Artificial intelligence is already entering routine diagnostic ultrasound, from acquisition guidance to automated measurement, reporting support and diagnostic decision support. Ultrasound remains valued for real-time imaging, portability and absence of ionising radiation, yet quality still depends heavily on operator skill and interpretation. A recent publication in WFUMB Ultrasound Open frames AI-enabled ultrasound as a clinical reality that needs active professional stewardship. The promise includes more consistent imaging, wider access in resource-limited settings and faster workflows. The risks include bias, limited generalisability, opaque model outputs, evolving regulation, deskilling, unresolved liability and unequal access to proprietary tools.
Clinical Uses Already Enter Practice
AI ultrasound already spans tools that assist image acquisition, automate measurements and reporting and support diagnostic decisions. In image acquisition, AI systems guide sonographers and novice operators towards standard imaging planes in real time through on-screen cues or automated feedback on probe adjustment. The strongest validation sits in echocardiography and obstetric ultrasound. AI-guided cardiac acquisition tools have received regulatory clearance and have enabled people without formal echocardiography training to obtain diagnostically interpretable images. In foetal anomaly screening, automated plane detection algorithms identify standard obstetric views with performance comparable to experienced sonographers.
Automated measurement is the most mature category. Foetal biometry algorithms for biparietal diameter, head circumference, abdominal circumference and femur length show reproducibility that meets or exceeds experienced operators under controlled conditions. In cardiac imaging, AI-enabled automated ejection fraction calculation shows strong agreement with expert-derived measurements and has entered commercial imaging platforms. For repetitive measurement tasks, these tools reduce inter-observer variability, a long-standing limitation of ultrasound. Diagnostic support tools risk-stratify thyroid nodules, characterise focal liver lesions and assess ovarian masses for malignancy potential as second-reader systems. Their role depends on appropriate validation and careful integration into clinical workflows.
Must Read: Deep Learning Advances in Ultrasound Segmentation
Potential Benefits for Patients and Systems
The most direct clinical benefit is reduction in operator dependency. Ultrasound quality depends on the skill and experience of the sonographer and on supervision in training environments. AI-guided acquisition and AI-assisted interpretation do not remove the need for trained professionals. Instead, they may compress learning curves, provide real-time quality feedback and function as a consistent expert companion that supports human judgement.
The combination of AI with point-of-care ultrasound on portable and handheld devices carries particular importance for access in resource-limited settings. In low- and middle-income countries, disease burden can coexist with limited availability of imaging specialists. AI-enabled point-of-care ultrasound could extend imaging to rural clinics, community health workers and emergency responders. Pilot programmes in sub-Saharan Africa and Southeast Asia have shown that community health workers using AI-guided handheld ultrasound can identify obstetric complications, cardiac dysfunction and peritoneal pathology with limited formal training. The implications extend to maternal mortality, infectious disease management and emergency triage.
AI-assisted ultrasound also addresses capacity pressures across health systems. Automated worklist prioritisation, AI-generated preliminary reports and real-time quality assurance flags can shorten turnaround, reduce non-diagnostic or inadequate scans requiring repeat examination and support radiologists and sonologists working under high volume and limited time. These effects matter amid global sonographer workforce shortages and rising imaging demand.
Standards Must Address Risk and Equity
Current AI ultrasound systems face unresolved problems that require local validation, professional oversight and clear clinical accountability. Algorithmic bias and generalisability are central concerns because many models use datasets from academic medical centres in high-income countries and patient groups with similar demographic and anthropometric characteristics. Performance may degrade with different body habitus, skin characteristics, comorbidities, paediatric populations, altered tissue acoustic properties or equipment differences. Regulatory clearance or single-centre validation cannot stand in for broad clinical generalisability.
Opaque model outputs further limit trust. Deep learning systems often cannot explain their classifications in a way clinicians can evaluate, validate or override with confidence. Explainable AI methods, including gradient-weighted class activation mapping and attention visualisation, remain early in clinical ultrasound implementation. Without interpretable outputs, automation bias becomes a practical risk. Regulation also remains incomplete, particularly when algorithms may be updated, retrained or modified after clearance in ways that alter performance.
Deskilling and liability add further pressure. Trainees may lose foundational scanning and pattern-recognition skills if AI performs measurements, identifies planes and flags pathology. Existing medico-legal standards still place accountability on interpreting physicians. Premium software licences and proprietary device features may also put AI tools beyond reach of rural and under-resourced settings. Affordable access, open-source development and cost-effectiveness validation remain central to any claim that AI-enabled ultrasound can widen access.
AI in ultrasound already affects devices, diagnoses and clinical workflows. Its value depends on profession-led stewardship rather than passive adoption. Professional societies, including WFUMB, AIUM, EFSUMB and SRU, need guidance on implementation, validation and scope-of-practice boundaries. Publication standards need prospective multi-site validation, transparent dataset reporting, subgroup performance metrics and comparison with existing clinical standards. Training must preserve unaided competence while adding AI literacy. Manufacturers need transparent monitoring, disclosure of limitations and validation in diverse settings. Regulation must also account for evolving algorithms, post-clearance changes and international use.
Source: WFUMB Ultrasound Open
Image Credit: iStock
References:
Dighe M (2026) Artificial Intelligence in Ultrasound Imaging: Inevitable, Promising, and In Need of Responsible Stewardship. WFUMB Ultrasound Open: In Press.
