The role of artificial intelligence in fetal ultrasonography, from automated biometrics to standard plane detection, from quality control to cardiac anomaly screening: evidence, commercial tools, and limitations honestly considered.
⚕️ Target audience: This content is aimed at healthcare professionals and is for educational/current-information purposes. AI tools offer clinical decision support; It does not replace clinician responsibility and approved clinical workflows.
Fetal ultrasonography is an examination that is highly dependent on variables such as image quality, fetal position, maternal habitus and especially operator experience. Different interpretations of the same image by different practitioners (interobserver variability) and the difficulty in obtaining standard planes directly affect the accuracy of the scan. In the last decade, deep learning-based methods have become one of the most active research areas in obstetric imaging, with the promise of reducing this variability and automating routine tasks. This text provides a balanced account of the concrete application areas, level of evidence, and real limitations of artificial intelligence (AI) in fetal ultrasonography.
The architecture used in most applications is the convolutional neural network (CNN—a deep learning architecture that learns features in the image layer by layer, such as edges, texture, and shape) and its derivatives, which are successful in image recognition. In Fiorentino et al.'s comprehensive review of 153 studies published after 2017, these studies were grouped under three main headings: standard plane detection, anatomical structure analysis/segmentation and biometry parameter estimation (Fiorentino et al., Med Image Anal 2023). These topics are discussed below from a clinical perspective.
Fetal biometry (biparietal diameter [BPD], head circumference [HC], abdominal circumference [AC], femur length [FL]) measurements are subject to both intraobserver and interobserver variability; Measurement error is directly reflected in the calculation of gestational age and estimated fetal weight (EFW). AI-based systems aim to make the measurement semi- or fully automatic by automatically recognizing the standard plane and segmenting the relevant structure.
In a study involving temporal (prospective) external validation by Goetz-Fu et al. from Strasbourg University Hospital, a deep neural network-based system first identified the standard HC/AC/FL planes, then performed the measurements and calculated EFW using the Hadlock formula. The system demonstrated a similar absolute relative measurement error compared with an expert sonographer, and a significantly lower absolute relative measurement error compared with less experienced operators (e.g., mean absolute relative error ~1.56% in the AI–expert comparison for AC; ~3.10% for EFW) (Goetz-Fu et al., J Gynecol Obstet Hum Reprod 2025, validation study). This finding suggests that the primary contribution of AI may not be to "outperform the expert," but to bring the performance and reproducibility of less experienced operators closer to the expert level.
Similarly, Płotka et al compared a deep learning method that directly predicts TFA using only a video scan of the fetal abdomen and gestational age against six human readers in a prospectively collected external test set of 100 cases. The mean absolute percentage error between six readers and YZ was 3.75±2.00%; compared to experts only, it was found to be 2.59±1.11%, and no systematic deviation was detected in the Bland-Altman analysis (Płotka et al., Am J Obstet Gynecol MFM 2023). The interesting aspect of this approach is that it can measure from video of a single anatomical region instead of three separate standard planes.
Correct acquisition of the standard plane is a prerequisite for a valid measurement. The anchor dataset in this field is Burgos-Artizzu et al.'s publicly available FETAL_PLANES dataset: over 12,400 images from 1,792 patients at two different hospitals, collected by multiple operators using multiple devices and labeled by an expert maternal-fetal clinician. In this study, current deep CNN architectures showed similar performance to research technicians in classifying the most commonly used fetal planes (abdomen, brain, femur, thorax) and cervix; However, it has been emphasized that performance in fine discrimination of fetal brain subplanes (transthalamic, transcerebellar, transventricular) is still open to improvement (Burgos-Artizzu et al., Sci Rep 2020). The fact that the data set is publicly available has formed the basis of comparison for many subsequent studies.
Plane detection is also directly linked to real-time quality control: systems that instantly mark whether the image obtained during scanning belongs to the correct anatomical plane and quality criteria (e.g. visibility of necessary structures) have the potential to increase both documentation quality and standardization. This acts as a “second eye” providing feedback to the practitioner, especially in training and high-volume scanning environments.
Evaluation of fetal brain structures in second trimester routine screening is also an area suitable for automation. Coronado-Gutiérrez and colleagues from the BCNatal group developed a deep learning pipeline that classifies the brain plane (transventricular/transthalamic/transcerebellar), segments nine structures, and automatically generates measurements on a dataset of routine mid-trimester images. Although plane classification accuracy is high (~98.6%), error rates vary significantly by structure: while the absolute error for four standard head biometrics is below 3.5%, the error increases to ~26% for more difficult measurements such as the degree of Sylvian fissure operculization (Coronado-Gutiérrez et al., Fetal Diagn Ther 2023). This suggests that a “single accuracy figure” can be misleading; This is an important warning that performance varies critically depending on the structure being measured.
Gestational age determination (dating) is one of the areas where AI is relatively mature. In Naz et al.'s systematic review and meta-analysis, the pooled mean error in gestational age estimation of ultrasound image/blind-sweep video-based AI models was reported to be ~4.32 days in 2D images and ~2.55 days in blind-sweep videos; statistical heterogeneity is high (Naz et al., Front Glob Womens Health 2025, meta-analysis). The authors emphasize that this performance may be particularly valuable in low-resource environments where trained interpreters are limited.
The area of AI with the highest potential for clinical impact—and also the one that requires the most attention—is structural anomaly and, in particular, congenital heart disease (CHD) screening. Fetal heart scanning represents an exemplary paradox in this regard: while the five cardiac images recommended by the guidelines together can capture approximately 90% of complex CHD, in real practice sensitivity can be as low as 30%.
In the groundbreaking study by Arnaout et al., an ensemble of artificial neural networks trained on 107,823 images obtained from 1,326 retrospective echocardiograms and screening ultrasounds first recognized the proposed cardiac images and then distinguished normal hearts from complex CHD. In the internal test set (4,108 fetal scans, 0.9% CHD), the model achieved an AUC of 0.99 in normal-abnormal discrimination, sensitivity of 95% (CI 84–99%), specificity of 96%, and negative predictive value of 100%; performance was also preserved on out-of-hospital and low-quality images (Arnaout et al., Nat Med 2021). This is among the strongest evidence that AI, when applied to guideline-compliant imaging, can close a critical global diagnostic gap—but it should be noted that this is an internal test set performance and is distinct from prospective clinical outcome data.
There are also promising findings for finely differentiated cardiac anomalies. Taksøe-Vester et al. developed a model that predicts the risk of aortic coarctation (CoA) using automated cardiac biometric measurements (RV/LV area ratio, main pulmonary artery/ascending aorta diameter ratio, etc.) during anomaly screening at 18–22 weeks; In a case-control design with 73 CoA cases and 7,300 healthy controls, the model achieved an AUC of 0.96, 90.4% sensitivity and 88.9% specificity (Taksøe-Vester et al., Ultrasound Obstet Gynecol 2024). This is a clinically relevant target for CoA, which is often missed prenatally; however, the case-control design may overestimate performance (especially the false-positive burden) in the actual screening population.
🌍 Low-resource settings: A significant proportion of maternal deaths, neonatal deaths and stillbirths occur during the intrapartum period, predominantly in low- and middle-income countries; Routine ultrasound use is limited by a lack of trained sonographers. Blind-sweep protocols + AI-based automated interpretation and tele-echocardiography have the potential to expand expert access (Bai et al., Med Image Anal 2026; Suha et al., Medicina 2025). But if the tools are validated only in high-resource populations, they may deepen inequality rather than reduce it.
Beyond research prototypes, there are also tools available for clinical use and regulatory approval. It is important to accurately convey the scope of their approval; The phrase “FDA-approved” often describes Class II software that is authorized to assist the clinician and not to make independent diagnoses.
One thing to note in this list: some AI-assisted image-acquisition tools on the market (e.g., UltraSight) are FDA cleared for adult cardiac echocardiography and are not approved for obstetric/fetal use. Therefore, they should not be presented as a fetal ultrasound tool—off-label generalization is a common source of misinformation.
With its AI Positioning Statement dated November 2025, the International Society for Ultrasound in Obstetrics and Gynecology (ISUOG) recognizes the potential of AI to contribute to women's health, while emphasizing the need for careful evaluation before implementation and assurances to protect patient/clinician safety. Declaration; It sets out principles such as transparency (clear description of the methods and data sets used), equity and avoidance of bias (demographic and ethnic representativeness of the training data), and risk stratification (administrative/low-risk → clinician-assisted/medium-risk → autonomous decision/high-risk). It also requires the preservation of human supervision, standard operating procedures for the interpretation of inconsistencies, and continuous post-implementation surveillance (ISUOG, 2025).
While the above success figures are promising, the field's evidence base is significantly immature. Honest consideration of these limitations is a requirement of professional responsibility in this field that is prone to exaggeration:
At today's level of evidence, AI in fetal ultrasonography should be positioned as a decision-support and standardization tool that empowers the clinician, not as a diagnostician that replaces the clinician. Its most concrete and immediate contributions are: (1) reducing operator dependency and inter-observer variability in biometrics, especially bringing the inexperienced operator closer to the expert level; (2) improving documentation quality with real-time quality control and standard plane marking; (3) support sensitivity as a “second reader” in high-miss-rate scans such as cardiac; (4) expand expert access for gestational age determination and basic screening in low-resource settings.
Practical principles for responsible use: knowing a tool's cleared indication and scope of clearance (assistive only vs. autonomous); preferring tools externally validated on data similar to your own population whenever possible; always evaluating AI output in clinical context and under human oversight; and prioritizing tools with prospective evidence that adhere to reporting standards (e.g., TRIPOD-AI/CONSORT-AI for AI studies). ISUOG's framework supports the same direction: transparency, equity, human oversight, and post-deployment surveillance.
In summary, AI in fetal ultrasonography has real potential—with strong preliminary evidence demonstrated in biometrics, plane detection, and cardiac screening support; However, the translation of this potential into clinical benefit depends on large-scale external validation, prospective outcome studies, and careful integration that considers equity. Excitement should not get in the way of evidence.
Last updated: . Sources have been verified against their publication dates and current versions. The regulatory status of commercial tools (FDA 510(k) / commercial availability) has been verified from manufacturer and official regulatory sources.