Artificial intelligence to support ultrasound in the detection of developmental dysplasia of the hip

Abstract

Aims

Ultrasound is a highly sensitive method to detect developmental dysplasia of the hip (DDH). However, the cost of expert sonographers performing the tests is an important factor preventing wider adoption of this technique. AI has been used to enable non-specialists to undertake imaging in other disciplines, lowering costs and improving patient access. This scoping review aims to assess and map the available evidence pertaining to AI-assisted ultrasound for DDH detection.

Methods

The Association for Computing Machinery (ACM) Digital Library, EMBASE, OVID MEDLINE, PubMed, COCHRANE library, Cumulative Index to Nursing and Allied Health Literature (CINAHL), Australian New Zealand Clinical Trials Registry (ANZCTR), and Institute of Electrical and Electronics Engineers (IEEE) Xplore databases were searched. Additionally, we searched clinical trial registries from International Committee of Medical Journal Editors (ICMJE) and the World Health Organization (WHO), as well as publicly available commercial material from Exo.AI.

Results

A total of 600 records were identified, of which 41 discussed the use of AI-assisted ultrasound for DDH. The search of commercial sources revealed one FDA-approved device, which featured in two of the records scrutinized in the review. Common challenges across studies included limited access to large, high-quality imaging datasets, resulting in poor generalizability due to the small sample sizes used in model training and testing. There was poor transparency in the patient population selected, insufficient reporting on model inference time, and no cost-effectiveness analyses.

Conclusion

AI-based assistance for ultrasound detection of DDH shows promise, but the evidence base is incomplete. Future research should focus on standardizing development processes, improving transparency in study reporting, considering the varied perspectives of multidisciplinary teams, and prioritize comprehensive health economic evaluations. Addressing these challenges will be critical for successful development and integration of AI into this area of clinical practice.

Cite this article: Bone Jt Open 2026;7(2):223–233.

Introduction

Developmental dysplasia of the hip (DDH) represents a spectrum of disease, ranging from mild dysplasia to a complete dislocation of the hip.^1,2 This condition may affect one or both hips and usually manifests in early infancy. Delayed treatment of DDH requires more invasive treatment, and can lead to chronic hip pain, a permanent limp, and the early onset of osteoarthritis.^1-3 DDH underlies up to 9% of all primary hip arthroplasties and as many as 29% of those performed for individuals aged under 60 years.¹ Early detection and timely intervention are crucial in preventing these long-term sequalae.

Woodacre et al,⁴ in their cost analysis of DDH treatment, highlighted that late detection significantly escalates treatment complexity, leading to a sevenfold increase in short-term treatment costs compared to early detection and successful management with a Pavlik harness. The analysis revealed that successful management with a Pavlik harness incurs a cost of £601 per child, while surgical interventions, such as open or closed reductions, rise to £4,352 to £7,052 per child.

A systematic review and meta-analysis reported an increase in the incidence of DDH following the revision of the Newborn and Infant Physical Examination (NIPE) guidelines,⁵ rising from 6.5 to 9.4 per 1,000 live births.⁶

However, even in expert hands, clinical examinations (Barlow and Ortolani tests) may fail to detect dislocated and unstable hips.^2,7,8 A diagnostic accuracy study demonstrated that a screening examination using the Barlow and Ortolani tests in 1,000 hips of newborns will identify four dislocations, and result in five missed diagnoses and nine unnecessary follow-up diagnostic ultrasound scans. Alternatively, a screening approach using limited hip abduction would identify one dislocated hip, with eight missed diagnoses and 33 unnecessary scans.⁷

As a result, ultrasound has become the cornerstone for DDH screening, offering a systematic, reliable, and non-invasive method for assessing hip anatomy in infants. Many countries have adopted national ultrasound screening programmes for DDH, including Austria, Germany, and Poland, reflecting the recognition of its efficacy in early detection. However, in countries like the UK, the screening programmes remains selective, targeting infants with specific risk factors or abnormal clinical findings.^1,9

Nonetheless, the accuracy of the diagnosis depends heavily on the expertise of the operator.^10-12 Delivering a universal ultrasound screening programme would require skilled sonographers, which is both prohibitively expensive and resource-intensive.

The Austrian Society for Orthopaedics and Orthopaedic Surgery (ÖGO), in collaboration with the Austrian Society for Ultrasound in Medicine (ÖGUM), has established standardized training requirements for hip ultrasound diagnostics.¹³

To obtain an Austrian hip ultrasound diploma, candidates must complete a series of basic and advanced courses and perform a minimum of 200 supervised hip ultrasound examinations under the guidance of a qualified supervisor. Competency must be assessed and confirmed by an external course leader accredited by ÖGUM. To qualify as an external course leader, an individual must have successfully completed all required courses, be actively involved in teaching, and regularly perform hip ultrasound examinations, with a documented record of more than 3,000 hip scans.¹³ Additionally, to maintain certification and stay updated with advancements in hip ultrasound screening, practitioners are required to complete a refresher course every five years.¹³

In the UK there are no nationally established guidelines equivalent to those in Austria. However, the Guidelines for Professional Ultrasound Practice by the Society of Radiographers (SoR) and the British Medical Ultrasound Society (BMUS) highly recommended that all staff performing Ultrasound of hips attend Graf Ultrasound Hip course to ensure competency.¹⁴

The use of AI assistance for ultrasound image acquisition and interpretation is expanding rapidly. This technology has been applied to diagnostic radiology,^15,16 obstetrics and gynaecology,¹⁷ and regional anaesthesia.¹⁸ It has the potential to improve diagnostic accuracy, streamline workflows, and enhance patient care.^16-18 In DDH screening, AI could enable screening by a non-expert sonographer, improving cost-effectiveness, patient access, and the reliability of image acquisition/interpretation.

This scoping review aims to map the available evidence for AI-assisted ultrasound detection of DDH. We synthesize the current literature into a single coherent summary, identifying critical issues that remain in the evaluation and implementation process.

Methods

Scoping review protocol and question

The objectives, inclusion criteria, and methods were defined in a protocol published ahead of conducting this study.¹⁹ The methods follow guidance provided by the Joanna Briggs Institute (JBI) and PRISMA – Extension for Scoping Reviews (PRISMA-ScR).^20,21 The question we aimed to answer was ‘What is the evidence for the use of artificial intelligence to support ultrasound in the detection of developmental dysplasia of the hip?’

Inclusion criteria

The review included data from sources centred on detecting DDH in infants younger than six months, which is the age when ultrasound becomes a less effective tool for this purpose, owing to the hip development and morphology.^22,23 Data from animal and bench studies were excluded. Included sources relate to AI-supported ultrasound detection of DDH and the utility of these systems with regard to clinical practice.

To ensure a comprehensive review, we did not impose limitations on the type of AI involvement in the detection process. AI-supported ‘detection’ encompasses both image acquisition in any form and AI-driven scan interpretation for screening, diagnosis, or classification of hip morphology. Sources were required to relate to ultrasound scanning in the context of DDH.

Types of sources

All publicly available data, including academic publications, information from clinical and specialist societies, and commercial product literature, were included if published during or after 1980. This year was used as a benchmark because it marks the introduction of the primary ultrasound scanning technique used for DDH detection.^24,25 For any sources published in languages other than English, the lead investigator (RY) contacted the corresponding author to request a translated version.

No restrictions were placed on the type of ultrasound machine or probe used in the studies. All relevant studies were included regardless of the scanning technique employed to obtain the images.

The review used multiple databases in an attempt to identify all relevant literature: Association for Computing Machinery (ACM) Digital Library, EMBASE, OVID MEDLINE, PubMed, COCHRANE library, Cumulative Index to Nursing and Allied Health Literature (CINAHL), Australian New Zealand Clinical Trials Registry (ANZCTR), and Institute of Electrical and Electronics Engineers (IEEE) Xplore.

Other resources – including The International Committee of Medical Journal Editors (ICMJE) and World Health Organization (WHO) clinical trial registries – were searched to identify ongoing or completed studies in this field, capturing unpublished research. In addition, the websites of the British Society for Children’s Orthopaedic Surgery (BSCOS) and the Paediatric Orthopaedic Society of North America (POSNA) were reviewed.

As AI is an area of intense commercial research, other evidence and commercial data were also sought to minimize publication bias. Publicly available information from Exo Imaging (USA) was reviewed as it is, to our knowledge, the only commercial entity with a Food and Drug Administration (FDA)-approved medical device in this field. A full list of searched databases and their results is available in Table I.

Table I.

Full list of searched databases and their results.

Database/resource	Date of search	Results
Medline (Ovid)	12 July 2024	195
Embase (Ovid)	12 July 2024	289
CINAHL (EBSCOhost)	12 July 2024	109
PubMed	12 July 2024	40
The Cochrane Library – Clinical Trials (CENTRAL)	12 July 2024 (Issue 7 of 12, July 2024)	12
IEEE Xplore	12July 2024	17
ACM Digital Library	12 July 2024	SEARCH 1: 27 SEARCH 2: 82
ANZCTR	12 July 2024	1
ClinicalTrials.gov	12 Jul 2024	24
CTIS	12 July 2024	0
EU Clinical Trials Register	12 July 2024	SEARCH 1: 6 SEARCH 2: 1 SEARCH 3: 5 SEARCH 4: 1 SEARCH 5: 75 SEARCH 6: 9
ISRCTN Registry	12 July 2024	2
OMON	12 July 2024	SEARCH 1: 17 SEARCH 2: 10 SEARCH 3: 6
UMIN Clinical Trials Registry (UMIN-CTR) (English Version)	12 July 2024	SEARCH 1: 1 SEARCH 2: 1 SEARCH 3: 0 SEARCH 4: 1 SEARCH 5: 0 SEARCH 6: 1
WHO ICTRP	12 July 2024	3
BSCOS website	12 July 2024	0
POSNA website	12 July 2024	0
Exo.AI website	12 July 2024	1
	Total before deduplication	936
	Total after deduplication	600

ACM, Association for Computing Machinery; ANZCTR, Australian New Zealand Clinical Trials Registry; BSCOS, British Society of Children’s Orthopaedic Surgery; CTIS, Clinical Trials Information System; EU, European Union; IEEE, Institute of Electrical and Electronics Engineers; ISRCTN, International Standard Randomised Controlled Trial Number; OMON, Overview of Medical Research in the Netherlands; POSNA, Paediatric Orthopaedic Society of North America; WHO ICTRP, World Health Organization International Clinical Trials Registry Platform.

Search strategy

The search strategy was developed by the lead investigator (RY) and a clinical librarian (NH), then reviewed and modified by the senior team (DCP, JSB).

Search terms

The keywords outlined in Table II were employed to identify relevant manuscripts. Keywords were searched in titles, abstracts, author keywords, and subject headings, considering abbreviations and combinations of these words. To maximize the sensitivity of the search, the reference list of each paper was manually reviewed by RY to identify any relevant literature. Additionally, Google Scholar was used (also by RY) to locate pertinent citing literature.

Table II.

Search strategy. The main terms identified to be searched for in title, abstract, author keywords (if available), and subject headings. We also considered abbreviations and incorporated different combinations of these terms.

Artificial intelligence	And	Ultrasound	And	Developmental Dysplasia of the Hip
“Artificial intelligence”		“Ultrasound”		“DDH*”
“Computer vision”		“Sonography”		“Developmental Dysplasia of the Hip”
“Machine intelligence”				“Congenital Hip Dislocation”
“Deep learning”				“Hip Dysplasia”
“Machine learning”				“Hip Subluxation"
“Neural network”				“Hip Dislocation”
“Computer aided”
“Computer aided”
“AI*”
“Machine vision”

Data extraction

After removing duplicates, two investigators (RY, MK) independently screened titles and abstracts to determine their relevance to the scoping review. If a title was ambiguous, the source was included for further abstract review. In cases where the assessors disagreed, a third investigator (DCP) provided an independent adjudication.

The full texts of all relevant sources were scrutinized by the lead author (RY) to confirm their continued suitability for inclusion and data extraction. Data were systematically recorded using a standardized data extraction form, as detailed in the published protocol.¹⁹

To ensure a comprehensive and expert evaluation of the literature, a machine-learning (ML) engineer (LS) was enlisted to provide expertise in the assessment of computer science and engineering papers.

Results

A total of 41 sources were included for analysis. The search and screening process, along with the number of sources considered at each stage, is detailed in Figure 1.

Fig. 1.

PRISMA flowchart showing the number of sources considered at each stage of the search and screening process.

Data acquisition

In total, 16 (39.0%) papers reported patient inclusion criteria, such as suspected DDH on clinical examination with referral to ultrasound clinic, age below six months, and the presence of risk factors for DDH.^26-40

Common exclusion criteria included underlying congenital disorders (e.g. cerebral palsy, multiple joint contractures, and syndromes), patients with a history of hip infections or septic arthritis in the neonatal or infant stages, and significant muscle abnormalities.^32,37,41-43

In total, 12 (29.3%) of the studies published data on patient demographic characteristics, primarily age and sex.^{26-29,37,39,40,42,44-47}

Ten (24.4%) papers outlined their image inclusion criteria, such as scan quality, echogenicity, and the presence of key anatomical landmarks, in alignment with Graf’s standard plane checklist.^{27,28,30-32,40,43,48-50} Four studies employed multiple physicians, sonographers, or trained experts to label or review the dataset. In instances where there was labelling or interpretation disagreement, those sequences were typically excluded.^32,36,51,52

One paper stated that they rejected inadequate scans without reporting the inclusion/exclusion criteria.⁴¹

The most common scanning technique was the Graf method, reported in 26 papers (63.4%).^{29,31,34,36,37,40,42-61} The Graf method focuses on assessing hip morphology by detecting key anatomical landmarks in the standard Graf plane. These landmarks are used to measure the α and β angles, which are critical for classifying hip development and detecting dysplasia.

The hip sweep technique was used in five papers (12.2%).^{26,27,30,38,45} Three (7.3%) collected data using 3D probes,^26-28,30 and one used dynamic scanning detect instability while manipulating the hip.³⁰

The scanning technique employed, or the type of ultrasound machine and probes used for image acquisition, were not reported in the remaining papers. This prompts for clearer documentation in future studies to ensure data comparability and reproducibility.

Artificial intelligence technology

Of the 41 relevant studies, 30 (73.1%) used supervised ML-based approaches,^{26-37,41-44,48,49,51,53,57,58,62-65} two (4.9%) used pre-trained models as part of commercially available software,^38,45 four (9.8%) used basic unsupervised ML (such as curve fitting and K-means clustering),^39,40,46,59 and five (12.2%) exclusively used algorithmic methods (e.g. programmatic image manipulation/interpretation).^{47,50,60,61,66} Five papers used multiple models and model types.^{30,37,44,49,59}

As a result, there are a total of 38 non-algorithmic models in the 41 sources. A comprehensive summary of the model types used in each paper is provided in the Supplementary Material.

Of the studies using supervised ML approaches, convolutional neural networks (CNNs) were the approach of choice, comprising 28/38 models (73.7%) reported across all papers. These were used either for direct classification, or as part of a wider analytical pipeline (e.g. to segment an image prior to classification via other means). The types of models used over time are detailed in Figure 2. A total of 23/28 (82.1%) of the CNN models used a pre-existing architecture (e.g. U-Net,⁶⁷ Visual Geometry Group (VGG),⁶⁸ Mask R-CNN⁶⁹).

Fig. 2.

Model types used over time. CNN, convolutional neural network; FCN, fully convolutional network; RNN, recurrent neural network; SVM, support vector machine.

Models are frequently used within a larger analysis pipeline in two major configurations. A total of 20/38 (52.6%) were used in a configuration where the model was used to identify anatomical features from which diagnostic criteria were extracted, e.g. highlighting key points on the image from which α and β angles can be calculated.²⁹ In 15/38 (39.5%), the models are used to perform diagnostic classification either from ultrasound images or from preceding analysis, e.g. directly providing a classification of dysplastic/non-dysplastic/incorrect probe placement.⁴⁸

Model input formats, detailed in Figure 3 alongside respective output formats, included structured data, 2D images, and 3D Figures. In total, 33/38 (86.8%) operated directly on 2D ultrasound images. Only one study used the 3D volume directly (exploring this with three different models),³⁰ with the others splitting the 3D volume into individual 2D frames.

Fig. 3.

Overview of AI model inputs and corresponding output outcomes. The diagram illustrates the various input methods used in AI models (structured data, 2D ultrasound image, 3D ultrasound image) and maps these to the respective output outcomes. The numbers within the figure indicate the number of models in our scoping review employing each method, highlighting trends and preferences in AI application.

Reference standard (ground truth)

‘Ground truth’ is used in all four major stages of developing a supervised ML model: training, validation, testing (‘test’ data can also be referred to as ‘holdout’), and verification. This can vary depending on different models and training schemes (and some stages may be excluded entirely), and the terminology used for these stages, particularly the distinction between the terms ‘verification’ and ‘validation’, can differ between fields and often even between teams. We therefore provide a brief summary of how these stages are used in this paper.

Training: Ground truth forms the target for learning to which model predictions are compared and adjusted. The model parameters are dynamically updated to match predictions to the ground truth.

Validation: Throughout training, the model is tested against a set of unseen data, typically randomly sampled and from the same population as the training set, to monitor performance and to ensure the model is learning generalized patterns as opposed to overfitting and regurgitating learnt answers.

Test/holdout: After training, the model is tested on another set of unseen data, typically from the same population as the training set but sampled to ensure diversity across key subgroups (e.g. demographics). This is used to report how well the model has been trained and its typical accuracy.

Verification: The model is assessed for its clinical value, typically in the context of its real-world intended usage. This could, for example, include expert review of model predictions or comparison to gold-standard ground truth, e.g. clinical outcome.

Result interpretation: In supervised ML, each input data item (e.g. an ultrasound image) must be paired to corresponding reference standard (or ‘ground truth’), used to define the ideal model prediction output for that input. Seven of the reviewed studies (17.1%) did not use clinical experts to generate ground truth but used engineering students, non-clinical academics, or unspecified ‘trained users’ or similar.^{30,41,44,51,54,64,65} Four do not mention the training of their ground-truther(s).^26,32,62,63 This is particularly true for segmentation problems, as generating mask-type ground-truth is both time-consuming and not a task that clinicians are routinely trained for. However, even in diagnostic classification tasks, the source of ground-truth varied considerably; seven sources recruited multiple contributors to provide the reference standard,^{33,37,43,54-57} and 14 relied on the expertise of a single individual.^{29-31,34-36,40-42,44,51,53,64,65}

Where multiple contributors were involved, there was variation in their specialties and levels of experience. These contributors were identified as orthopaedic doctors, sonographers, radiologists, or computer science engineers. Their expertise was often labelled with broad terms such as ‘trained’, ‘experienced’, ‘expert’, ‘professional’, or ‘specialist’. Other studies used automated ground truth⁵⁸ or the clinical outcome as ground truth.⁴⁹

Data reporting

There was considerable variability in how datasets were reported. While 13 studies reported the number of unique patients and images used in each dataset, these were in the minority.^{26,29-31,34,41,43,54,57,58,63-65}

Moreover, it is often unclear how the training/validation/testing sets have been split, and whether the same patient may appear in more than one set (which can lead to inflated accuracy reports). Nine studies (30%) reported using a validation set (during training),^{32,33,43,44,48,53,62,64,65} and only one study made a distinction between the test and verification datasets (representing a fully held-out subset of training data, and a representative set from clinical practice),⁵⁶ if independent verification was performed at all.

In six studies (14.6%), the training data contained more than 1,000 images.^{29,32,44,54,55,65} Of the 35 with ≤ 1,000 training images, six reported handling this by using cross-validation as a training method, which mitigates the risk of overfitting.^{31,34,36,51,55,56}

Reporting on model performance, studies reported on a range of measures to quantify ‘success’, including percentage accuracy,^{26,31,32,36,42,48,55,57,58,62-65} area under the curve,^{29,36,48,51,57,64} Dice similarity coefficient,^{30,33,34,41,43,57} Hausdorff metric,^{30,33,41,43,49,57} intersection over union,³⁰ precision,^33,39 recall,³³ F-measure,³² Youden Index,³⁶ point error,⁵⁵ mean average error,^29,42,55 sensitivity,^{26,27,30-33,35,36,41,48,54,55,57,63} specificity,^{26,27,30-33,35,36,41,43,48,54,57} positive predictive value,^26,27,36 negative predictive value,^26,27,36 various measures of interobserver agreement,^{31,32,34,39-42,44} Euclidean distance,^{30,41,44,49,56} and Jaccard Index.⁴¹

For the successful integration of AI-assisted ultrasound screening and diagnosis of DDH, the technology must be time-efficient, cost-effective, and should require minimal training for inexperienced users. In our scoping review, we focused on evaluating the reporting of these critical parameters.

Time efficiency, model training, and data processing

There was considerable variability in how time parameters were reported across studies. Some papers made general claims, such as AI devices being faster than conventional methods^46,62 or operating in near-real-time.⁵² Others provided specific comparisons, including AI processing times versus manual expert performance,⁶⁴ average time costs for measurements across expertise levels,³⁷ and training and testing times of proposed algorithms against others.³⁶ Additionally, one study examined the average running times of algorithms for detecting hip landmarks.⁵⁶ Several papers detailed model training times^27,33,34,57 or processing times,^{28,41,42,47,58,60,65,66} often using graphics processing unit (GPU)-accelerated systems, which may not reflect the performance on standard hardware like laptops or ultrasound machines.

Operator training

None of the studies provided comprehensive details about the training or level of experience necessary for personnel to effectively operate the technology. Two offered vague statements, such as describing their models as suitable for lightly trained users.^46,51 While some envisioned their models as tools for users with minimal training,^26,27 most did not state the extent of training necessary. This lack of clarity poses a barrier to wider adoption and consistent performance across different settings.

Cost-effectiveness

None of the studies addressed the costs associated with developing these systems, such as those related to technology, manpower, implementation, and maintenance. As a result, the health economics of these prototypes remains uncertain.

Discussion

The scoping review highlights increasing interest and ongoing efforts to integrate AI-powered ultrasound screening for DDH in infants. The integration of AI has the potential to address critical limitations in the current diagnostic pathway, which includes the reliance on operator expertise, inconsistent detection rates, and resource-intensive processes. However, critical gaps remain in the evaluation and standardization of the technology, which prohibits wider implementation. Addressing these gaps could have considerable benefit for individuals and the broader population.

Key challenges

The limitations and challenges outlined are diverse, and highlight numerous critical issues in the development of AI-based medical technologies.

Studies exhibited notable variability in the patient selection criteria, image acquisition techniques, and metrics to evaluate the model performance. The variation in ‘success’ measure poses challenges in comparing different systems, preventing the ability to identify the most effective tools and set benchmarks for future developments.

Patient demographics are typically absent from reporting, and detailed descriptions of dysplastic versus normal hips are also frequently missing. Comprehensive data describing the study sample are imperative to ensure that the data on which models are developed and tested represent real-world clinical scenarios. As it stands, we may have models that perform well in controlled environments, but fail to generalize effectively to diverse, real-world patient populations, thereby limiting their applicability and reliability in clinical practice.

There is no established guideline on the optimal number of images required for AI to effectively perform deep learning on clinical images, nor is there a standardized test set on which to transparently build or validate tools. Furthermore, few studies distinguished between the training and testing datasets, raising concerns about potential overlap, which could lead to overfitting and biased outcomes, ultimately undermining the reliability of the findings. A standardized high-quality data resource for model training and testing would enable greater transparency in the development process.

A key claimed advantage for integrating AI is its potential to enable accurate scanning by operators without background training in DDH scanning, making it more feasible to adopt a national screening programme. This would be achieved either by automating the detection of key anatomical landmarks required for diagnosis, calculating α and β angles, or directly classifying images as dysplastic or non-dysplastic.

However, studies have largely overlooked the level of personnel training required to operate these models. For successful future implementation, these tools must be highly reproducible and user-friendly, particularly for clinicians with limited experience in hip scanning.

While studies focused on the technical performance and accuracy of AI models, none addressed the financial feasibility of developing, maintaining, and implementing these systems at scale. The perceived benefits of the AI tools suggest potential efficiencies (i.e. time costs), though the cost of the tools, the implementation cost, and the change in the diagnostic performance are essential factors to understand in upscaling and implementing these technologies. Cost-effectiveness is a key determinant in healthcare decision-making, particularly in public health interventions in resource-constrained environments.

Suggestions for future work

The development of standardized protocols should outline best practices for model development, testing, and evaluation, ensuring that results are consistent, comparable, and reliable across different studies. As AI adoption continues to expand, emerging frameworks have been developed to standardize and streamline this process. One example of such a framework is the Practice Integration of Artificial Intelligence (CPI-AI) framework, a five-stage approach to the clinical practice adoption of AI in the setting of Trauma & Orthopaedics, based on the IDEAL principles.⁷⁰

A key step will be the development of an open-access dataset of representative high-quality ultrasound scans. Given the substantial difficulties in obtaining sufficiently large datasets for effective training of modern ML models, the development of a large open-access dataset of ultrasound images, both 2D and 3D, would provide great value in this field. This shared resource would allow all researchers to use the same data for validation studies, facilitating a more standardized and transparent evaluation process.

Future studies should focus on correlating the effects of AI implementation on clinical outcomes through longitudinal studies. By tracking the impact of AI on patient outcomes over time, researchers can better understand the real-world benefits and limitations of these technologies, ensuring that they not only perform well in controlled settings, but also improve patient care in clinical practice.

Comprehensive evaluations of cost-effectiveness of existing DDH pathways are required to determine how the adoption of AI technology to alter the screening pathway may influence the cost-effectiveness. Without economic and logistical considerations, even highly accurate models will struggle to gain widespread adoption. These models should consider not only the initial costs of AI system development, but take a life-course approach, modelling the potential savings from earlier diagnosis and the reduced need for invasive treatments and long-term arthritis.

AI-driven ultrasound: commercialization and industry adoption

The lucrative commercial prospects in this field are fuelling increasing industry interest in developing these technologies and making them more accessible. Exo recently acquired MEDO AI (Canada) as part of its strategy to commercialize handheld ultrasound devices that are both affordable and user-friendly.⁷¹ MEDO AI, a FDA-cleared device, leverages its proprietary sweep technology alongside AI to deliver high-quality diagnostic ultrasound scans, specifically aimed at screening for DDH. This technology has shown promise in reducing variability between novice and expert users.⁴⁵ Additionally, a feasibility study piloted its use as a point-of-care screening tool in primary care, yielding encouraging results which could indicate suitability to consider broader implementation.³⁷

In conclusion, in the future, AI may hold the key to improving DDH detection by reducing operator-dependent variability and supporting the establishment of a national screening programme. However, several challenges and limitations must be addressed before this potential can be fully realized. A major obstacle is the lack of a structured framework and standardization in model developing, validation, and reporting. Addressing these barriers will be essential for the successful integration of AI into routine clinical practice.

Take home message

- AI-assisted ultrasound may enhance the detection of developmental dysplasia of the hip by improving reproducibility, reducing operator dependency, and supporting clinical decision-making, particularly within screening pathways; however, several challenges remain before widespread clinical implementation.

Data Availability

All data generated or analyzed during this study are included in the published article and/or in the supplementary material.