Sage Journals: Discover world-class research

Abstract

Artificial intelligence (AI) is increasingly integrated into point-of-care ultrasound (POCUS) to enhance its utility in critical care settings. This manuscript explores the current state of AI applications in POCUS, focusing on key domains such as image acquisition, image interpretation, education, task automation, procedural guidance, program development, and quality assurance. AI-driven tools can potentially improve image quality, provide real-time feedback, and assist in the interpretation of ultrasound images, thereby democratizing the use of POCUS across varying levels of operator expertise. This narrative review highlights relevant studies demonstrating the clinical utility of AI in POCUS, discusses the challenges that remain, and provides insights into future developments. The goal is to equip intensivists with a comprehensive understanding of how AI can support POCUS practice today and what advancements are on the horizon.

Keywords

artificial intelligence ultrasound medical education

Introduction

It's in your phone, your car, your thermostat, and now your ultrasound machine. You didn’t ask for it, but there it is - primed and ready to opine on whether your patient really needs that extra bolus of fluid. Artificial intelligence (AI) is quietly permeating our everyday lives, and the Intensive Care Unit (ICU) is not excluded. Point-of-care ultrasound (POCUS) is an indispensable tool to the modern-day intensivist, bringing real-time diagnosis and monitoring to the bedside. However, ultrasound quality and interpretation are highly operator-dependent, posing challenges in the ICU where operator experience varies widely.¹ AI is positioned to bridge this expertise gap and democratize the use of POCUS, with handheld and cart-based devices now routinely incorporating AI-driven features for image optimization, anatomical labelling, and automated measurements.

AI is a term that broadly describes any task performed by a computer meant to mimic human intelligence.² Machine learning (ML) is a more defined subset of AI that mimics not only human intelligence, but human learning. Recall the last time you were studying for an exam using a bank of practice questions. You would make an educated guess, then flip to the back of the textbook to check the answer; if incorrect, you would update your understanding of the concept being tested, and only after the exam would you know whether the question bank was worth the cost you paid. In this analogy, the practice questions are the training data that engineers use to build a ML model. The exam itself is a subset of data set aside that the model has not seen, used to test the model. Whether or not the question bank was worth it is referred to as the ‘fit’ of the model - a good fit means the question bank was reflective of the content in the exam. Finally, imagine going to another school to write their version of the exam. How well you perform here represents the ‘generalizability’ of the knowledge you gained from your practice questions.²

AI is presently able to enhance POCUS across a spectrum of functional domains relevant to critical care. This paper will examine the current state of AI for critical care ultrasound, organized by key domains: image acquisition, image interpretation, education, task automation, procedural guidance, program development, and quality assurance. We highlight current capabilities, relevant studies demonstrating clinical utility, and remaining challenges.³ We aim to provide intensivists with a comprehensive understanding of how AI can support POCUS practice today, and what future developments are on the near horizon.

Applications of Artificial Intelligence in Point-of-Care Ultrasound

Image Acquisition

We all remember struggling to obtain respectable cardiac windows, and AI-guided image acquisition tries to help us forget by providing live feedback on transducer movement and orientation.⁴ Deep learning systems for cardiac ultrasound currently exist, recognizing anatomical structures in real time and displaying on-screen prompts to help probe alignment for standard echocardiographic views.⁵ Beyond this particular tool there have been numerous studies with novice scanners in domains spanning from echocardiography, deep vein thrombosis (DVT) protocols, and Focused Assessment with Sonography with Trauma (FAST) that show improved image acquisition and thereafter increased use of POCUS for routine clinical assessments.^6–10 These systems come timely as critical care fellowships are increasingly expected to include ultrasound training, but faculty resources are limited. For better or for worse, the development of AI applications for POCUS are quickly outpacing the establishment of much needed formal POCUS educational programs.^11–13

These studies underscore that AI guidance can “skill-enable” clinicians, extending ultrasound to situations when human experts are not available. Early evidence supports AI functioning as a capable bedside coach, helping novice users to acquire high-quality images.

Image Interpretation

The most active area of research for AI use in POCUS is image interpretation, and we highlight specific examples by organ system. The selected examples described below are promising tools that can help automate interpretation of POCUS images, whether as a primary assessment for novice scanners or a second pair of eyes for those more experienced. The benefit of consistent performance by AI models cannot be understated: they are immune to fatigue and are not subject to various cognitive biases. However, overreliance on AI tools must be avoided, just as over reliance on POCUS findings without clinical integration is dangerous.¹⁴ As these models continue to mature with more diverse datasets, they exemplify how pattern recognition by AI can flag critical findings to expedite additional investigations and management.

Cardiovascular Ultrasound

AI models have been calculating left ventricular ejection fraction (EF) since at least 2020, and have been shown to agree within 5% of experts in most cases.¹⁵ Such technology has been incorporated into POCUS devices, with handheld models from different companies demonstrating expert-level EF predictions even with images obtained by non-experts.¹⁶ Beyond EF estimation, AI algorithms are interpreting other cardiac findings including pericardial effusions, valvular abnormalities, and diastolic dysfunction.^17–19 AI also has the potential to automate and standardize pulse checks during Advanced Cardiovascular Life Support (ACLS). Knowing the unreliability of pulse checks among healthcare providers,²⁰ a model was trained to use carotid artery compressibility to predict return of spontaneous circulation (ROSC)²¹ and was found to be 96% accurate.

Lung Ultrasound

The COVID-19 pandemic was an impetus to develop automated lung ultrasound (LUS) solutions. B-line detection was an early candidate given its visually simplistic appearance, with highly accurate models cropping up that performed well on multi-institutional data, and some even with the ability to differentiate between COVID and non-COVID pneumonia with Area Under the Curve (AUC) of >0.9.^22,23 Identifying ultrasonic ‘fingerprints’ to differentiate causes of respiratory failure can serve as a low-cost, low-risk, rapid screening tool in any setting, but the jury is out on the utility and accuracy of these models given extensive variability in agreement scores when compared to experts.^24,25 The heterogeneity of model performance using similar AI tools highlights the need for highly diverse, multi-center data to produce more generalizable models.

For pneumothorax detection, one LUS model demonstrated similar diagnostic accuracy to chest x-ray (CXR) but shorter time to diagnosis, providing valuable lead time to intervention.²⁶ Another pneumothorax detection model was validated prospectively on ICU patients with sensitivity and specificity of 92% and 80%, respectively,²⁷ further pushing the envelope to forgo burdensome x-ray machines during times where rapid diagnosis is needed.

With POCUS being the superior tool for effusion and consolidation diagnosis,²⁸ automating detection can provide an accurate, low-cost solution. Existing models are already able to identify and localize both consolidations and effusions with accuracies approaching 90%,^29,30 and a future state combining these kinds of models with guided image acquisition algorithms (as discussed above) would allow accurate LUS in both resource and expertise limited settings to become a reality.

Trauma Ultrasound

A positive FAST exam in the correct clinical context can mobilize resources like few other imaging results, and AI can assist by highlighting subtle anechoic fluid slices that a rushed human may overlook, with 98% accuracy.³¹ The Extended FAST (eFAST) exam is a multi-organ trauma evaluation that adds pneumothorax and pericardial effusion assessment, both of which are currently being studied from an AI perspective. A future state combining these models would result in an all-in-one trauma assessment AI suite for rapid and automated identification of life-threatening injuries.

Vascular Ultrasound

In the undifferentiated patient in shock or respiratory failure, the diagnosis of pulmonary embolism is never far from the mind. Investigation of deep vein thrombosis (DVT) should be routine in these situations,³² and manual (non-AI) 2-point POCUS DVT studies are a cheap and non-invasive screening method with excellent performance characteristics.³³ With respect to AI options, one automated DVT detection algorithm using images obtained by non-experts achieved sensitivities and specificities as high as 96% and 85%, respectively.³⁴ Again, consider the power of guided image acquisition with automated interpretation to democratize DVT exams for non-experts or in areas where radiology expertise is not available.

Neurological Ultrasound

Intensivists caring for patients with acute neurologic injuries must maintain a high degree of suspicion for elevated intracranial pressures (ICP).³⁵ Aside from the gold standard ICP monitors, clinicians rely on bedside examination and computerized tomography (CT) imaging as surrogate markers for ICP, neither of which is sensitive or specific.³⁶ Optic nerve sheath diameter (ONSD), measured using POCUS, is a non-invasive method of evaluating elevated ICP which has significant correlation with invasive measurements.³⁷ Automating ONSD measurement could be a helpful tool to screen for elevated ICPs without need for transport or radiation, and one model demonstrated significant correlation (r = 0.7) with expert measurements.³⁸

Task Automation

Deriving quantitative measurements from POCUS can potentially enrich understanding of deranged physiologic states. Tracing Doppler waveforms or measuring distances are a subjective skill - deciding which part of the envelope to include or optimal positioning for linear measurements. Additionally, measurement errors are compounded by formulae that double or square the measured values. AI can potentially reduce intra- and inter-operator variability, while also saving time.

Non-invasive hemodynamic measurements are increasingly used for personalized shock and volume management,³⁹ and ultrasound manufacturers have responded by offering automated measurement tools, such as auto-velocity time integral (VTI) and auto-inferior vena cava (IVC). Accurate and consistent measurement of LVOT VTI is crucial, especially if it is being serially trended to assess volume responsiveness (VR).⁴⁰ Several studies have demonstrated good agreement between these tools when tested against expert users (κ = 0.498-0.655),^41,42 however conflicting studies suggest poor agreement and systematic underestimation of measurements.^42,43

Intensivists are ever in pursuit of an accurate measure of VR. Although fraught with challenges, measuring IVC collapsibility and distensibility is a commonly used method,^44,45 but despite its seeming simplicity, the inter-rater reliability is poor.⁴⁶ Early attempts to automate this process with AI suggest excellent agreement with manual expert measurements, in addition to near perfect intra- and inter-observer reliability (ICC 0.96-0.99),⁴⁷ but conflicting studies also exist.⁴¹

The dizzying amount of variability in these highlighted studies that test the same automated measurement tool suggests there is: 1) room for improvement in these FDA-approved devices and 2) still a degree of operator dependency that requires solid fundamental image acquisition. When paired with automated image guidance algorithms as above, these task automation tools have the potential to minimize variability in measurements while saving time. It also opens the door to those less familiar with quantitative assessments. Clinicians must still be vigilant about the caveats that come with quantitative assessments. At the same time there are other quantitative POCUS techniques that could benefit from AI automation, including the Venous Excess Ultrasound Score (VExUS) that is gaining popularity to assess solid organ congestion and fluid tolerance.⁴⁸

Procedural Guidance

POCUS has tremendous safety benefits related to performing procedures in the ICU.^49,50 AI can potentially act as a heads-up display, highlighting pertinent anatomic structures on a live ultrasound image, and such algorithms can successfully identify vessels, bones, tendons, and nerves on upper extremity ultrasound clips for vessel cannulation, for example.⁵¹ Handheld devices also offer visualization tools that improve operator awareness during vascular access,⁵² and AI systems have been shown to improve cannulation time in difficult targets.⁵³

Other procedural applications include localizing safe pockets for tube thoracostomy or paracentesis. Algorithms have already been developed to map normal anatomy in LUS images, such as rib shadows and pleurae in addition to pathologic findings such as consolidations.^29,54 This concept can be easily applied for locating safe ascitic pockets for paracentesis, with proof-of-concept models already developed.⁵⁵

At the time of this review, many assistive features currently include clinicians “in the loop”, meaning they help the proceduralist rather than fully automating the action. As these AI tools become more vigorously vetted against diverse patient populations, they provide an opportunity for safer, more accessible avenues for procedures in the ICU. As with diagnostic interpretation, rigorous clinical trials and careful implementation will determine how quickly and broadly these tools are adopted.

Program Development and Quality Assurance

Ongoing establishment of POCUS use has necessitated reliable archiving, quality assurance, and reporting.⁵⁶ An expert who oversees a POCUS program will assume responsibility for reviewing the quality of images, confirming appropriate interpretation, and accurate documentation of findings. AI can improve the structure and delivery of POCUS programs by automating existing workflows. Many models have been developed to accurately identify various organs and specific views in scanning protocols.^57–59 Automated view classification not only enables powerful filtering of a POCUS archiving database for educational or research purposes, but can also improve AI workflow efficiencies.⁵⁸ This has the potential to accelerate data labelling in AI research by providing clinicians with preliminary labels, similar to an ECG computer interpretation, that are to be reviewed and adjusted if necessary.

The various applications of AI in POCUS discussed up until this point sets the stage for automated report generation. A preliminary report that includes image quality, image interpretation, and automated measurements for expert review can improve the efficiency of the documentation process. Automated report generation has been developed for CXRs, with one AI model producing up to 77% reports deemed equivalent or preferable to clinician reports, increasing to 94% for cases without abnormal findings.⁶⁰ These models can also be deployed as a triage tool; for example a prospective CXR AI model was able to triage urgent cases with 99% specificity to reduce turnaround times by 77%.⁶¹ Automating report generation also has the added benefit of reducing the administrative burden for billing.⁶² Given the multi-organ nature of POCUS, automated report generation can weave together insightful cardio-pulmonary data - combining information about cardiac function, pulmonary pathology, and VExUS scores to present an ultrasonic profile of a patient that can guide resuscitative measures.

AI can improve workflow efficiencies in existing POCUS program infrastructure by intelligently archiving clips, identifying quality issues, generating preliminary reports, triaging abnormal scans, and improving documentation and billing. Offloading POCUS program leaders from these tasks can free them to coach scanning at the bedside and continue innovating their programs. The promise of this technology can also facilitate the development of remote POCUS programs without local expertise.

Discussion and Future Directions

With the potential to revolutionize POCUS, AI technologies must undergo an important vetting process prior to real-world deployment. The power of AI algorithms depends on the quality of data they are trained on, and just as with clinical research having a diverse representation of patients in training data is crucial for AI models to perform without bias. The generalizability (or external validity) of an AI model refers to how well it can perform on data it has never seen before, typically from different institutions.⁶³ This is particularly important for POCUS given a high degree of variability in images owing to a wide range of ultrasound machines, probes, image presets, and operator scanning preferences.⁶⁴ Only 6–7% of studies investigating AI algorithms for conventional radiology imaging have validated their models on external datasets,^65,66 of which 81% report a performance degradation.⁶⁷ Some AI algorithms for POCUS have already been pressure tested against multi-center data,⁶⁸ but this remains the exception rather than the rule. Multi-center validation and fine-tuning of lung ultrasound models have also been reported.^69,70

Public datasets of deidentified medical imaging from Picture Archiving Communication Systems (PACS) have accelerated AI development by democratizing data for algorithm development.^69–72 With POCUS archiving solutions being inconsistent or absent across hospital systems,⁷³ public datasets have been crucial to crowdsource precious images that capture the variability inherent in point-of-care imaging. The COVID-19 pandemic ignited efforts to gather lung ultrasound data for development of AI screening tools.^70,74 Ongoing contribution to these public datasets for POCUS images will be central to developing highly generalizable AI models.

Using patient data to build algorithms or databases requires attention to data privacy. A major step in data preparation for AI research includes deidentification of images- removing sensitive information including name, medical record number, and date of birth.⁷⁵ National policies, such as the U.S. Health Insurance Portability and Accountability Act (HIPAA), outline identifiers that must be removed to preserve patient privacy. Imaging formats must be considered as they may have embedded patient information, such as the Digital Imaging and Communication in Medicine (DICOM) standard that ultrasound images may be stored as.⁷⁶ Deidentification pipelines for ultrasound DICOM images already exist, however a significant bottleneck in AI research is encountered once investigators must navigate data sharing agreements requiring legal and IT expertise.^77,78 A novel solution that maintains patient privacy by removing the need for inter-institutional data transferring is federated learning. Instead of multiple hospitals sending data to a single institution for AI development, an AI model is sent to participating hospitals to be trained on local data, after which the individual models are amalgamated into a final model.⁷⁹ Although this will require local IT expertise to implement, this decentralized method of AI training can avoid sharing patient data and time delays relating to data sharing agreements, ultimately accelerating AI research.

An inevitability of AI algorithms is performance degradation over time given changes in the data it will see, such as introduction of new ultrasound machines, probes, software preprocessing, and changes in disease patterns. This concept is referred to as data drift.⁸⁰ This was illustrated by analyzing CXRs before and after the emergence of COVID-19, detecting global, measurable changes in CXR features.⁸¹ Anticipation and defending against data drift should be undertaken by surveilling model performance and routinely re-training AI models. Machine learning operations (MLOps) is an existing framework that defines a continual learning pipeline for AI models, providing guiding principles for model monitoring, data management, model versioning, and deployment.⁸² Just as clinicians must keep up to date with new evidence and guidelines, AI models must keep up with the constant evolution of data.

An emerging technique called multi-modal AI addresses the old adage of ‘clinical correlation required’. These models ingest multiple types of data (imaging, text, laboratory, waveform, etc) to generate conclusions,⁸³ just as an intensivist would evaluate a patient in shock with ultrasound, perfusion makers, and vital signs. One multi-modal model, for example, was able to differentiate between Alzheimer's disease, mild cognitive impairment, and normal controls using MRI, clinical notes, patient demographic data, and clinical prediction scores.⁸⁴ Compared to unimodal models that use only images or text, superior performance was demonstrated across five separate public datasets.⁸⁴ Development of multi-modal models that include POCUS will help intensivists navigate complex multi-organ interdependencies when assessing shock states or volume status.

Despite the bright future of AI, it is not meant to replace clinical reasoning. A misinterpretation of AI's role could lead to over-reliance, and clinicians must be able to contextualize AI findings using their domain expertise. Thus, there will still be value in knowing the principles and caveats behind POCUS applications. The value of these algorithms lies not only within the technology itself but also the deliberate, expert-led evaluation by POCUS experts and educators to ensure clinically relevant and safe application. In critical care, the stakes are high, so dual validation (AI and clinician) is likely to remain the standard approach.

Conclusion

We appear to be at an inflection point where AI in POCUS transitions from novelty to routine assistant. In the ICU of tomorrow, AI may guide a resident's hand to get the ultrasound view, analyze the image, fill the report, and even suggest the next intervention – but the skilled intensivist will always be at the helm, interpreting those suggestions in light of the whole patient. The synergy of clinician expertise and AI precision holds great promise for improving critical care outcomes. The path forward will require continuing to rigorously evaluate these technologies and ensuring they are implemented in a way that maximizes benefit and minimizes risk. With careful oversight, AI will become an integral part of the critical care ultrasound landscape, helping fulfill the potential of POCUS as a tool for faster, smarter, and more equitable care at the bedside.

Footnotes

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Derek Wu

Robert Arntfield

Scott J. Millington

References

Adelman

Deshwal

Pradhan

. Critical care ultrasound competency of fellows and faculty in pulmonary and critical care medicine: A nationwide survey. POCUS J. 2023;8(2):202-211.

Janiesch

Zschech

Heinrich

. Machine learning and deep learning. Electron Mark. 2021;31(3):685-695.

Geis

Brady

, et al. Ethics of artificial intelligence in radiology: Summary of the joint European and north American multisociety statement. Can Assoc Radiol J. 2019;70(2):329-334.

Holden

Portillo

Salame

. Skills classification in cardiac ultrasound with temporal convolution and domain knowledge using a low-cost probe tracker. Ultrasound Med Biol. 2021;47(10):3002-3013.

Narang

Bae

Hong

, et al. Utility of a deep-learning algorithm to guide novices to acquire echocardiograms for limited diagnostic use. JAMA Cardiol. 2021;6(6):624-632.

Mears

Kaleem

Panchamia

, et al. Leveraging the capabilities of AI: Novice neurology-trained operators performing cardiac POCUS in patients with acute brain injury. Neurocrit Care. 2024;41(2):523-532.

Baum

Tandel

Ren

, et al. Acquisition of cardiac point-of-care ultrasound images with deep learning: A randomized trial for educational outcomes with novices. CHEST Pulmonary. 2023;1(3):100023.

Mor-Avi

Genovese

Yamat

, et al. Real-Time artificial intelligence-based guidance of echocardiographic imaging by novices: Image quality and suitability for diagnostic interpretation and quantitative analysis. Circ Cardiovasc Imaging. 2023;16(11):e015569.

Nothnagel

Aslam

. Evaluating the benefits of machine learning for diagnosing deep vein thrombosis compared with gold standard ultrasound: a feasibility study. BJGP Open. 2024;8(4):1-8.

10.

Kulik

Bell

Holden

. FAST Skill assessment from kinematics data using convolutional neural networks. Int J Comput Assist Radiol Surg. 2024;19(1):43-49.

11.

Stowell

Kessler

, et al. Critical care ultrasound: A national survey across specialties. J Clin Ultrasound. 2018;46(3):167-177.

12.

Eisen

Leung

Gallagher

Kvetan

. Barriers to ultrasound training in critical care medicine fellowships: A survey of program directors. Crit Care Med. 2010;38(10):1978-1983.

13.

Mosier

Malo

Stolz

, et al. Critical care ultrasound training: A survey of US fellowship directors. J Crit Care. 2014;29(4):645-649.

14.

Prager

Olivo-Vergara

Arntfield

, et al. Medico-legal risks of point-of-care ultrasound: A closed-case analysis of Canadian medical protective association medico-legal cases. Ultrasound J. 2024;16(1):16.

15.

Ouyang

Ghorbani

, et al. Video-based AI for beat-to-beat assessment of cardiac function. Nature. 2020;580(7802):252-256.

16.

Papadopoulou

S-L

Kyriakou

Koushiappa

, et al. Artificial intelligence-assisted evaluation of cardiac function by oncology staff in chemotherapy patients. Eur Heart J Digit Health. 2024;5(3):278-287.

17.

Yıldız Potter

Leo

Vaziri

Feldman

. Automated detection and localization of pericardial effusion from point-of-care cardiac ultrasound examination. Med Biol Eng Comput. 2023;61(8):1947-1959.

18.

Holste

Oikonomou

Mortazavi

, et al. Severe aortic stenosis detection by deep learning applied to echocardiography. Eur Heart J. 2023;44(43):4592-4604.

19.

Gottlieb

Schraft

O’Brien

Patel

. Diagnostic accuracy of artificial intelligence for identifying systolic and diastolic cardiac dysfunction in the emergency department. Am J Emerg Med. 2024;86:115-119.

20.

Eberle

Dick

Schneider

Wisser

Doetsch

. Checking the carotid pulse check: Diagnostic accuracy of first responders in patients with and without a pulse. Resuscitation. 1996;33(2):107-116.

21.

Park

Lee

Kim

, et al. Artificial intelligence-based evaluation of carotid artery compressibility via point-of-care ultrasound in determining the return of spontaneous circulation during cardiopulmonary resuscitation. Resuscitation. 2024;202:110302.

22.

Arntfield

Tschirhart

, et al. Automation of lung ultrasound interpretation via deep learning for the classification of normal versus abnormal lung parenchyma: A multicenter study. Diagnostics (Basel). 2021;11(11):1-17.

23.

Arntfield

VanBerlo

Alaifan

, et al. Development of a convolutional neural network to differentiate among the etiology of similar appearing pathological B lines on lung ultrasound: A deep learning study. BMJ Open. 2021;11(1):e045120.

24.

Moore

Wang

Battisti

, et al. Interobserver agreement and correlation of an automated algorithm for B-line identification and quantification with expert sonologist review in a handheld ultrasound device. J Ultrasound Med. 2022;41(9):2487-2495.

25.

Russell

Ehrman

Cosby

, et al. B-line quantification: Comparing learners novice to lung ultrasound assisted by machine artificial intelligence technology to expert review. The Ultrasound Journal. 2021;13(1):1-7.

26.

Yang

Zhao

Wang

Gao

. Comparison of lung ultrasound assisted by artificial intelligence to radiology examination in pneumothorax. J Clin Ultrasound. 2024;52(8):1051-1055.

27.

Clausdorff Fiedler

Prager

Smith

, et al. Automated real-time detection of lung sliding using artificial intelligence: A prospective diagnostic accuracy study. Chest. 2024;166(2):362-370.

28.

Nazerian

Volpicelli

Vanni

, et al. Accuracy of lung ultrasound for the diagnosis of consolidations when compared to chest computed tomography. Am J Emerg Med. 2015;33(5):620-625.

29.

Kessler

Zhu

Gregory

, et al. Development and testing of a deep learning algorithm to detect lung consolidation among children with pneumonia using hand-held ultrasound. PLoS One. 2024;19(9):e0309109.

30.

Vukovic

Stojic

Jakovljevic

, et al. Automatic deep learning-based pleural effusion segmentation in lung ultrasound images. BMC Med Inform Decis Mak. 2023;23(1):274.

31.

Levy

Castle

Virodov

, et al. Artificial intelligence evaluation of focused assessment with sonography in trauma. J Trauma Acute Care Surg. 2023;95(4):706-712.

32.

Varrias

Palaiodimos

Balasubramanian

, et al. The use of point-of-care ultrasound (POCUS) in the diagnosis of deep vein thrombosis. J Clin Med. 2021;10(17):3903.

33.

Zaki

Bashir

Shaban

, et al. Deep venous thrombosis (DVT) diagnostics: Gleaning insights from point-of-care ultrasound (PoCUS) techniques in emergencies: A systematic review and meta-analysis. Ultrasound J. 2024;16(1):37.

34.

Kainz

Makrapoulos

Oppenheimer

, et al. Non-invasive diagnosis of deep vein thrombosis from ultrasound imaging with machine learning. npj Digit Med. 2021;4(1):1-18.

35.

Badri

Chen

Barber

, et al. Mortality and long-term functional outcome associated with intracranial pressure after traumatic brain injury. Intensive Care Med. 2012;38(11):1800-1809.

36.

Rosenberg

Shiloh

Savel

Eisen

. Non-invasive methods of estimating intracranial pressure. Neurocrit Care. 2011;15(3):599-608.

37.

Wang

, et al. Ultrasonographic optic nerve sheath diameter correlation with ICP and accuracy as a tool for noninvasive surrogate ICP measurement in patients with decompressive craniotomy. J Neurosurg. 2020;133(2):514-520.

38.

Hirzallah

Bose

Maltz

. Automation of ultrasonographic optic nerve sheath diameter measurement using convolutional neural networks. J Neuroimaging. 2023;33(6):898-903.

39.

Yoshida

Fujiwara

Takei

, et al. Diagnostic accuracy of point-of-care ultrasound for shock: A systematic review and meta-analysis. Crit Care. 2023;27(1):1-11.

40.

Desai

Garry

. Assessing dynamic fluid-responsiveness using transthoracic echocardiography in intensive care. BJA Educ. 2018;18(7):218-226.

41.

Gohar

Herling

Mazuz

Tsaban

Gat

Kobal

. Artificial intelligence (AI) versus POCUS expert: A validation study of three automatic AI-based, real-time, hemodynamic echocardiographic assessment tools. J Clin Med. 2023;12(4):1-13.

42.

Shaikh

Berman

Ahmed

, et al. Measuring the accuracy of cardiac output using POCUS: The introduction of artificial intelligence into routine care. Ultrasound J. 2022;14(7):1-8.

43.

Louart

Muller

Emond

Boulet

Roger

. Agreement between manual and automatic ultrasound measurement of the velocity–time integral in the left ventricular outflow tract in intensive care patients: Evaluation of the AUTO-VTI® tool. J Clin Monit Comput. 2024;39(2):355-364.

44.

Millington

. Ultrasound assessment of the inferior vena cava for fluid responsiveness: Easy, fun, but unlikely to be helpful. Can J Anesth. 2019;66(6):633-638.

45.

Orso

Paoli

Piani

, et al. Accuracy of ultrasonographic measurements of inferior vena cava to determine fluid responsiveness: A systematic review and meta-analysis. J Intensive Care Med. 2020;35(4):354-363.

46.

Bussmann

Sharma

Mcgregor

Hulme

Harris

. Observational study in healthy volunteers to define interobserver reliability of ultrasound haemodynamic monitoring techniques performed by trainee doctors. Eur J Emerg Med. 2019;26(3):217-223.

47.

Damodaran

Kulkarni

Gunaseelan

Raj

Kanchi

. Automated versus manual B-lines counting, left ventricular outflow tract velocity time integral and inferior vena cava collapsibility index in COVID-19 patients. Indian J Anaesth. 2022;66(5):368-374.

48.

Prager

Arntfield

Olivo-Vergara

, et al. Venous congestion in septic shock quantified with point-of-care ultrasound: A pilot prospective multicentre cohort study. Can J Anesth. 2024;71(5):640-649.

49.

Hrics

Wilber

Blanda

Gallo

. Ultrasound-assisted internal jugular vein catheterization in the ED. Am J Emerg Med. 1998;16(4):401-403.

50.

Lichtenstein

Hulot

Rabiller

Tostivint

Mezière

. Feasibility and safety of ultrasound-aided thoracentesis in mechanically ventilated patients. Intensive Care Med. 1999;25(9):955-958.

51.

Blaivas

Arntfield

White

. Creation and testing of a deep learning algorithm to automatically identify and label vessels, nerves, tendons, and bones on cross-sectional point-of-care ultrasound scans for peripheral intravenous catheter placement by novices. J Ultrasound Med. 2020;39(9):1721-1727.

52.

iQ3 Ultrasound. https://www.butterflynetwork.com/iq3?srsltid=AfmBOooApWhXJShQQ8HFuOZ-GNtP3_iEk3s-591JKod4hQfGsfEcu3Bj.

53.

Abe-Doi

Murayama

Kajiwara

, et al. Effects of ultrasound with an automatic vessel detection system using artificial intelligence on the selection of puncture points among ultrasound beginner clinical nurses. J Vasc Access. 2024;25(6):1252-1260.

54.

Howell

Ingram

Lapham

Morrell

McLaughlan

. Deep learning for real-time multi-class segmentation of artefacts in lung ultrasound. Ultrasonics. 2024;140:107251.

55.

Lin

Cao

, et al. Deep learning for emergency ascites diagnosis using ultrasonography images. J Appl Clin Med Phys. 2022;23(9):e13695.

56.

Arntfield

Millington

Ainsworth

, et al. Canadian Recommendations for critical care ultrasound training and competency. Can Respir J. 2014;21(6):341-345.

57.

Blaivas

Arntfield

White

. DIY AI, deep learning network development for automated image classification in a point-of-care ultrasound quality assurance program. J Am Coll Emerg Physicians Open. 2020;1(2):124-131.

58.

VanBerlo

Chen

, et al. Enhancing annotation efficiency with machine learning: Automated partitioning of a lung ultrasound dataset by view. Diagnostics. 2022;12(10):2351.

59.

Naser

Chandrashekar

Maamari

, et al. Artificial intelligence-based classification of echocardiographic views. Eur Heart J Digit Health. 2024;5(2):260-269.

60.

Tanno

Arcaspy

Hou

, et al. Collaboration between clinicians and vision–language models in radiology report generation. Nat Med. 2024;31(2):599-608.

61.

Sridharan

Channin

Sadowski

, et al. Real-World evaluation of an AI triaging system for chest X-rays: A prospective clinical study. Eur J Radiol. 2024;181:111783.

62.

Flannigan

Adhikari

. Point-of-Care ultrasound work flow innovation: Impact on documentation and billing. J Ultrasound Med. 2017;36(11):2467-2474.

63.

LeCun

Bengio

Hinton

. Deep learning. Nature. 2015;521(7551):436-444.

64.

Sonko

Arnold

Kuznetsov

. Machine learning in point of care ultrasound. POCUS J. 2022;7(1):78-87.

65.

Kim

Jang

Kim

Shin

Park

. Design characteristics of studies reporting the performance of artificial intelligence algorithms for diagnostic analysis of medical images: Results from recently published papers. Korean J Radiol. 2019;20(3):405-410.

66.

Oeding

Krych

Pearle

Kelly

Kunze

. Medical imaging applications developed using artificial intelligence demonstrate high internal validity yet are limited in scope and lack external validation. Arthroscopy. 2025;41(1):455-472.

67.

Yao

Cheng

Pan

Kitamura

. Deep learning in neuroradiology: A systematic review of current algorithms and approaches for the new wave of imaging technology. Radiol Artif Intell. 2020;2(2):1-6. doi:10.1148/ryai.2020190026

68.

Kagiyama

Abe

Kusunose

, et al. Multicenter validation study for automated left ventricular ejection fraction assessment using a handheld ultrasound with artificial intelligence. Sci Rep. 2024;14(1):1-8.

69.

EchoNet-Dynamic Cardiac Ultrasound. Center for Artificial Intelligence in Medicine & Imaging https://aimi.stanford.edu/datasets/echonet-dynamic-cardiac-ultrasound.

70.

CheXpert Plus. Center for Artificial Intelligence in Medicine & Imaging https://aimi.stanford.edu/datasets/chexpert-plus.

71.

CT Pulmonary Angiography. Center for Artificial Intelligence in Medicine & Imaging https://aimi.stanford.edu/datasets/CT-Pulmonary-Angiography.

72.

Johnson

AEW

Bulgarelli

Shen

, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10(1):1-9.

73.

Nathanson

Williams

Gupta

, et al. Current use and barriers to point-of-care ultrasound in primary care: A national survey of VA medical centers. Am J Med. 2023;136(6):592-595.e2.

74.

Roy

Menegazzi

Scuccimarri

, et al. Deep learning for classification and localization of COVID-19 markers in point-of-care lung ultrasound. IEEE Trans Med Imaging. 2020;39(8):2676-2687.

75.

Willemink

Koszek

Hardell

, et al. Preparing medical imaging data for machine learning. Radiology. 2020;295(1):1-15. doi:10.1148/radiol.2020192224

76.

Larobina

Murino

. Medical image file formats. J Digit Imaging. 2014;27(2):200-206.

77.

Monteiro

Costa

Oliveira

. A De-identification pipeline for ultrasound medical images in DICOM format. J Med Syst. 2017;41(1):1-16.

78.

Batlle

Dreyer

Allen

, et al. Data sharing of imaging in an evolving health care world: Report of the ACR data sharing workgroup, part 1: Data ethics of privacy, consent, and anonymization. J Am Coll Radiol. 2021;18(11):1646-1654.

79.

Kaissis

Makowski

Rückert

Braren

. Secure, privacy-preserving and federated machine learning in medical imaging. Nat Mach Intell. 2020;2(1):305-311.

80.

Sahiner

Chen

Samala

Petrick

. Data drift in medical machine learning: Implications and potential remedies. Br J Radiol. 2023;96(1148):20220878.

81.

Kore

Bavil

Subasri

, et al. Empirical data drift detection experiments on real-world medical imaging data. Nat Commun. 2024;15(1):1887.

82.

Kreuzberger

Kühl

Hirschl

. Machine learning operations (MLOps): Overview, definition, and architecture. IEEE Access. 2023;11:31866-31879.

83.

Baltrusaitis

Ahuja

Morency

L-P

. Multimodal machine learning: A survey and taxonomy. IEEE Trans Pattern Anal Mach Intell. 2019;41(2):423-443.

84.

Chen

Hong

. AliFuse: Aligning and fusing multi-modal medical data for computer-aided diagnosis. arXiv [cs.CV] (2024) doi:10.48550/ARXIV.2401.01074.