Variability between human experts and artificial intelligence in identification of anatomical structures by ultrasound in regional anaesthesia: a framework for evaluation of assistive artificial intelligence

James S Bowness; Robert Morse; Owen Lewis; James Lloyd; David Burckett-St Laurent; Boyne Bellew; Alan J R Macfarlane; Amit Pawa; Alasdair Taylor; J Alison Noble; Helen Higham

doi:10.1016/j.bja.2023.09.023

Variability between human experts and artificial intelligence in identification of anatomical structures by ultrasound in regional anaesthesia: a framework for evaluation of assistive artificial intelligence

Br J Anaesth. 2023 Oct 26;132(5):1063-1072. doi: 10.1016/j.bja.2023.09.023. Online ahead of print.

Authors

Affiliations

¹ Nuffield Department of Clinical Anaesthesia, University of Oxford, Oxford, UK; Department of Anaesthesia, Aneurin Bevan University Health Board, Newport, UK. Electronic address: james.bowness@jesus.ox.ac.uk.
² Intelligent Ultrasound, Cardiff, UK.
³ Department of Anaesthesia, Aneurin Bevan University Health Board, Newport, UK.
⁴ Department of Anaesthesia, Royal Cornwall Hospitals NHS Trust, Truro, UK.
⁵ Department of Surgery & Cancer, Imperial College London, London, UK; Department of Anaesthesia, Imperial College Healthcare NHS Trust, London, UK.
⁶ Department of Anaesthesia, NHS Greater Glasgow & Clyde, Glasgow, UK; School of Medicine, Dentistry & Nursing, University of Glasgow, Glasgow, UK.
⁷ Department of Anaesthesia, Guy's & St Thomas' NHS Foundation Trust, London, UK; Faculty of Life Sciences and Medicine, King's College London, London, UK.
⁸ Department of Anaesthesia, NHS Tayside, Dundee, UK.
⁹ Institute for Biomedical Engineering, University of Oxford, Oxford, UK.
¹⁰ Nuffield Department of Clinical Anaesthesia, University of Oxford, Oxford, UK; Department of Anaesthesia, Oxford University Hospitals NHS Foundation Trust, Oxford, UK.

Abstract

Background: ScanNav^TMAnatomy Peripheral Nerve Block (ScanNav™) is an artificial intelligence (AI)-based device that produces a colour overlay on real-time B-mode ultrasound to highlight key anatomical structures for regional anaesthesia. This study compares consistency of identification of sono-anatomical structures between expert ultrasonographers and ScanNav™.

Methods: Nineteen experts in ultrasound-guided regional anaesthesia (UGRA) annotated 100 structures in 30 ultrasound videos across six anatomical regions. These annotations were compared with each other to produce a quantitative assessment of the level of agreement amongst human experts. The AI colour overlay was then compared with all expert annotations. Differences in human-human and human-AI agreement are presented for each structure class (artery, muscle, nerve, fascia/serosal plane) and structure. Clinical context is provided through subjective assessment data from UGRA experts.

Results: For human-human and human-AI annotations, agreement was highest for arteries (mean Dice score 0.88/0.86), then muscles (0.80/0.77), and lowest for nerves (0.48/0.41). Wide discrepancy exists in consistency for different structures, both with human-human and human-AI comparisons; highest for sartorius muscle (0.91/0.92) and lowest for the radial nerve (0.21/0.27).

Conclusions: Human experts and the AI system both showed the same pattern of agreement in sono-anatomical structure identification. The clinical significance of the differences presented must be explored; however the perception that human expert opinion is uniform must be challenged. Elements of this assessment framework could be used for other devices to allow consistent evaluations that inform clinical training and practice. Anaesthetists should be actively engaged in the development and adoption of new AI technology.

Keywords: artificial intelligence; machine learning; medical devices; regional anaesthesia; sono-anatomy; ultrasonography; validation.