Quick Buzz Feed
Artificial intelligence (AI) is a promising solution to enhance fetal ultrasound access in low-resource settings, addressing shortages of trained staff, limited infrastructure, and unequal diagnostic imaging. This review examines AI's clinical applications, implementation experiences, challenges, and future prospects in fetal ultrasound. AI can support various stages, including gestational age estimation, automated biometry, image quality assessment, anomaly detection, fetal monitoring, image enhancement, and edge deployment, demonstrating encouraging diagnostic performance and potential for task-shifting to non-specialist healthcare providers. However, challenges include underrepresentation of low-resource populations in training datasets, limited prospective validation, infrastructure constraints, training needs, and unresolved regulatory and ethical issues. Future research should focus on local datasets, multicenter validation, resource-efficient AI models, robust regulations, and health economic evaluations to ensure equitable access to quality antenatal imaging and improved maternal and fetal healthcare.
Fetal ultrasound is a crucial, cost-effective antenatal care intervention, recommended by the WHO, but access is highly unequal in low- and middle-income countries due to lack of equipment, electricity, and trained personnel. AI, particularly deep learning, offers a rapid solution for automating image acquisition, biometric measurement, and anomaly detection, potentially enabling non-expert operators through 'blind-sweep' protocols. This review aims to synthesize evidence on AI-assisted fetal ultrasound in low-resource settings, covering its applications, implementation, challenges, and future directions.
A comprehensive literature search was conducted across PubMed/MEDLINE, Scopus, and Google Scholar for studies published between January 2017 and July 2026. The search focused on artificial intelligence, machine learning, or deep learning applications in fetal or obstetric ultrasound, or wearable fetal monitoring, specifically concerning their relevance to low-resource settings. Eligible studies included peer-reviewed original research, reviews, protocols, and guidelines, while studies exclusively from high-income settings, conference abstracts, or non-peer-reviewed preprints were excluded.
Neonatal and perinatal disorders remain leading causes of life years lost, especially in sub-Saharan Africa and South Asia. Access gaps are evident in low prenatal detection rates for conditions like congenital heart defects in low- and middle-income countries. While task-shifting point-of-care ultrasound to non-specialists is feasible, it faces hurdles like training costs, unreliable internet, and unstable electricity, underscoring the need for AI-assisted solutions to reduce dependence on continuous expert oversight.
Accurate gestational age (GA) estimation, often operator-skill dependent, can be revolutionized by AI. Early work in Ethiopia showed deep learning could automatically detect fetal heads and estimate circumference from scans by minimally trained workers. Later meta-analyses reported pooled mean errors for AI models, though with a representation gap for low-income countries. Multicenter studies in Australia, India, UK, USA, and Kenya have demonstrated AI models estimating GA from non-targeted images and blind-sweep videos with high accuracy, even outperforming traditional biometry and showing noninferiority to expert sonographers, supporting scalable deployment in low-resource settings.
To ensure accurate measurements, AI focuses on automating anatomical plane recognition and post-hoc quality assurance. Early deep-learning solutions verified neurosonographic and cardiac four-chamber planes against clinical standards. More recent large-scale evaluations in China demonstrated significant and sustained improvements in fetal ultrasound image quality across 34 hospitals due to AI-based quality control. Initiatives like AMAL-For-Qatar also integrate segmentation models, automated reporting, and super-resolution for low-resource deployment, though clinical benefit beyond process metrics still requires further study.
AI models are being developed to detect specific structural anomalies beyond basic biometry. In Ethiopia, a deep-learning pipeline achieved 98% accuracy in segmenting and classifying microcephaly and macrocephaly from ultrasound images, addressing the shortage of trained personnel. The international CAIFE study aims to build models for differentiating normal fetal hearts from congenital heart defects, specifically to aid detection in non-specialist or low-resource settings. Reviews indicate that AI models integrating various data types can achieve accuracy comparable to experienced clinicians in detecting fetal growth disorders, though retrospective designs and small sample sizes are common limitations.
AI is being applied to continuous fetal surveillance outside specialist units. AutoFHR, an interpretable neural sequential model, accurately estimates fetal heart rate from low-cost, portable Doppler ultrasound signals, outperforming conventional algorithms. A lightweight deep-learning framework for fetal movement monitoring, deployable on resource-constrained microcontrollers, reduces memory footprint and enables extended battery-powered operation, specifically designed for low-resource contexts. Wearable devices for antenatal fetal monitoring are also being reviewed for their potential in continuous, real-time data collection in resource-limited settings.
AI can mitigate the lower image quality from low-cost ultrasound probes often used in low-resource settings. Super-resolution techniques like Real-ESRGAN consistently improve image quality and anatomical plane classification accuracy, even with limited datasets, providing a practical strategy to offset hardware constraints. Additionally, compact multimodal AI architectures, including Siamese networks and quantization, are being developed to reduce model size for edge deployment, enabling prenatal anomaly detection without continuous cloud connectivity.
Early real-world implementation data shows promising results. A mixed-methods study in Sierra Leone evaluated an AI-enabled smartphone-based obstetric ultrasound device, finding 83.8% of scans were of sufficient quality for AI analysis and good acceptance by users. However, infrastructural constraints, supply-chain issues, and lack of comprehensive guidelines were identified as barriers. In Uganda, a pilot protocol is creating an African-specific obstetric Doppler ultrasound dataset to train AI models for pre-eclampsia complication prediction, addressing the global representation gap in obstetric AI datasets.
AI-assisted fetal ultrasound offers several key opportunities: it lowers the skill threshold for image acquisition through AI-guided protocols, enabling task-shifting to non-specialist healthcare workers. Automated quality control provides scalable supervision, as evidenced by improved image standardization in China. Computationally efficient, edge-deployable models address hardware and connectivity limitations in low-resource settings. Crucially, efforts are underway to generate training and validation data from underrepresented populations in sub-Saharan Africa, addressing the bias of commercial AI tools developed predominantly in high-income settings.
Significant challenges hinder widespread AI adoption. Data representativeness and generalizability are major concerns, with most studies originating from high-income settings, raising doubts about accuracy in diverse low-resource populations. Methodological quality is also limited, with many studies being retrospective, single-center, and lacking external validation. Clinical benefit beyond technical process metrics is not yet fully demonstrated. Infrastructural barriers like unreliable connectivity, supply chain issues, and funding gaps persist. Regulatory approval, data privacy, algorithmic bias, and clinician trust also remain unresolved considerations.
Future efforts should prioritize prospective, multicenter validation studies conducted directly in diverse low-resource settings, linking AI performance to tangible perinatal outcomes. Continued investment in computationally efficient, edge-deployable architectures and diversified datasets from underrepresented populations is crucial. Weakly supervised and self-supervised learning approaches can reduce the heavy reliance on manually labeled images. Implementation science frameworks should be proactively integrated into deployment studies to identify and address adoption barriers. Finally, robust regulatory and governance frameworks specific to AI diagnostic tools for non-specialist providers in low-resource settings are essential to ensure safe, ethical, and equitable deployment.
AI-assisted fetal ultrasound holds significant promise for transforming antenatal imaging in low-resource settings by improving access and reliability despite shortages of skilled sonographers. Current evidence shows encouraging performance across various applications, including gestational age estimation, automated biometry, and anomaly detection. However, most studies are retrospective and lack real-world implementation evidence or long-term clinical effectiveness data in low-resource environments. Future research must prioritize prospective multicenter validation in diverse populations, development of locally representative datasets, and the creation of robust regulatory and governance frameworks to ensure safe and equitable deployment. Furthermore, comprehensive health economic evaluations are necessary to determine the affordability and sustainability of these innovations for resource-constrained healthcare systems, ultimately translating promising AI into scalable solutions for improving maternal and fetal health.