Table 2.

Artificial intelligence method and its purpose.

Study citation, first author	Machine learning model	Purpose of ML	Benefits	Risks and limitations
Bowness et al. (18)	ScanNav Anatomy Peripheral Nerve Block system (Intelligent Ultrasound Ltd [IUL], Cardiff, UK) - deep convolutional neural networks based on the U-Net architecture	To identify anatomical regions	Identifying the specific anatomical structures, correct ultrasound view to anesthetists and standardization of clinical procedure	Model-related: Recognizes only anatomical structures on images
Alkhatib et al. (9)	Adaptive Median Binary Pattern approach Joint Adaptive Median Binary Pattern approach Three tracking algorithms: particles filter, Mean Shift and Kanade-Lucas-Tomasi (KLT) techniques	To imrove tracking procedure	Automatic detection and tracking of nerve structure, ROIs	Model-related: Nerve appearance might be similar to surroundings Difficulties in real-time tracking Risk of error after many iterations
Alkhatib et al. (10)	Deep learning methods: C-COT, ECO, CNT, MDNet SANet, SiameFC, CFNet, DCFNet, MCPF, HDT, HCFT CREST, DLT, PF-AMBP	Median and the sciatic nerves	Good performance Overcoming noise difficulties No need for pre-filtering images	Model-related: Nerve appearance might be similar to surroundings Failure of retracing the nerve
Gungor et al. (11)	Nerveblox, Smart Alfa Teknoloji San	Identify anatomical structures	A real-time interpretation of anatomic structures	Model-related: Low accuracy in pediatric/geriatric patients
Hetherington et al. (19)	SLIDE (Spine Level IDEntification) System based on deep convolutional neural network	transverse spinal ultrasound planes classification	Successful detection of vertebral regions at real-time speed	Model-related: Failure in identifying the difference between gap and bone images Real-time speed considerations
Huang et al. (20)	Deep learning model: U-Net	identify femoral nerve	Fast training and forecasting of the method Real-time segmentation	Study-related: Small sample size Limited number of images No data augmentation
Mwikirize et al. (12)	Deep learning (DL) based on convolution neural networks (CNNs)	Evaluate the new method	2D US data; deep convolution neural network usage detection data and intensity invariant feature maps	Model-related: Cannot systematically find the needle Relying on an expert sonographer
Oh et al. (13)	to detect the inter-spinous images	Localize L3/4	Confirm the sonographic images and structures. Time saving method Less possible complications	Study-related: Lack of a comparator arm Highly specific algorithm. The system is validated by current study population Absence of complex spinal anatomy, obesity, pediatric and geriatric patients. The risk of misinterpretation of fusion or reduced interspinous distance
Pesteie et al. (14)	CNN-based machine learning technique	Evaluate the convolutional network architecture	Few outliers in detecting the needles Performance is better compared with others	Model-related: Not running in real time
Smistad et al. (15)	Deep convolutional neural network – U-Net	Identify musculocutaneous, median, ulnar, and radial nerves and blood vessels	Accurate detection of blood vessels, median and ulnar nerves Real-time identification Direct comparison of 4 methods	Study-related: Small sample size Low precision and recall values Poor identification of musculocutaneous, radial nerves
Tran et al. (21)	MATLAB algorithm	Detect the LF depth	Helps to find the epidural space and measure the skin-to-LF depth An implementation in a wide range of ultrasound machines.	Model-related: Insignificant errors and failures to detect the LF mean Poor image quality might result in unsatisfactory outcomes
Bowness et al. (16)	ML/DL	Identification of the anatomical structures	Potential to support non-experts in training /clinical practice, as well as experts in teaching UGRA. It may promote the uptake and spread of UGRA.	Model-related: Experts reported an increase in risk
Bowness et al. (22)	DL (based on U-Net architecture)	Identification of the anatomical structures; highlighting anatomical structures of interest	High TP/TN and low FP/FN rates in key anatomical structure identification	Model-related: UGRA itself has not reduced the incidence of nerve injury; Study-related: remote expert were not present when the subjects were scanned.
Yang et al. (23)	DL		The developed model located the “interscalene brachial plexus” more accurately compared to nonexperts.
Liu et al. (17)	DL, SegNet Model	to optimize regional anesthesia puncture path;	DL ultrasound guided imaging for scapular nerve block in scapular fracture surgery was more efficient, significantly shortened the time of performing nerve block and reduced complications compared to traditional method.

ML, machine learning; PPV, positive predictive value; NPV, Negative predictive value; AUC-area under the curve; FP-false-positive; FN-false-negative.