, and descending aorta in AAV [25]. 3. Development of HeartAssist 3.1. Data acquisition All the acquired images were stored on the computer in the format of digital imaging and communications in medicine (DICOM). Subsequently, these images were classified into 10 cardiac views and labeled [ground-truth (GT)], regardless of the GA, by one expert in fetal echocardiography (M.Y.L). The segmentation label (GT) was also manually delineated by same expert (M.Y.L). 3.2. Overview of HeartAssist We developed HeartAssist (Samsung Medison Co., Ltd.), which features classification and segmentation models based on the Vision Transformer (ViT) [31] and the CNNs [32]. 3.2.1. ViT for classification The ViT architecture is a model that has been extensively validated in natural language processing and has been successfully extended to various domains by leveraging the strengths of the transformer framework . Additionally, ViT offers scalability , allowing easy adjustment of the number of layers and parameters , which helps balance performance and computational cost . Based on these advantages, we utilized the ViT for image classification . 3.2.2. CNN for segmentation For the segmentation model, the CNN architecture was chosen to better capture fine anatomical structures . Because ViT processes images in discrete patches , it may be less effective in preserving localized spatial information compared to CNNs. Given that fetal cardiac imaging often involves high-resolution images with numerous small structures, CNNs are better suited for this task, as they can more effectively analyze detailed anatomical features . Specifically, HeartAssist includes two main functions, namely, an automatic view recognition module (a classification algorithm model) and an anatomy segmentation module (a segmentation model used for annotation and measurement). Both modules were developed using deep learning algorithms under supervised learning. We illustrated the multi-stage deep learning-based framework for fetal echocardiogram analysis in Figure 1, which consists of frame selection, ViT-based classification, CNN-based segmentation, and measurement. 3.3. Model training and data augmentation To train the ViT-based classification model, we resized the input images to 224 × 224 while keeping the aspect ratio unchanged and performed CutMix [33] and MixUp [34]. CutMix and MixUp are strategies for image data augmentation. They arbitrarily cut out a portion of one image and place it on top of another, and then interpolate the pixel values between the two images. These two methods prevent the model from overfitting the training distribution and increase the likelihood that the model can generalize to examples outside the distribution. Furthermore, they also prevent the model from relying too heavily on specific features to learn. The segmentation model utilized a higher resolution (1,024 × 512) than the classification model because of the precision required for positioning. Supplementary Table S1 shows the hyper-parameters of the classification and segmentation models. The initial values were experimentally determined, with the optimal performance achieved at the values shown in the supplementary Table S1. Supplementary Table S2 presents the performance of our models in relation to their computational and space complexity, measured in terms of the number of FLOPs and parameters. The original dataset, consisting of 65,324 images, was divided into a training set (93.8% of the total images) and a validation set (6.2% of the total images) through random selection. There was no overlap between the training and validation datasets. Subsequently, data augmentation techniques were employed to improve the performance of the learning algorithm. These techniques included pixel brightness normalization, horizontal flipping, rotation, and scaling. The augmentation layer was randomly applied. Additionally, we implemented weight adjustment, assigning greater importance to underrepresented views to improve learning. 3.4. Annotation and measurement process After learning of the segmentation model, annotation and measurement can be performed. The annotation results directly use the center point of the segmented regions through the segmentation network; thus, no additional postprocessing is applied. The list of annotations of cardiac structures is presented in Table 2. In addition, the area and circumference measurement directly use the contour regions of the segmented results. However, the same method cannot be applied to diameter measurements. Thus, postprocessing based on the area where the structure is divided is employed, or the diameter is measured by dividing the position of the caliper. Area-based postprocessing involves finding the diameters of items that are easy to find within the area of the detected structure according to the measurement standards for each item. Furthermore, diameter items that are difficult to measure solely from the area of the structure are measured by dividing the area centered on both endpoints of the caliper. The size of this area is determined within a certain range depending on the caliper length. Each parameter has been applied using a proper technique summarized in Table 3. 4. Statistical analysis To assess the classification model performance, we calculated recall, precision, F1-score, and accuracy based on the true positive, true negative, false positive, and false negative values. Recall is the ability of the classifier to correctly identify all the positive samples. Precision reflects the ability of the classifier not to label a sample that is negative as positive. The F1-score is interpreted as a weighted harmonic mean of recall and precision. Accuracy is a simple ratio of true positive and true negative of the total sample. All four metrics are optimized when they reach a value of one and are least effective at a score of zero. The classification performance of HeartAssist was evaluated using a confusion matrix, illustrating the agreement between results and GT values. The detailed confusion matrix is provided in the supplementary Table S3. To evaluate the accuracy of the annotation, one expert (M.Y.L) visually checked the annotation results of HeartAssist to determine success or failure and calculate the success rate. To quantitatively evaluate the performance of the segmentation algorithm, the Dice similarity coefficient (DSC) [35, 36] and the error rate were used. The DSC is a popular metric for image segmentation, particularly in medical image analysis, as it determines whether the segmentation is correctly performed. The DSC is calculated separately for each anatomical structure rather than for all structures as a whole. In addition, the accuracy of the measurement was measured by comparing the automated measurements with the manual measurements by the expert (M.Y.L). The error on the automated measurement was calculated against the GT, which is defined as the mean of the expert’s measurements. The error rate (%) was calculated as: The resulting similarity value ranges from 0 to 1, with 1 indicating perfect similarity and the lower values indicating greater discrepancies between the measurements. Results We collected 65,324 randomly sampled images of fetal echocardiography from 2,985 fetuses over 34 months. In the classification model, 61,277 (93.8%) images were used for training and 4,047 (6.2%) for validation (Table 4). The performances of the classification model are presented in Table 5. The automatically classified cardiac images matched the expert’s classified answers by 99.4% on average. The average of recall, precision, and F1-score in all cardiac views were 0.931, 0.950, and 0.939, respectively. The classification performance of HeartAssist was evaluated using a confusion matrix, illustrating the agreement between results and GT values. The detailed confusion matrix is provided in the supplementary materials (supplementary Table S4). The datasets utilized for the annotations are presented in Table 6. The accuracy of annotations was excellent across all cardiac views, as shown in Table 7. In particular, the annotations in 4CV, LVOTV, and RVOTV were all correctly displayed (Figure 2A, 2B, and 2C). A total of 37,981 images were used for training and validation in the measurements (Table 8). Table 9 shows the results of the measurement analysis. All cardiac parameters, except for the descending aorta diameter, were automatically measured with a successful rate of over 90%. The mean error rate of all cardiac parameters was 7.624%. The error rates exceeded 10% for the ascending aorta in LVOTV, SVC and thymus diameter in 3VV, aortic isthmus and DA diameter in 3VTV, RPA diameter in 3VV PA, and transverse aortic arch and aortic isthmus diameter in AVV. The mean caliper similarity was 0.613. The cardiac axis in 4CV exhibited the highest similarity (0.825), while the thymus in 3VV exhibited the lowest (0.352). The DSC for area and circumference measurements showed excellent results. Discussion Our study presents several notable advancements. First, while previous AI-based approaches primarily focused on classification, HeartAssist integrates classification, segmentation, and measurement into a single comprehensive system. Second, it enables automated measurement of 43 cardiac parameters, providing a comprehensive assessment of the fetal heart. Finally, the model was trained on a large dataset, contributing to its robustness and potential clinical applicability. In our study, 10 cardiac views were selected and analyzed. The 3VV PA was added to the existing standard cardiac views. When any branch of the PA was visible, it was classified as 3VV PA. This view was included based on the guidelines of the American Institute of Ultrasound in Medicine to evaluate the size and orientation of the PA branches [ 5 ]. Recent advancements in AI have greatly improved fetal echocardiography, with deep learning models achieving higher accuracy in classification and segmentation. Various studies have explored AI applications using CNNs and hybrid architectures such as U-Net, ResNet, Mask R-CNN, and DW-Net for automated analysis, as summarized in supplementary Table S4. However, most prior studies focused on limited datasets, primarily 4CV, with relatively small image counts, limiting their clinical applicability. In contrast, our study integrated ViT and CNN architectures to evaluate 10 cardiac views using a significantly larger dataset of 65,324 images, achieving a classification F1-score of 0.939, annotation accuracy of 98.4%, and segmentation DSC of 0.974. These results indicate that our model achieved better performance compared to previous studies. The RVOT, 3VV, 3VT, and 3VV PA values were slightly lower compared with those of other cardiac views. These four cardiac views can be challenging to classify even for experts in fetal echocardiography. For example, RVOT and 3VV can be classified differently depending on the visibility of the RV and PV. They may be categorized as either RVOT or 3VV. In such cases, if either RVOT or 3VV is considered to be correct, the accuracy would improve. Furthermore, since our dataset primarily consists of retrospectively collected static images, the automatic recognition of RV and PV structures may have been more challenging. Training with prospectively collected video clips could improve classification performance by allowing the model to better identify cardiac structures in motion rather than relying only on static images. Although the success rates for the automated measurement of most cardiac parameters were high, some parameters exhibited lower success rates than the others. For thoracic area and circumference, the automated system failed to perform measurements when the entire thorax was not included in the image. Similarly, the success rates of the measurements in AAV were low. Among the parameters measured by AAV, the success rate for the descending aorta was the lowest. This could be attributed to the fact that most of the failed cases were in the prone position, which caused the descending aorta to not be clearly delineated. Our data exhibited a slightly higher mean error rate of 7.6% than the generally acceptable rate of 5% and a moderate to high level of caliper similarity with a value of 0.613. For parameters with small values (e.g., SVC in 3VV, aortic isthmus and DA diameter in 3VTV, RPA diameter in 3VV PA, and transverse arch and aortic isthmus in AA), even a minimal displacement of the cursor can significantly increase the error rate (Fig. 3 A, and 3 B). Nevertheless, the caliper similarity for these values exceeded 0.6, suggesting that the measurements were closely aligned. In the case of the thymus, the boundaries are often unclear, leading to greater discrepancies between measurements and, consequently, a higher error rate and lower caliper similarity. Nevertheless, upon reviewing the images measured by the automated system, we found that the measurements themselves were performed quite accurately (Fig. 4 A, and 4 B). Contrarily, for the measurement of the cardiac axis, the reference points are clearly defined, which results in a low error rate and a high caliper similarity between the automated and manual measurements. The strengths of this study include the application of a deep-learning process using a diverse set of images from both normal and abnormal fetuses. In addition, to ensure the highest accuracy of GT data, all images were evaluated and measured by a single expert (M.Y.L). However, this study also has several limitations. First, the performance of the automated measurements of cardiac parameters remains insufficient and may require additional manual corrections. We believe that incorporating a larger dataset and validating the measurements with multiple experts may further improve the accuracy of cardiac parameter assessment. Second, as most images were retrospectively obtained, accurate determination of the specific phase of the cardiac cycle captured in each image was challenging. The required phase of the cardiac cycle varies for different measurements, and some measurements may not have been taken at the appropriate cardiac cycle. Third, as this study was based on data from a single center, its applicability to broader populations may be limited. Additionally, the use of ultrasound machines from a single manufacturer may have influenced image quality and model performance. Despite these challenges, our results demonstrated comparable outcomes. At present, we are prospectively collecting video clips of the fetal heart to automatically detect each cardiac cycle and measure cardiac parameters in the relevant cardiac view. Future advancements to HeartAssist will focus on enhancing its ability to perform accurate real-time assessments while improving its capacity to distinguish between correctly and incorrectly captured images. These advancements will not only assist experts in conducting examinations more efficiently but also help beginners perform ultrasound assessments more effectively by following AI-assisted guidelines, thereby reducing disparities in diagnostic accuracy. Additionally, strengthening the tool’s capability to detect CHD will be a crucial step in maximizing its clinical utility. In conclusion, our study confirms that the novel automated program, HeartAssist, is a feasible and effective tool for fetal cardiac screening. With it high accuracy in the classification and annotation of cardiac views and structures, as well as measurement of cardiac parameters, HeartAssist mitigates the limitations of manual examination due to inter-examiner variability. It is particularly valuable for examiners with less expertise in fetal echocardiography. We anticipate that HeartAssist will significantly enhance the detection rate of CHDs, ultimately leading to improved perinatal outcomes and better overall care for fetal cardiac health. Declarations Author’s contributions Rina Kim: Designed and performed experiments, analyzed data, and wrote the paper. Mi-Young Lee: Designed and performed experiments, manually checked all data, analyzed and evaluated data, wrote the paper, and supervised the research. Yoo Jin Lee: Collected and organized data. Hye-Sung Won: Supervised the research and provided final approval of the version to be published. Jinki Park, Jihoon Lee, and Kwangyeon Choi: Developed the HeartAssist program and analyzed the results. Also, evaluated data and performed statistics. Data availability The data used to support the findings of this study are available from the corresponding author upon request. Competing interests The authors declare no competing interests. Ethics declarations The institutional review board of our institution approved this study (approval number: 2017-0688). Informed consent was obtained from the prospectively recruited subjects and was waived from the retrospectively recruited subjects. Funding None Acknowledgement The authors gratefully acknowledge the support of Samsung Medison for their contributions to this works. References Abuhamad, A. and R. Chaoui, A Practical Guide to Fetal Echocardiography: Normal and Abnormal Hearts . 4th ed. 2015, Philadelphia: Wolters Kluwer Health. Lee, W., et al., ISUOG consensus statement: what constitutes a fetal echocardiogram? Ultrasound Obstet Gynecol, 2008. 32 (2): p. 239-42. The International Society of Ultrasound in Obstetrics & Gynecology, Cardiac screening examination of the fetus: guidelines for performing the 'basic' and 'extended basic' cardiac scan. Ultrasound Obstet Gynecol, 2006. 27 (1): p. 107-13. Carvalho, J.S., et al., ISUOG Practice Guidelines (updated): sonographic screening examination of the fetal heart. Ultrasound Obstet Gynecol, 2013. 41 (3): p. 348-59. American Institute of Ultrasound in Medicine, AIUM Practice Parameter for the Performance of Fetal Echocardiography. J Ultrasound Med, 2020. 39 (1): p. E5-E16. Carvalho, J.S., et al., ISUOG Practice Guidelines (updated): fetal cardiac screening. Ultrasound Obstet Gynecol, 2023. 61 (6): p. 788-803. van Velzen, C.L., et al., Prenatal detection of congenital heart disease--results of a national screening programme. BJOG, 2016. 123 (3): p. 400-7. Waern, M., et al., Prenatal detection of congenital heart disease - results of a Swedish screening program 2013-2017. BMC Pregnancy Childbirth, 2021. 21 (1): p. 579. Ramirez Zegarra, R. and T. Ghi, Use of artificial intelligence and deep learning in fetal ultrasound imaging. Ultrasound Obstet Gynecol, 2023. 62 (2): p. 185-194. Shengfeng, L., et al., Deep learning in Medical Ultrasound Analysis: A Review. Engineering, 2019. 5 (2): p. 261-275. Day, T.G., et al., Artificial intelligence, fetal echocardiography, and congenital heart disease. Prenatal Diagnosis, 2021. 41 (6): p. 733-742. Vaanathi, S., et al., Automated characterization of the fetal heart in ultrasound images using fully convolutional neural networks. IEEE, 2017: p. 671-674. Philip, M.E., et al., Convolutional Neural Networks for Automated Fetal Cardiac Assessment using 4D B-Mode Ultrasound. 2019: p. 824-828. Sakai, A., et al., Medical Professional Enhancement Using Explainable Artificial Intelligence in Fetal Cardiac Ultrasound Screening. Biomedicines, 2022. 10 (3): p. 551. Wu, H., et al., Application of Artificial Intelligence in Anatomical Structure Recognition of Standard Section of Fetal Heart. Comput Math Methods Med, 2023. 2023 : p. 13. Taksoee-Vester, C.A., et al., AI supported fetal echocardiography with quality assessment. Sci Rep, 2024. 14 (1): p. 5809. Moon-Grady, A.J., et al., Guidelines and Recommendations for Performance of the Fetal Echocardiogram: An Update from the American Society of Echocardiography. J Am Soc Echocardiogr, 2023. 36 (7): p. 679-723. Paladini, D., S.K. Chita, and L.D. Allan, Prenatal measurement of cardiothoracic ratio in evaluation of heart disease. Arch Dis Child, 1990. 65 (1): p. 20-23. Sharland, G.K. and L.D. Allan, Normal fetal cardiac measurements derived by cross-sectional echocardiography. Ultrasound Obstet Gynecol, 1992. 2 (3): p. 175-81. Ramada S., S., et al., Ultrasonographic left cardiac axis deviation: a marker for fetal anomalies. Obstet Gynecol, 1995. 85 (2): p. 187-191. Shapiro, I., et al., Fetal cardiac measurements derived by transvaginal and transabdominal cross-sectional echocardiography from 14 weeks of gestation to term. Ultrasound Obstet Gynecol, 1998. 12 (6): p. 404-418. Yagel, S., et al., The three vessels and trachea view (3VT) in fetal cardiac scanning. Ultrasound Obstet Gynecol, 2002. 20 (4): p. 340-5. Theera, T. and T. Teerapong, Cardiothoracic ratio in the first half of pregnancy. J Clin Ultrasound, 2004. 32 (4): p. 186-9. Schneider, C., et al., Development of Z-scores for fetal cardiac dimensions from echocardiography. Ultrasound Obstet Gynecol, 2005. 26 (6): p. 599-605. Pasquini, L., et al., Z-scores of the fetal aortic isthmus and duct: an aid to assessing arch hypoplasia. Ultrasound Obstet Gynecol, 2007. 29 (6): p. 628-33. Cho, J.Y., et al., Diameter of the normal fetal thymus on ultrasound. Ultrasound Obstet Gynecol, 2007. 29 (6): p. 634-8. Lopez, L., et al., Recommendations for quantification methods during the performance of a pediatric echocardiogram: a report from the Pediatric Measurements Writing Group of the American Society of Echocardiography Pediatric and Congenital Heart Disease Council. J Am Soc Echocardiogr, 2010. 23 (5): p. 465-95; quiz 576-7. Li, X., et al., Z-score reference ranges for normal fetal heart sizes throughout pregnancy derived from fetal echocardiography. Prenat Diagn, 2015. 35 (2): p. 117-24. García-Otero, L., et al., Reference ranges for fetal cardiac, ventricular and atrial relative size, sphericity, ventricular dominance, wall asymmetry and relative wall thickness from 18 to 41 gestational weeks. Ultrasound Obstet Gynecol, 2021. 58 (3): p. 388-397. Lang, R.M., et al., Recommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging. J Am Soc Echocardiogr, 2015. 28 (1): p. 1-39 e14. Han, K., et al., A Survey on Vision Transformer. IEEE Trans Pattern Anal Mach Intell, 2023. 45 (1): p. 87-110. LeCun, Y., et al., Backpropagation Applied to Handwritten Zip Code Recognition. Neural Computation, 1989(1): p. 541-551. Sangdoo, Y., et al., CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features. arXiv, 2019. Hongyi, Z., et al., mixup: Beyond Empirical Risk Minimization. 2018. Dice, L.R., Measures of the Amount of Ecologic Association Between Species. Ecology, 1945. 26 (3): p. 297-302. Minaee, S., et al., Image Segmentation Using Deep Learning: A Survey. IEEE Trans Pattern Anal Mach Intell, 2022. 44 (7): p. 3523-3542. Tables Tables 1 to 9 are available in the Supplementary Files section Additional Declarations No competing interests reported. Supplementary Files Tables.docx SupplementaryFINAL.pdf Cite Share Download PDF Status: Published Journal Publication published 16 Apr, 2025 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Accepted 08 Apr, 2025 Reviewers agreed at journal 08 Apr, 2025 Reviews received at journal 05 Apr, 2025 Reviews received at journal 05 Apr, 2025 Reviewers agreed at journal 05 Apr, 2025 Reviews received at journal 03 Apr, 2025 Reviewers agreed at journal 03 Apr, 2025 Reviews received at journal 03 Apr, 2025 Reviewers agreed at journal 03 Apr, 2025 Reviewers agreed at journal 02 Apr, 2025 Reviewers invited by journal 02 Apr, 2025 Submission checks completed at journal 01 Apr, 2025 First submitted to journal 31 Mar, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5718511","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":437652707,"identity":"7a58a5c4-1a5d-4c76-8966-239d9411fe3a","order_by":0,"name":"Rina Kim","email":"","orcid":"","institution":"Jeju National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Rina","middleName":"","lastName":"Kim","suffix":""},{"id":437652708,"identity":"48bd60cc-6468-4714-93b0-dd959b1cf169","order_by":1,"name":"Mi-Young Lee","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA60lEQVRIiWNgGAWjYNACAwYDPgbmY2A2Gzth9YwNIC1sDGxpEC3MIDKBkBYGkBYeMwifkBZz9t7jj3kK7hizsfd8e/BzxzZ5PmYGtgcff+DWYtlzLrGZx+CZGRvP2e2GvWduG7YxM7AbzsBji8GNHMPmHIPDNmwSudskeNtuMwK1sEnzEKVF/s0zyb9tt+3BWv4QocWMTYKHTRpoSyJYCz7vg/wy+4/BYWM2njQzadm228ltzIxtkj1puLUAQ+zAxxl/Dhv2sx9+Jvm27bbt/PbmYxI/bPA4jIEHQwwcUaRpGQWjYBSMglGACgCnOkrvZJf4PQAAAABJRU5ErkJggg==","orcid":"","institution":"University of Ulsan College of Medicine, Asan Medical Center","correspondingAuthor":true,"prefix":"","firstName":"Mi-Young","middleName":"","lastName":"Lee","suffix":""},{"id":437652709,"identity":"22213342-2358-4ab1-9a8b-1d7a67bfb8f9","order_by":2,"name":"Yoo Jin Lee","email":"","orcid":"","institution":"University of Ulsan College of Medicine, Asan Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Yoo","middleName":"Jin","lastName":"Lee","suffix":""},{"id":437652710,"identity":"31433319-b84e-44a9-adce-e9859e0cf80d","order_by":3,"name":"Hye-Sung Won","email":"","orcid":"","institution":"University of Ulsan College of Medicine, Asan Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Hye-Sung","middleName":"","lastName":"Won","suffix":""},{"id":437652711,"identity":"2340626b-381e-4565-9d20-f6b8e5419d93","order_by":4,"name":"Jinki Park","email":"","orcid":"","institution":"Samsung Medison Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Jinki","middleName":"","lastName":"Park","suffix":""},{"id":437652712,"identity":"96337720-caaf-492e-bfd7-8c95b006e1dc","order_by":5,"name":"Jihoon Lee","email":"","orcid":"","institution":"Samsung Medison Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Jihoon","middleName":"","lastName":"Lee","suffix":""},{"id":437652713,"identity":"83e78d72-ef87-45c0-81a1-f77d80e1d16e","order_by":6,"name":"Kwangyeon Choi","email":"","orcid":"","institution":"Samsung Medison Co., Ltd","correspondingAuthor":false,"prefix":"","firstName":"Kwangyeon","middleName":"","lastName":"Choi","suffix":""}],"badges":[],"createdAt":"2024-12-27 02:23:10","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5718511/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5718511/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-025-97934-z","type":"published","date":"2025-04-16T15:56:58+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":79905999,"identity":"26edab24-3b6d-47ba-ab05-bf6202df2dc8","added_by":"auto","created_at":"2025-04-04 11:01:23","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":2194201,"visible":true,"origin":"","legend":"\u003cp\u003eDeep learning-based framework for fetal cardiac imaging: Classification, Segmentation, and Measurement. 3VV, three-vessel view; 4CV, four-chamber view; CNN, convolutional neural network; MLP, multi-layer perceptron; TAV, transverse abdominal view; ViT, Vision Transformer\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5718511/v1/0fcf527149f2fe1c4730e1bb.png"},{"id":79907021,"identity":"5cc63bc6-b7af-4f2f-bb58-505cc7eb1200","added_by":"auto","created_at":"2025-04-04 11:09:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":2178136,"visible":true,"origin":"","legend":"\u003cp\u003eResults of automatic classification and annotation\u003c/p\u003e\n\u003cp\u003eThe automatic labels are displayed centrally within the structures. A. Four-chamber view (4CV); B. Left ventricular outflow tract view (LVOTV); C. Right ventricular outflow tract view (RVOTV). All 2D gray-scale images were acquired transabdominally. LA, left atrium; RA, right atrium; LV, left ventricle; RV, right ventricle; Asc. Ao, ascending aorta.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5718511/v1/f91a70e92853b2322413d06a.png"},{"id":79906007,"identity":"49b9b4f5-b81c-41d2-b01e-a983318eb2a0","added_by":"auto","created_at":"2025-04-04 11:01:24","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3981948,"visible":true,"origin":"","legend":"\u003cp\u003eEffect of cursor displacement on error rate for small parameters.\u003c/p\u003e\n\u003cp\u003eComparison between automated measurement by HeartAssist (A) and manual measurement (B) on the aortic arch in the aortic arch view, acquired using 2D ultrasound transabdominally. Both the automated and manual measurements appear reasonably accurate; however, differences were observed in the measurements of the proximal transverse aortic arch and distal transverse aortic arch, with error rates of 13.0% and 20.3%, respectively. In addition, a discrepancy exists in the measured positions between the automated and manual measurements for the descending aorta(arrow).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5718511/v1/ded4f50990f9254ca1401edd.png"},{"id":79906010,"identity":"dac34c3f-5826-4834-ab56-8db9de04eb42","added_by":"auto","created_at":"2025-04-04 11:01:24","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":2311206,"visible":true,"origin":"","legend":"\u003cp\u003eMeasurements of the thymus diameter in the three-vessel view, acquired using 2D ultrasound transabdominally.\u003c/p\u003e\n\u003cp\u003eComparison between automated measurement by HeartAssist (A) and manual measurement (B). Although the automated measurement placed the cursor more internally relative to the thymus boundary (arrow), it still measured comparably well in comparison to the manual method.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5718511/v1/9eca80fd64aeed680c7c419a.png"},{"id":81050727,"identity":"ec15b8c3-4476-4523-b35a-4582c86f6104","added_by":"auto","created_at":"2025-04-21 16:02:34","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":11489436,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5718511/v1/31c2deac-4acd-4e59-8cf5-73ab6ebeeda3.pdf"},{"id":79905998,"identity":"d6bca7e8-7835-4a1f-a3c0-a765f7f7275b","added_by":"auto","created_at":"2025-04-04 11:01:23","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":58018,"visible":true,"origin":"","legend":"","description":"","filename":"Tables.docx","url":"https://assets-eu.researchsquare.com/files/rs-5718511/v1/0bf68c45faccd3d1971ab02b.docx"},{"id":79906004,"identity":"fdad5aae-fcb2-44b2-839f-8b9e9ccd6008","added_by":"auto","created_at":"2025-04-04 11:01:24","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":249499,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFINAL.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5718511/v1/f1bbf7304511b7976a86268a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial intelligence based automatic classification, annotation, and measurement of the fetal heart using HeartAssist","fulltext":[{"header":"Introduction","content":"\u003cp\u003eCongenital heart disease (CHD) is the most common congenital anomaly and a leading cause of infantile morbidity and mortality worldwide [1]. An accurate prenatal diagnosis of CHD can improve neonatal outcomes and enhance parental counseling. Fetal echocardiography is the main method for diagnosing CHD, particularly in women with increased risks for fetal cardiac anomalies [2].\u003c/p\u003e\n\u003cp\u003eDespite several guidelines recommending essential scanning views and cardiac parameters for fetal echocardiography [3-6], the prenatal detection rate of CHD remains variable [7, 8]. The following factors contribute to this challenge: 1) fetal heart is small with indistinct anatomical appearance; 2) fetus moves with relative movement of the probe; 3) fetal heartbeats are fast; 4) contrast artifacts make imaging difficult; and 5) some examiners lack expertise in fetal echocardiography. These factors may result in extended time in the acquisition of clear images and missed prenatal diagnosis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDeep learning has become increasingly popular in the field of fetal imaging [9]. With advances in artificial intelligence (AI) technology, automated techniques employing convolutional neural networks (CNNs) have been introduced in fetal echocardiography [10, 11]. However, many studies on AI applications in fetal echocardiography have predominantly focused on view classification, with limited data available on their capability to measure cardiac structures [12-16]. In addition, there is a study on the automatic classification of eight fetal\u0026nbsp;cardiac views using an AI system\u0026nbsp;[16]. While\u0026nbsp;this study encompasses the largest number of cardiac views reported\u0026nbsp;to date, if multiple guidelines are integrated, more than 10 cardiac views could potentially be\u0026nbsp;confirmed by fetal echocardiography\u0026nbsp;[1, 5, 6, 17].\u003c/p\u003e\n\u003cp\u003eTo address these limitations, we collaborated with Samsung Medison to develop an automated tool designed for classifying fetal cardiac views, annotating cardiac structures, and measuring cardiac parameters. This study aimed to evaluate the feasibility of this new system, HeartAssist (Samsung Medison Co., Ltd, Seoul, Korea).\u0026nbsp;\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003e\u003cu\u003e1. Study population\u003c/u\u003e\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study prospectively and retrospectively\u0026nbsp;analyzed the data of fetuses collected at Asan Medical Center, Seoul, Korea, from January 2016 to October 2018. The inclusion criteria were: 1) fetuses from both singleton and multiple pregnancies confirmed by prenatal ultrasonography; 2) required to undergo\u0026nbsp;detailed ultrasonography, including echocardiography by several specialists in maternal fetal medicine: and 3) between 20 and 40 weeks of gestational age (GA); and 4) regardless of the presence of fetal anomalies.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e\u003cu\u003e2. Data acquisition\u003c/u\u003e\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll prenatal ultrasonographic evaluations were conducted using WS80A or HERA W10 (Samsung Medison Co., Ltd, Seoul, Korea) with associated transabdominal probes.\u0026nbsp;This study was approved by the Institutional Review Board\u0026nbsp;of\u0026nbsp;Asan Medical Center (approval number: 2017-0688). Informed consent was obtained from the prospectively recruited subjects and was waived for the retrospectively recruited subjects as the study used pre-collected, anonymized data without direct patient interaction\u0026nbsp;by the Institutional Review Board\u0026nbsp;of\u0026nbsp;Asan Medical Center. The waiver was granted in accordance with the Institutional Review Board\u0026nbsp;of Asan Medical Center, the Declaration of Helsinki, and ethical guidelines.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp; 2.1. 10 cardiac views and 43 parameters\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe recording data included 10 classified views;\u0026nbsp;transverse abdominal view, four-chamber view (4CV), left ventricular outflow tract view (LVOTV), right ventricular outflow tract view (RVOTV), three-vessel view (3VV), three-vessel\u0026nbsp;trachea\u0026nbsp;view (3VTV), three-vessel view with pulmonary artery branches (3VV PA), aortic arch view (AAV), ductal arch view, and bicaval view. Among the 10 views, seven views (4CV, LVOTV, RVOTV, 3VV, 3VTV, 3VV PA, and AAV) were used to measure 43 parameters that were classified into two categories (Table 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe measurement data (35) included cardiac axis, thoracic area, cardiac area, thoracic circumference, cardiac circumference, thoracic transverse diameter, cardiac transverse diameter, left atrium (LA) width, right atrium (RA) width, LA area, RA area, left ventricle (LV) width, right ventricle (RV) width, LV length, RV length, LV area, RV area, mitral valve (MV) annulus diameter, and tricuspid valve (TV) annulus diameter on 4CV; aortic valve (AV) annulus diameter and ascending aorta diameter on LVOTV; pulmonary valve (PV) annulus diameter on RVOTV; PA diameter, aorta diameter, superior vena cava (SVC) diameter, and thymus diameter on 3VV; aortic isthmus diameter and ductus arteriosus (DA) diameter on 3VTV; main pulmonary artery (MPA) diameter and right pulmonary artery (RPA) diameter on 3VV PA; and ascending aorta diameter, proximal transverse aortic arch diameter, distal transverse aortic arch diameter, aortic isthmus diameter, and descending aorta diameter on AAV.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe calculation data (8) included cardiothoracic area ratio (CTAR), cardiothoracic circumference ratio (CTCR), cardiothoracic ratio (CTR), left to right ventricular width ratio [LV/RV (W)], left to right ventricular length ratio [LV/RV (L)], left to right ventricular area ratio [LV/RV (A)], aortic valve to aorta ratio (AV/Aorta), aortic to pulmonary valve ratio (AV/PV).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u0026nbsp; 2.2. The measurement methods of the cardiac parameters\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe cardiac parameters were measured based on previously reported studies [4, 5, 18-30]. All cardiac parameters were measured from the inner to inner edge.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.2.1. Parameters in 4CV\u003c/p\u003e\n\u003cp\u003eThe following criteria were used to measure the parameters in 4CV.\u0026nbsp;①\u0026nbsp;The cardiac axis refers to the angle formed by the line dividing the thorax in half and the extension line of the interventricular septum of the heart.\u0026nbsp;②\u0026nbsp;The area and circumference of the heart were measured with maximal distention of the heart at end-diastole (just after the atrioventricular valves close) along the outer border of the heart.\u0026nbsp;③\u0026nbsp;The cardiac transverse diameter was measured at the level of the atrioventricular valves at end-diastole.\u0026nbsp;④\u0026nbsp;The thoracic area, thoracic circumference, and thoracic transverse diameter were measured at the outer border from rib to rib at the same aforementioned cardiac cycle.\u0026nbsp;⑤\u0026nbsp;The CTAR and CTCR were calculated as the ratio of the area and circumference of the heart and thorax, respectively, and CTR was calculated as the ratio of the transverse diameter of the heart and thorax.\u0026nbsp;⑥\u0026nbsp;The atrial diameters and areas were measured when both atria showed maximal expansion at end-systole (just before the atrioventricular valves opened).\u0026nbsp;⑦\u0026nbsp;The LA and RA widths were measured in the largest transverse diameter.\u0026nbsp;⑧\u0026nbsp;The LA and RA areas were measured, excluding neither the pulmonary veins nor the atrioventricular valves.\u0026nbsp;⑨\u0026nbsp;The ventricular diameters and areas were measured when both ventricles showed maximal expansion at end-diastole.\u0026nbsp;⑩\u0026nbsp;The LV and RV widths were measured below the MV and TV, respectively.\u0026nbsp;⑪\u0026nbsp;The LV and RV lengths were measured from the basal to the apex of the ventricles, including the MV and TV.\u0026nbsp;⑫\u0026nbsp;The LV and RV areas were measured including the endocardium and the moderator band inside the RV cavity.\u0026nbsp;⑬\u0026nbsp;LV/RV (W), LV/RV (L), and LV/RV (A) were calculated using the respective result values from the LV and RV.\u0026nbsp;⑭\u0026nbsp;The TV and MV annulus diameters were measured between the hinge points of the leaflet attachment with maximal expansion when both atrioventricular valves opened at early-diastole.\u003c/p\u003e\n\u003cp\u003e2.2.2. Parameters in the outflow tract views\u003c/p\u003e\n\u003cp\u003eThe following criteria were used to measure the parameters in the outflow tract views.\u0026nbsp;①\u0026nbsp;The AV and PV annuli diameters were measured perpendicular to the great vessels at the level of the valves at mid-systole (when the AV and PV valves opened) in LVOTV and RVOTV, respectively. ②\u0026nbsp;The ascending\u0026nbsp;aorta diameter was measured at maximal size above the AV annulus.\u0026nbsp;③\u0026nbsp;AV/aorta and AV/PV were obtained by calculating the ratio of each value.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.2.3. Parameters in the 3VV\u003c/p\u003e\n\u003cp\u003eThe following criteria were used to measure the parameters in the 3VV.\u0026nbsp;①\u0026nbsp;The diameters of PA, aorta, and SVC were measured at the widest diameter perpendicular to each vessel.\u0026nbsp;②\u0026nbsp;The thymus diameter was measured at the largest transverse diameter.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.2.4. Parameters in the 3VTV\u003c/p\u003e\n\u003cp\u003eThe following criteria were used to measure the parameters in the 3VTV.\u0026nbsp;①\u0026nbsp;The diameters of the aortic isthmus and DA were measured at the largest diameter just before its junction into the descending aorta.\u0026nbsp;②\u0026nbsp;The MPA and RPA diameters were measured at the widest diameter in 3VV PA.\u0026nbsp;③\u0026nbsp;The RPA diameter was measured just after bifurcation of the MPA.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e2.2.5. Parameters in AAV\u003c/p\u003e\n\u003cp\u003eThe aortic diameters were measured at five levels: ascending, proximal transverse, distal transverse, isthmus, and descending aorta in AAV [25].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e\u003cu\u003e3. Development of HeartAssist\u003c/u\u003e\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.1. Data acquisition\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll the acquired images were stored on the computer in the format of digital imaging and communications in medicine (DICOM). Subsequently, these images were classified into 10 cardiac views and labeled [ground-truth (GT)], regardless of the GA, by one expert in fetal echocardiography (M.Y.L). The segmentation label (GT) was also manually delineated by same expert (M.Y.L).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.2. Overview of HeartAssist\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe developed HeartAssist\u0026nbsp;(Samsung Medison Co., Ltd.), which features\u003csup\u003e\u0026nbsp;\u003c/sup\u003eclassification\u0026nbsp;and segmentation models based on the Vision Transformer (ViT) [31] and the CNNs [32].\u003c/p\u003e\n\u003cp\u003e3.2.1. ViT for classification\u003c/p\u003e\n\u003cp\u003eThe \u003cstrong\u003eViT architecture\u003c/strong\u003e is a model that has been extensively validated in \u003cstrong\u003enatural language processing\u0026nbsp;\u003c/strong\u003eand has been successfully extended to various domains by leveraging the strengths of the \u003cstrong\u003etransformer framework\u003c/strong\u003e. Additionally, ViT offers \u003cstrong\u003escalability\u003c/strong\u003e, allowing easy adjustment of \u003cstrong\u003ethe number of layers and parameters\u003c/strong\u003e, which helps balance \u003cstrong\u003eperformance and computational cost\u003c/strong\u003e. Based on these advantages, we utilized the \u003cstrong\u003eViT for image classification\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e3.2.2. CNN for segmentation\u003c/p\u003e\n\u003cp\u003eFor the segmentation model, the \u003cstrong\u003eCNN architecture\u003c/strong\u003e was chosen to better capture \u003cstrong\u003efine anatomical structures\u003c/strong\u003e. Because \u003cstrong\u003eViT processes images in discrete patches\u003c/strong\u003e, it may be less effective in preserving \u003cstrong\u003elocalized spatial information\u003c/strong\u003e compared to CNNs. Given that \u003cstrong\u003efetal cardiac imaging often involves high-resolution images with numerous small structures,\u0026nbsp;\u003c/strong\u003eCNNs are better suited for this task, as they can more effectively analyze \u003cstrong\u003edetailed anatomical features\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003eSpecifically, HeartAssist\u0026nbsp;includes two main functions, namely, an automatic view recognition module (a classification algorithm model) and an anatomy\u0026nbsp;segmentation\u0026nbsp;module (a segmentation model used for annotation and measurement). Both modules were developed using deep learning algorithms under supervised learning. We illustrated the multi-stage deep learning-based framework for fetal echocardiogram analysis in Figure 1, which consists of frame selection, ViT-based classification, CNN-based segmentation, and measurement.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3. Model training and data augmentation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo train the ViT-based classification model, we resized the input images to 224 \u0026times; 224 while keeping the aspect ratio unchanged and performed CutMix [33] and MixUp [34]. CutMix and MixUp are strategies for image data augmentation. They arbitrarily cut out a portion of one image and place it on top of another, and then interpolate the pixel values between the two images. These two methods prevent the model from overfitting the training distribution and increase the likelihood that the model can generalize to examples outside the distribution. Furthermore, they also prevent the model from relying too heavily on specific features to learn.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe segmentation model utilized a higher resolution (1,024 \u0026times; 512) than the classification model because of the precision required for positioning. Supplementary Table S1 shows the hyper-parameters of the classification and segmentation models. The initial values were experimentally determined, with the optimal performance achieved at the values shown in the supplementary Table S1. Supplementary Table S2 presents the performance of our models in relation to their computational and space complexity, measured in terms of the number of FLOPs and parameters.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe original dataset, consisting of 65,324 images, was divided into a training set (93.8% of the total images) and a validation set (6.2% of the total images) through random selection. There was no overlap between the training and validation datasets. Subsequently, data augmentation techniques were employed to improve the performance of the learning algorithm. These techniques included pixel brightness normalization, horizontal flipping, rotation, and scaling. The augmentation layer was randomly applied. Additionally, we implemented weight adjustment, assigning greater importance to underrepresented views to improve learning.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4. Annotation and measurement process\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAfter learning of the segmentation model, annotation and measurement can be performed. The annotation results directly use the center point of the segmented regions through the segmentation network; thus, no additional postprocessing is applied. The list of annotations of cardiac structures is presented in Table 2.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn addition, the area and circumference measurement directly use the contour regions of the segmented results. However, the same method cannot be applied to diameter measurements. Thus, postprocessing based on the area where the structure is divided is employed, or the diameter is measured by dividing the position of the caliper. Area-based postprocessing involves finding the diameters of items that are easy to find within the area of the detected structure according to the measurement standards for each item. Furthermore, diameter items that are difficult to measure solely from the area of the structure are measured by dividing the area centered on both endpoints of the caliper. The size of this area is determined within a certain range depending on the caliper length. Each parameter has been applied using a proper technique summarized in Table 3.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e\u003cu\u003e4.\u0026nbsp;Statistical analysis\u003c/u\u003e\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo assess the classification model performance, we calculated recall, precision, F1-score, and accuracy based on the true positive, true negative, false positive, and false negative values. Recall is the ability of the classifier to correctly identify all the positive samples. Precision reflects the ability of the classifier not to label a sample that is negative as positive. The F1-score is interpreted as a weighted harmonic mean of recall and precision. Accuracy is a simple ratio of true positive and true negative of the total sample. All four metrics are optimized when they reach a value of one and are least effective at a score of zero. The classification performance of HeartAssist was evaluated using a confusion matrix, illustrating the agreement between results and GT values. The detailed confusion matrix is provided in the supplementary Table S3.\u003c/p\u003e\n\u003cp\u003eTo evaluate the accuracy of the annotation, one expert (M.Y.L) visually checked the annotation results of HeartAssist to determine success or failure and calculate the success rate.\u003c/p\u003e\n\u003cp\u003eTo quantitatively evaluate the performance of the segmentation algorithm, the Dice similarity coefficient (DSC) [35, 36] and the error rate were used.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003eThe DSC is a popular metric for image segmentation, particularly in medical image analysis, as it determines whether the segmentation is correctly performed. The DSC is calculated separately for each anatomical structure rather than for all structures as a whole.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn addition, the accuracy of the measurement was measured by comparing the automated measurements with the manual measurements by the expert (M.Y.L). The error on the automated measurement was calculated against the GT, which is defined as the mean of the expert\u0026rsquo;s measurements.\u0026nbsp;\u003cu\u003eThe error rate (%) was calculated as:\u003c/u\u003e\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\"\u003e\u003c/p\u003e\n\u003cp\u003eThe resulting similarity value ranges from 0 to 1, with 1 indicating perfect similarity and the lower values indicating greater discrepancies between the measurements.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eWe collected 65,324 randomly sampled images of fetal echocardiography from 2,985 fetuses over 34 months. In the\u0026nbsp;classification model, 61,277 (93.8%) images were used for training and 4,047 (6.2%) for validation (Table 4). The performances of the classification model are presented in Table 5. The automatically classified cardiac images matched the expert’s classified answers by 99.4% on average. The average of recall, precision, and F1-score in all cardiac views were 0.931, 0.950, and 0.939, respectively. The classification performance of HeartAssist was evaluated using a confusion matrix, illustrating the agreement between results and GT values. The detailed confusion matrix is provided in the supplementary materials (supplementary Table S4).\u003c/p\u003e\n\u003cp\u003eThe datasets utilized for the\u0026nbsp;annotations are presented in Table 6.\u0026nbsp;The accuracy of annotations was excellent across all cardiac views, as shown in Table 7. In particular, the annotations in 4CV, LVOTV, and RVOTV were all correctly displayed (Figure 2A, 2B, and 2C).\u003c/p\u003e\n\u003cp\u003eA total of 37,981 images were used for training and validation in the measurements (Table 8). Table 9 shows the results of the measurement analysis. All cardiac parameters, except for the descending aorta diameter, were automatically measured with a successful rate of over 90%. The mean error rate of all cardiac parameters was 7.624%. The error rates exceeded 10% for the ascending aorta in LVOTV, SVC and thymus diameter in 3VV, aortic isthmus and DA diameter in 3VTV, RPA diameter in 3VV PA, and transverse aortic arch and aortic isthmus diameter in AVV. The mean caliper similarity was 0.613. The cardiac axis in 4CV exhibited the highest similarity (0.825), while the thymus in 3VV exhibited the lowest (0.352). The DSC for area and circumference measurements showed excellent results.\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur study presents several notable advancements. First, while previous AI-based approaches primarily focused on classification, HeartAssist integrates classification, segmentation, and measurement into a single comprehensive system. Second, it enables automated measurement of 43 cardiac parameters, providing a comprehensive assessment of the fetal heart. Finally, the model was trained on a large dataset, contributing to its robustness and potential clinical applicability.\u003c/p\u003e \u003cp\u003eIn our study, 10 cardiac views were selected and analyzed. The 3VV PA was added to the existing standard cardiac views. When any branch of the PA was visible, it was classified as 3VV PA. This view was included based on the guidelines of the American Institute of Ultrasound in Medicine to evaluate the size and orientation of the PA branches [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRecent advancements in AI have greatly improved fetal echocardiography, with deep learning models achieving higher accuracy in classification and segmentation. Various studies have explored AI applications using CNNs and hybrid architectures such as U-Net, ResNet, Mask R-CNN, and DW-Net for automated analysis, as summarized in supplementary Table S4. However, most prior studies focused on limited datasets, primarily 4CV, with relatively small image counts, limiting their clinical applicability. In contrast, our study integrated ViT and CNN architectures to evaluate 10 cardiac views using a significantly larger dataset of 65,324 images, achieving a classification F1-score of 0.939, annotation accuracy of 98.4%, and segmentation DSC of 0.974. These results indicate that our model achieved better performance compared to previous studies.\u003c/p\u003e \u003cp\u003eThe RVOT, 3VV, 3VT, and 3VV PA values were slightly lower compared with those of other cardiac views. These four cardiac views can be challenging to classify even for experts in fetal echocardiography. For example, RVOT and 3VV can be classified differently depending on the visibility of the RV and PV. They may be categorized as either RVOT or 3VV. In such cases, if either RVOT or 3VV is considered to be correct, the accuracy would improve. Furthermore, since our dataset primarily consists of retrospectively collected static images, the automatic recognition of RV and PV structures may have been more challenging. Training with prospectively collected video clips could improve classification performance by allowing the model to better identify cardiac structures in motion rather than relying only on static images.\u003c/p\u003e \u003cp\u003eAlthough the success rates for the automated measurement of most cardiac parameters were high, some parameters exhibited lower success rates than the others. For thoracic area and circumference, the automated system failed to perform measurements when the entire thorax was not included in the image. Similarly, the success rates of the measurements in AAV were low. Among the parameters measured by AAV, the success rate for the descending aorta was the lowest. This could be attributed to the fact that most of the failed cases were in the prone position, which caused the descending aorta to not be clearly delineated.\u003c/p\u003e \u003cp\u003eOur data exhibited a slightly higher mean error rate of 7.6% than the generally acceptable rate of 5% and a moderate to high level of caliper similarity with a value of 0.613. For parameters with small values (e.g., SVC in 3VV, aortic isthmus and DA diameter in 3VTV, RPA diameter in 3VV PA, and transverse arch and aortic isthmus in AA), even a minimal displacement of the cursor can significantly increase the error rate (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA, and \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eB). Nevertheless, the caliper similarity for these values exceeded 0.6, suggesting that the measurements were closely aligned. In the case of the thymus, the boundaries are often unclear, leading to greater discrepancies between measurements and, consequently, a higher error rate and lower caliper similarity. Nevertheless, upon reviewing the images measured by the automated system, we found that the measurements themselves were performed quite accurately (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA, and \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB). Contrarily, for the measurement of the cardiac axis, the reference points are clearly defined, which results in a low error rate and a high caliper similarity between the automated and manual measurements.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe strengths of this study include the application of a deep-learning process using a diverse set of images from both normal and abnormal fetuses. In addition, to ensure the highest accuracy of GT data, all images were evaluated and measured by a single expert (M.Y.L). However, this study also has several limitations. First, the performance of the automated measurements of cardiac parameters remains insufficient and may require additional manual corrections. We believe that incorporating a larger dataset and validating the measurements with multiple experts may further improve the accuracy of cardiac parameter assessment. Second, as most images were retrospectively obtained, accurate determination of the specific phase of the cardiac cycle captured in each image was challenging. The required phase of the cardiac cycle varies for different measurements, and some measurements may not have been taken at the appropriate cardiac cycle. Third, as this study was based on data from a single center, its applicability to broader populations may be limited. Additionally, the use of ultrasound machines from a single manufacturer may have influenced image quality and model performance. Despite these challenges, our results demonstrated comparable outcomes. At present, we are prospectively collecting video clips of the fetal heart to automatically detect each cardiac cycle and measure cardiac parameters in the relevant cardiac view.\u003c/p\u003e \u003cp\u003eFuture advancements to HeartAssist will focus on enhancing its ability to perform accurate real-time assessments while improving its capacity to distinguish between correctly and incorrectly captured images. These advancements will not only assist experts in conducting examinations more efficiently but also help beginners perform ultrasound assessments more effectively by following AI-assisted guidelines, thereby reducing disparities in diagnostic accuracy. Additionally, strengthening the tool\u0026rsquo;s capability to detect CHD will be a crucial step in maximizing its clinical utility.\u003c/p\u003e \u003cp\u003eIn conclusion, our study confirms that the novel automated program, HeartAssist, is a feasible and effective tool for fetal cardiac screening. With it high accuracy in the classification and annotation of cardiac views and structures, as well as measurement of cardiac parameters, HeartAssist mitigates the limitations of manual examination due to inter-examiner variability. It is particularly valuable for examiners with less expertise in fetal echocardiography. We anticipate that HeartAssist will significantly enhance the detection rate of CHDs, ultimately leading to improved perinatal outcomes and better overall care for fetal cardiac health.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor\u0026rsquo;s contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRina Kim: Designed and performed experiments, analyzed data, and wrote the paper.\u003c/p\u003e\n\u003cp\u003eMi-Young Lee: Designed and performed experiments, manually checked all data, analyzed and evaluated data, wrote the paper, and supervised the research.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eYoo Jin Lee: Collected and organized data.\u003c/p\u003e\n\u003cp\u003eHye-Sung Won: Supervised the research and provided final approval of the version to be published.\u003c/p\u003e\n\u003cp\u003eJinki Park, Jihoon Lee, and Kwangyeon Choi: Developed the HeartAssist program and analyzed the results. Also, evaluated data and performed statistics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data used to support the findings of this study are available from the corresponding author upon request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics declarations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe institutional review board of our institution approved this study (approval number: 2017-0688). Informed consent was obtained from the prospectively recruited subjects and was waived from the retrospectively recruited subjects.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNone\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors gratefully acknowledge the support of Samsung Medison for their contributions to this works.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cbr\u003e \u003c/strong\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAbuhamad, A. and R. Chaoui, \u003cem\u003eA Practical Guide to Fetal Echocardiography: Normal and Abnormal Hearts\u003c/em\u003e. 4th ed. 2015, Philadelphia: Wolters Kluwer Health.\u003c/li\u003e\n\u003cli\u003eLee, W., et al., \u003cem\u003eISUOG consensus statement: what constitutes a fetal echocardiogram?\u003c/em\u003e Ultrasound Obstet Gynecol, 2008. \u003cstrong\u003e32\u003c/strong\u003e(2): p. 239-42.\u003c/li\u003e\n\u003cli\u003eThe International Society of Ultrasound in Obstetrics \u0026amp; Gynecology, \u003cem\u003eCardiac screening examination of the fetus: guidelines for performing the \u0026apos;basic\u0026apos; and \u0026apos;extended basic\u0026apos; cardiac scan.\u003c/em\u003e Ultrasound Obstet Gynecol, 2006. \u003cstrong\u003e27\u003c/strong\u003e(1): p. 107-13.\u003c/li\u003e\n\u003cli\u003eCarvalho, J.S., et al., \u003cem\u003eISUOG Practice Guidelines (updated): sonographic screening examination of the fetal heart.\u003c/em\u003e Ultrasound Obstet Gynecol, 2013. \u003cstrong\u003e41\u003c/strong\u003e(3): p. 348-59.\u003c/li\u003e\n\u003cli\u003eAmerican Institute of Ultrasound in Medicine, \u003cem\u003eAIUM Practice Parameter for the Performance of Fetal Echocardiography.\u003c/em\u003e J Ultrasound Med, 2020. \u003cstrong\u003e39\u003c/strong\u003e(1): p. E5-E16.\u003c/li\u003e\n\u003cli\u003eCarvalho, J.S., et al., \u003cem\u003eISUOG Practice Guidelines (updated): fetal cardiac screening.\u003c/em\u003e Ultrasound Obstet Gynecol, 2023. \u003cstrong\u003e61\u003c/strong\u003e(6): p. 788-803.\u003c/li\u003e\n\u003cli\u003evan Velzen, C.L., et al., \u003cem\u003ePrenatal detection of congenital heart disease--results of a national screening programme.\u003c/em\u003e BJOG, 2016. \u003cstrong\u003e123\u003c/strong\u003e(3): p. 400-7.\u003c/li\u003e\n\u003cli\u003eWaern, M., et al., \u003cem\u003ePrenatal detection of congenital heart disease - results of a Swedish screening program 2013-2017.\u003c/em\u003e BMC Pregnancy Childbirth, 2021. \u003cstrong\u003e21\u003c/strong\u003e(1): p. 579.\u003c/li\u003e\n\u003cli\u003eRamirez Zegarra, R. and T. Ghi, \u003cem\u003eUse of artificial intelligence and deep learning in fetal ultrasound imaging.\u003c/em\u003e Ultrasound Obstet Gynecol, 2023. \u003cstrong\u003e62\u003c/strong\u003e(2): p. 185-194.\u003c/li\u003e\n\u003cli\u003eShengfeng, L., et al., \u003cem\u003eDeep learning in Medical Ultrasound Analysis: A Review.\u003c/em\u003e Engineering, 2019. \u003cstrong\u003e5\u003c/strong\u003e(2): p. 261-275.\u003c/li\u003e\n\u003cli\u003eDay, T.G., et al., \u003cem\u003eArtificial intelligence, fetal echocardiography, and congenital heart disease.\u003c/em\u003e Prenatal Diagnosis, 2021. \u003cstrong\u003e41\u003c/strong\u003e(6): p. 733-742.\u003c/li\u003e\n\u003cli\u003eVaanathi, S., et al., \u003cem\u003eAutomated characterization of the fetal heart in ultrasound images using fully convolutional neural networks.\u003c/em\u003e IEEE, 2017: p. 671-674.\u003c/li\u003e\n\u003cli\u003ePhilip, M.E., et al., \u003cem\u003eConvolutional Neural Networks for Automated Fetal Cardiac Assessment using 4D B-Mode Ultrasound.\u003c/em\u003e 2019: p. 824-828.\u003c/li\u003e\n\u003cli\u003eSakai, A., et al., \u003cem\u003eMedical Professional Enhancement Using Explainable Artificial Intelligence in Fetal Cardiac Ultrasound Screening.\u003c/em\u003e Biomedicines, 2022. \u003cstrong\u003e10\u003c/strong\u003e(3): p. 551.\u003c/li\u003e\n\u003cli\u003eWu, H., et al., \u003cem\u003eApplication of Artificial Intelligence in Anatomical Structure Recognition of Standard Section of Fetal Heart.\u003c/em\u003e Comput Math Methods Med, 2023. \u003cstrong\u003e2023\u003c/strong\u003e: p. 13.\u003c/li\u003e\n\u003cli\u003eTaksoee-Vester, C.A., et al., \u003cem\u003eAI supported fetal echocardiography with quality assessment.\u003c/em\u003e Sci Rep, 2024. \u003cstrong\u003e14\u003c/strong\u003e(1): p. 5809.\u003c/li\u003e\n\u003cli\u003eMoon-Grady, A.J., et al., \u003cem\u003eGuidelines and Recommendations for Performance of the Fetal Echocardiogram: An Update from the American Society of Echocardiography.\u003c/em\u003e J Am Soc Echocardiogr, 2023. \u003cstrong\u003e36\u003c/strong\u003e(7): p. 679-723.\u003c/li\u003e\n\u003cli\u003ePaladini, D., S.K. Chita, and L.D. Allan, \u003cem\u003ePrenatal measurement of cardiothoracic ratio in evaluation of heart disease.\u003c/em\u003e Arch Dis Child, 1990. \u003cstrong\u003e65\u003c/strong\u003e(1): p. 20-23.\u003c/li\u003e\n\u003cli\u003eSharland, G.K. and L.D. Allan, \u003cem\u003eNormal fetal cardiac measurements derived by cross-sectional echocardiography.\u003c/em\u003e Ultrasound Obstet Gynecol, 1992. \u003cstrong\u003e2\u003c/strong\u003e(3): p. 175-81.\u003c/li\u003e\n\u003cli\u003eRamada S., S., et al., \u003cem\u003eUltrasonographic left cardiac axis deviation: a marker for fetal anomalies.\u003c/em\u003e Obstet Gynecol, 1995. \u003cstrong\u003e85\u003c/strong\u003e(2): p. 187-191.\u003c/li\u003e\n\u003cli\u003eShapiro, I., et al., \u003cem\u003eFetal cardiac measurements derived by transvaginal and transabdominal cross-sectional echocardiography from 14 weeks of gestation to term.\u003c/em\u003e Ultrasound Obstet Gynecol, 1998. \u003cstrong\u003e12\u003c/strong\u003e(6): p. 404-418.\u003c/li\u003e\n\u003cli\u003eYagel, S., et al., \u003cem\u003eThe three vessels and trachea view (3VT) in fetal cardiac scanning.\u003c/em\u003e Ultrasound Obstet Gynecol, 2002. \u003cstrong\u003e20\u003c/strong\u003e(4): p. 340-5.\u003c/li\u003e\n\u003cli\u003eTheera, T. and T. Teerapong, \u003cem\u003eCardiothoracic ratio in the first half of pregnancy.\u003c/em\u003e J Clin Ultrasound, 2004. \u003cstrong\u003e32\u003c/strong\u003e(4): p. 186-9.\u003c/li\u003e\n\u003cli\u003eSchneider, C., et al., \u003cem\u003eDevelopment of Z-scores for fetal cardiac dimensions from echocardiography.\u003c/em\u003e Ultrasound Obstet Gynecol, 2005. \u003cstrong\u003e26\u003c/strong\u003e(6): p. 599-605.\u003c/li\u003e\n\u003cli\u003ePasquini, L., et al., \u003cem\u003eZ-scores of the fetal aortic isthmus and duct: an aid to assessing arch hypoplasia.\u003c/em\u003e Ultrasound Obstet Gynecol, 2007. \u003cstrong\u003e29\u003c/strong\u003e(6): p. 628-33.\u003c/li\u003e\n\u003cli\u003eCho, J.Y., et al., \u003cem\u003eDiameter of the normal fetal thymus on ultrasound.\u003c/em\u003e Ultrasound Obstet Gynecol, 2007. \u003cstrong\u003e29\u003c/strong\u003e(6): p. 634-8.\u003c/li\u003e\n\u003cli\u003eLopez, L., et al., \u003cem\u003eRecommendations for quantification methods during the performance of a pediatric echocardiogram: a report from the Pediatric Measurements Writing Group of the American Society of Echocardiography Pediatric and Congenital Heart Disease Council.\u003c/em\u003e J Am Soc Echocardiogr, 2010. \u003cstrong\u003e23\u003c/strong\u003e(5): p. 465-95; quiz 576-7.\u003c/li\u003e\n\u003cli\u003eLi, X., et al., \u003cem\u003eZ-score reference ranges for normal fetal heart sizes throughout pregnancy derived from fetal echocardiography.\u003c/em\u003e Prenat Diagn, 2015. \u003cstrong\u003e35\u003c/strong\u003e(2): p. 117-24.\u003c/li\u003e\n\u003cli\u003eGarc\u0026iacute;a-Otero, L., et al., \u003cem\u003eReference ranges for fetal cardiac, ventricular and atrial relative size, sphericity, ventricular dominance, wall asymmetry and relative wall thickness from 18 to 41 gestational weeks.\u003c/em\u003e Ultrasound Obstet Gynecol, 2021. \u003cstrong\u003e58\u003c/strong\u003e(3): p. 388-397.\u003c/li\u003e\n\u003cli\u003eLang, R.M., et al., \u003cem\u003eRecommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging.\u003c/em\u003e J Am Soc Echocardiogr, 2015. \u003cstrong\u003e28\u003c/strong\u003e(1): p. 1-39 e14.\u003c/li\u003e\n\u003cli\u003eHan, K., et al., \u003cem\u003eA Survey on Vision Transformer.\u003c/em\u003e IEEE Trans Pattern Anal Mach Intell, 2023. \u003cstrong\u003e45\u003c/strong\u003e(1): p. 87-110.\u003c/li\u003e\n\u003cli\u003eLeCun, Y., et al., \u003cem\u003eBackpropagation Applied to Handwritten Zip Code Recognition.\u003c/em\u003e Neural Computation, 1989(1): p. 541-551.\u003c/li\u003e\n\u003cli\u003eSangdoo, Y., et al., \u003cem\u003eCutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features.\u003c/em\u003e arXiv, 2019.\u003c/li\u003e\n\u003cli\u003eHongyi, Z., et al., \u003cem\u003emixup: Beyond Empirical Risk Minimization.\u003c/em\u003e 2018.\u003c/li\u003e\n\u003cli\u003eDice, L.R., \u003cem\u003eMeasures of the Amount of Ecologic Association Between Species.\u003c/em\u003e Ecology, 1945. \u003cstrong\u003e26\u003c/strong\u003e(3): p. 297-302.\u003c/li\u003e\n\u003cli\u003eMinaee, S., et al., \u003cem\u003eImage Segmentation Using Deep Learning: A Survey.\u003c/em\u003e IEEE Trans Pattern Anal Mach Intell, 2022. \u003cstrong\u003e44\u003c/strong\u003e(7): p. 3523-3542.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 to 9 are available in the Supplementary Files section\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"artificial intelligence, classification, congenital heart disease, fetal cardiac measurement, fetal echocardiography, image interpretation, computer-assisted","lastPublishedDoi":"10.21203/rs.3.rs-5718511/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5718511/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eObjectives:\u003c/strong\u003e This study evaluated the feasibility of HeartAssist, a novel automated tool designed for classification of fetal cardiac views, annotation of cardiac structures, and measurement of cardiac parameters. Unlike previous AI tools that primarily focused on classification, HeartAssist integrates classification, annotation and measurement capabilities, enabling a more comprehensive fetal cardiac assessment.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e Cardiac images from fetuses (gestational ages 20–40 weeks) were collected at Asan Medical Center between January 2016 and October 2018. HeartAssist was developed using convolutional neural networks to classify 10 cardiac views, annotate 26 structures, and measure 43 parameters. One expert performed manual classifications, annotations, and measurements, which were then compared to HeartAssist outputs to assess feasibility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e A total of 65,324 images from 2,985 fetuses were analyzed. HeartAssist achieved 99.4% classification accuracy, with recall, precision, and F1-score of 0.93, 0.95, and 0.94, respectively. Annotation accuracy was 98.4%, while the automatic measurement success rate was 97.6%, with an error rate of 7.62% and caliper similarity of 0.613.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions: \u003c/strong\u003eHeartAssist is a reliable tool for fetal cardiac screening, demonstrating high accuracy in classifying cardiac views and annotating structures, with comparable outcomes in measuring cardiac parameters. This tool could enhance prenatal detection of congenital heart disease and improve perinatal outcomes.\u003c/p\u003e","manuscriptTitle":"Artificial intelligence based automatic classification, annotation, and measurement of the fetal heart using HeartAssist","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-04 11:01:19","doi":"10.21203/rs.3.rs-5718511/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Accepted","date":"2025-04-08T07:19:45+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"244490140433877992870884047979305349527","date":"2025-04-08T05:21:04+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-06T01:01:01+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-05T20:52:43+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"64728624007326680793484108450499073100","date":"2025-04-05T09:58:53+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-03T15:26:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"234291349944279422845929232127735546690","date":"2025-04-03T14:59:42+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-03T09:25:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"175111877072845197035472391089533766760","date":"2025-04-03T08:46:32+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"1731200067122666016134813170736983910","date":"2025-04-02T23:41:42+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-04-02T16:05:38+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-01T07:43:14+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-03-31T05:59:24+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f230d2ec-b5a3-4a06-b5fc-e55c33c49aaa","owner":[],"postedDate":"April 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":46603453,"name":"Health sciences/Cardiology/Cardiovascular biology/Heart development"},{"id":46603454,"name":"Health sciences/Health care/Medical imaging/Ultrasonography/Echocardiography"}],"tags":[],"updatedAt":"2025-04-21T15:58:27+00:00","versionOfRecord":{"articleIdentity":"rs-5718511","link":"https://doi.org/10.1038/s41598-025-97934-z","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2025-04-16 15:56:58","publishedOnDateReadable":"April 16th, 2025"},"versionCreatedAt":"2025-04-04 11:01:19","video":"","vorDoi":"10.1038/s41598-025-97934-z","vorDoiUrl":"https://doi.org/10.1038/s41598-025-97934-z","workflowStages":[]},"version":"v1","identity":"rs-5718511","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5718511","identity":"rs-5718511","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}