Automation of ultrasound image acquisition with artificial intelligence

AI systems now optimize gain and depth, guide operators to standard planes, and even take over part of probe positioning, shortening the learning curve for point-of-care ultrasound.

When people talk about artificial intelligence in ultrasound, most attention goes to automated diagnosis: algorithms that classify nodules, calculate scores, or flag suspicious findings on an image that has already been acquired. There is, however, an earlier and equally transformative layer that acts during the actual scanning of the patient. This is acquisition automation: systems that adjust image parameters in real time, recognize whether the captured plane is correct, guide the operator to the ideal acoustic window, and, in some settings, even physically control the probe. This article focuses specifically on that capture stage, leaving AI-assisted interpretation and reporting to this site's companion article on automated diagnosis.

What it means to automate acquisition

The term automation covers a spectrum, not a single feature. At the most basic level is image auto-optimization: algorithms that automatically adjust gain, depth, focus, and dynamic range with every probe or patient change, a task that previously required constant manual adjustment by the examiner. A second level is automatic view and plane recognition: the software identifies, from the real-time pixel pattern, whether the image corresponds to an apical four-chamber view, a fetal transventricular plane, or a thyroid short axis, and tells the operator when the target has been reached. A third level adds active guidance, with arrows or voice instructions indicating which way to angle or slide the probe to reach the correct window. At the top of the spectrum are automatic segmentation and measurement during the exam itself — for example, left ventricular contouring with simultaneous ejection fraction calculation — and, finally, semi-autonomous or robotic systems that take over part or all of the physical probe positioning, controlled locally or remotely via tele-guidance.

Echocardiography: view recognition and automated ejection fraction

The most mature and best-documented example of acquisition automation is point-of-care echocardiography. In February 2020, the FDA authorized, via the De Novo pathway, the Caption Guidance software (Caption Health), indicated to assist clinicians in acquiring standardized two-dimensional transthoracic echocardiography images in adults. The system uses machine learning to give real-time instructions on how to move the probe to obtain diagnostic-quality views, and it automatically captures the best clip from each window. A study published in JAMA Cardiology tested the system with eight nurses who had no prior echocardiography experience; they performed 240 AI-guided exams on 240 patients at Northwestern Memorial Hospital and the Minneapolis Heart Institute. Agreement between the images obtained by the AI-guided nurses and reference exams performed by sonographers was 92.5%, with 98.8% accuracy for assessing right ventricular size and function and the presence of pericardial effusion, a result consistent across sex, race, and body mass index.

More recently, automation has moved from view acquisition to measurement itself. In June 2026, the FDA cleared Clarius Ejection Fraction AI, a tool built into the app for Clarius's wireless transducers (PA, PAL, and C3 HD3 models) that automatically calculates left ventricular ejection fraction from the standard cardiac views themselves, without requiring manual tracing of endocardial borders. The manufacturer describes the goal of delivering a numeric result within the roughly 90-second window used to guide resuscitation decisions, reducing reliance on subjective visual estimation — historically one of the most operator-dependent steps in emergency echocardiography.

Obstetrics: fetal biometry and standard-plane detection

Obstetric ultrasound relies heavily on capturing very specific anatomical planes — biparietal diameter, abdominal circumference, femur length — and getting them right is one of the most time-consuming aspects of training in obstetrics and gynecology. Recent reviews on AI-assisted localization of fetal standard planes describe convolutional neural networks that can automatically identify, in real time, whether the displayed section corresponds to the transventricular, transcerebellar, or transthalamic plane of the fetal head, alerting the operator when the plane is adequate for biometry. Studies on automatic identification of fetal abdominal planes using deep learning report classification performance comparable to experienced examiners in distinguishing adequate from inadequate planes for abdominal circumference measurement. More recent models for automatic fetal biometry from ultrasound video, with temporal validation in independent cohorts, aim not only to identify the plane but also to perform structure measurement continuously throughout the exam, rather than relying on a single static frame manually chosen by the operator.

Thyroid and breast: automated volume, segmentation, and scoring

Outside cardiology and obstetrics, automation has also advanced in glandular and breast volume exams. The Automated Breast Volume Scanner (ABVS) sweeps the entire breast in a single motorized protocol, generating a three-dimensional dataset that removes the need for manual quadrant-by-quadrant scanning and allows standardized later review; early comparisons between ABVS and manual B-mode, including case series of several dozen patients, showed comparable performance in lesion detection, with the added benefit of inter-observer reproducibility. More recently, object-detection network models (such as architectures from the YOLO family) have been applied to ABVS images to automatically detect and classify benign and malignant breast lesions. In the thyroid, reviews of AI tools in the radiology literature (including a special series in the American Journal of Roentgenology on AI applications) describe software capable of automatically segmenting nodules, measuring their three axes, and suggesting a classification under TI-RADS-type risk stratification systems, acting as an added layer of standardization on top of the examiner's own measurement rather than a substitute for the clinical decision to biopsy.

Point-of-care use in the emergency department and ICU

Perhaps the strongest clinical argument for acquisition automation lies in the emergency department and intensive care unit, environments where point-of-care ultrasound must be performed by physicians and nurses without formal training in echocardiography or radiology, under time pressure. Recent reviews of AI in cardiac point-of-care ultrasound describe how image-recognition algorithms help non-specialist operators quickly identify pericardial effusion, severe ventricular dysfunction, and other critical findings that guide immediate decisions, acting as a safety net over image capture. A systematic review of controlled trials on AI-assisted point-of-care ultrasound training in novice learners found, consistently across the studies evaluated, improved image quality obtained by novice operators compared with conventional training without AI support, although the authors note methodological heterogeneity among the trials and the need for larger, more standardized studies. In lung point-of-care ultrasound, acquisition-support software has been explored to help less experienced operators recognize patterns such as B-lines and consolidations during the bedside exam itself, reducing reliance on an on-site specialist to validate every image.

Robotic and tele-guided ultrasound: bringing the probe where specialists are scarce

The most advanced level of automation combines robotics with remote control. The Melody system (AdEchoTech, France), a three-degree-of-freedom robotic arm that manipulates the probe under the command of a remote sonographer, is approved for abdominal, pelvic, obstetric, vascular, and musculoskeletal ultrasound. During the 2020 COVID-19 outbreak, the system was used to maintain obstetric care at the health centre in La Loche, an isolated Indigenous community in northern Saskatchewan, Canada, when regular flights were suspended: a sonographer 605 km away performed 21 obstetric exams on 18 patients between April and June of that year, remotely controlling the robotic arm. Limited exams achieved adequate image quality in 81% of cases, but complete second-trimester exams had a lower adequacy rate (20% adequate, 50% inadequate), showing that the technique still faces real limitations with anatomically more demanding exams, along with technical delays recorded in roughly a quarter of sessions. A review published in the Methodist DeBakey Cardiovascular Journal cites roughly ten telerobotic ultrasound clinical studies, totaling about 800 patients, with operator-to-patient distances ranging from 3 to 7,000 kilometers, and mentions gains such as a reduction in time to diagnosis from 114 to 26 days reported in a program in rural Sweden. The same review points to unresolved obstacles: the need for very low-latency networks, underdeveloped haptic feedback, the lack of specific billing codes in most health systems, and operator licensing regulations that vary by region.

What automation actually solves — and what it does not

Taken together, this evidence shows that the most consistent benefit of acquisition automation is not replacing the examiner but shrinking the gap between a novice operator and an exam that is good enough for an initial clinical decision. This is particularly relevant in settings of specialist scarcity — an emergency department without an on-call radiologist, rural units without a resident sonographer, communities served only by occasional visits — where the realistic alternative to AI is not a perfect exam by an expert, but no timely exam at all. The Caption Guidance data show that nurses with no prior echocardiography training can, with AI guidance, obtain images with over 90% agreement against reference exams; the telerobotics data show a real reduction in time to diagnosis in remote settings. At the same time, the studies themselves make the limits clear: adequacy rates drop noticeably for anatomically more complex exams, such as complete second-trimester obstetric studies performed by telerobotics; plane-recognition automation trained mostly on typical anatomy tends to behave less predictably with anatomical variants, difficult body habitus, or atypical pathology not represented in the training data; and none of the systems approved so far replaces the final clinical interpretation — they optimize or guide image capture, but the diagnostic decision remains with the responsible physician. Regulatorily, it is also worth noting that most of these tools were cleared for specific, narrow tasks (one cardiac view, one ejection-fraction calculation, one exam class), not for full autonomous scanning, and that the current standard of care still requires human oversight of the process.

Practical relevance for those learning ultrasound

For those building competence in musculoskeletal or general ultrasound, the practical value of these tools today lies more in standardization assistance than in replacing training itself. A system that flags excessive gain, suggests a depth adjustment, or confirms that a plane is technically adequate can accelerate the initial phase of familiarization with the equipment, freeing up more study time for what AI still does not do well: recognizing normal variation in individual anatomy, correlating the sonographic finding with the clinical history, and deciding on management in the face of an ambiguous result. In other words, automation tends to compress the learning curve for the technical part of the exam — the mechanics of obtaining a technically correct image — without eliminating the need for the clinical part, which still depends on supervised exposure, deliberate practice, and clinical reasoning. Understanding this boundary clearly is what allows these tools to be used as training accelerators, rather than as a shortcut that replaces the real building of sonographic competence.