Artificial Intelligence for Anatomical Structure Identification on Ultrasound in Regional Anaesthesia: A Scoping Review Protocol

Background

Ultrasound image guidance for regional anaesthesia was first described in 1989.¹ It has since become the main method to guide needle insertion and direct the delivery of local anaesthetic to inhibit the action of specific peripheral nerves.^{2, 3} Ultrasound-guided regional anaesthesia (UGRA) intends to use ultrasound to identify relevant sono-anatomy, to aid in directing the needle toward the desired target(s) and avoid unwanted trauma of important proximate structure(s).^4-6 It is also utilised to image the spine and surrounding tissues ahead of performing central neuraxial blockade (CNB; spinal and epidural).⁷ Fundamental to this is a thorough knowledge of the underlying anatomy, which informs the appropriate acquisition and interpretation of ultrasound images.^{6, 8, 9} However, anaesthetists’ knowledge of clinical anatomy is variable,¹⁰ ultrasound is known to be an operator-dependent imaging modality,⁹ and fallibility in human interpretation of medical images is recognised.¹¹

Artificial intelligence (AI) is becoming a mainstream methodology for medical image interpretation. It is a general term which refers to the ability of computers to perform tasks traditionally associated with human intelligence.¹² Computer vision – a field dealing with the ability of computers to interact with the visual world – often involves the use of machine learning (ML). This includes techniques which a enable computer programs to improve performance for a given task (learn) with increased exposure to training data. Deep learning (DL), a field within ML, uses artificial neural networks (interconnected mathematical functions) arranged in layers, allowing progressively more information to be extracted from the input data at each layer. Examples of AI-based ultrasound image analyses are emerging in clinical practice.^13,14

Deep learning for image analysis involves developing an algorithm (a rule-based problem-solving function executed by a computer) and exposing it to training data. These training data can be ‘labelled’ with the correct answer or features of interest (supervised ML) or presented alone (unsupervised ML). In supervised ML, as more training data is presented to the algorithm, the modifiable (learnable) parameters of the algorithm change as the algorithm ‘learns’ to make associations between the input data and training labels. When training is completed, and the fixed algorithm is deployed in its desired use case, it can be used to predict the appropriate label(s) for new image data. In unsupervised ML, the training data are presented without paired labels. The algorithm learns to identify clusters in the data (areas of similarity) that can then be assigned to a category (or label). Training stops when the identification of these clusters is considered suitable for the desired use case, and the algorithm can be deployed in the same manner as with supervised ML.

Recent publications present the case for the use of AI for ultrasound scanning in UGRA; peripheral nerve blocks (PNB) and CNB.^{15, 16} Potential benefits may include aiding the performance of ultrasound scanning in regional anaesthesia, training and learning these skills, and facilitating the use of UGRA outside the traditional environment of anaesthesia (e.g., the emergency department).^{15, 17} There are few systems with regulatory approval for clinical use and this is an emerging field – the first system gained regulatory approval for clinical use in PNBs in April 2021.¹⁸ Recently, academic reporting guidelines for studies involving medical AI devices have been published,^{19, 20} but these make little recommendation as to how studies should be structured. Also, there are no formal standards which guide consistency in regulatory approval of these devices. Thus, evaluations of such devices for the identification of anatomical structures on ultrasound in relation to UGRA are heterogeneous. Available data include proof-of-concept studies in pre-clinical, experimental work^{16, 21, 22} to early assessments of commercially-available systems.^{17, 23-25}

Furthermore, this is an interdisciplinary field with expertise drawn from computer science, engineering, and clinical medicine. Specialists may focus on different endpoints or use different approaches to assess the same endpoint. As such, opportunities for communication between clinicians, anatomists, computer scientists, engineers, mathematicians, and industry/manufacturers are likely to be missed. There is no single source of information which brings together different thought processes in this field to analyse early studies and assess areas of strength or weakness. Importantly, there are few recommendations on how this field can move forward, with a particular focus on the consistent and standardised evaluation of such new devices.

Thus, it is important to survey the literature and combine the findings into a coherent summary. This scoping review aims to identify and map the available evidence in this field. It is intended to inform future work by identifying areas where development is needed and proposing a formal, standard assessment framework which future evaluation processes can follow. Data included from adult subjects and from the point of introduction of ultrasound image guidance in regional anaesthesia (1989) will be considered.

A preliminary search (conducted on 4^th May 2022) was unable to identify scoping or systematic reviews registered with the Joanna Briggs Institute (JBI) Database of Systematic Reviews and Implementation Reports (https://csu-sfsu.primo.exlibrisgroup.com/discovery/fulldisplay?vid=01CALS_SFR:01CALS_SFR&tab=jsearch_slot&docid=alma991067665362702901&searchScope=EVERYTHING&context=L&lang=en), Cochrane Database of Systematic Reviews (https://www.cochranelibrary.com/cdsr/reviews), or the International Prospective Register of Systematic Reviews (PROSPERO; https://www.crd.york.ac.uk/prospero).

Methods

Discussion

The proposed scoping review aims to summarise the available evidence on the accuracy and utility of AI systems for ultrasound scanning in regional anaesthesia.

The amount and content of available data are likely to reflect that this is an emerging field, and this study may highlight a lack of consistency between studies. Therefore, we hope this review can serve as a concise summary of the state of the art and provide guidance for the direction and structure of future works.

Limitations

Despite the systematic search strategy summarised above, we anticipate challenges in the proposed study. We recognise the possibility of publication bias in the available work and that studies conducted as part of theses may not be found, despite the thorough search strategy. Given the pace of progress in this sector, further studies may be in progress but not in the public domain. Industry data may not be available, intentionally or otherwise, as companies may not always have a commercial incentive to share data.

As there are few systems available in clinical practice, and the first of these was only approved in 2021,²⁹ we anticipate that a large number of the included sources will be proof-of-concept studies with few scan images/subjects. It is likely that few, if any, will include patient outcome data.