Effects of convolutional neural network models on the segmentation of the interscalene brachial plexus in ultrasound imaging for radiomics

Introduction

Ultrasound (US)-guided regional anesthesia is an essential modality for perioperative analgesia and has been demonstrated to significantly influence clinical outcomes (1,2). Interscalene brachial plexus block (ISB) is widely applied in surgical procedures involving the shoulder joint and upper arm regions (3,4). The anatomical characteristics of the interscalene brachial plexus differ from those of the femoral nerve, axillary brachial plexus, and popliteal sciatic nerve. The interscalene brachial plexus is located between the anterior scalene muscle (ASM) and the middle scalene muscle (MSM), and lacks adjacent vascular landmarks that can be used to guide needle placement. In addition, substantial interindividual variability exists in the ultrasonographic appearance of the interscalene brachial plexus (5,6). These characteristics can pose challenges for anesthesiologists, particularly those with limited experience, as difficulties in image acquisition or anatomical identification may impair probe–needle coordination, potentially resulting in repeated needle insertions, increased patient discomfort, and unintended injury to adjacent critical structures. In severe cases, such factors may contribute to nerve injury (7). Previous studies have reported that ISB is among the most frequently implicated techniques in litigation related to nerve block procedures (8).

Although the application of artificial intelligence (AI) in medicine has progressed at a slower rate compared to fields such as computing and engineering, significant advances have been achieved in recent years, particularly in radiology and pathology (9,10). AI research within the field of anesthesiology has expanded gradually. Owing to considerable interindividual variability in patient conditions and the presence of multiple confounding factors in clinical practice, most AI applications in this field remain focused on model development and decision-support tools. Current AI research in anesthesiology focuses predominantly on four areas: (I) monitoring the depth of anesthesia; (II) image-guided procedural techniques; (III) prediction of anesthesia-related events and risks; and (IV) automated control of drug delivery systems (11).

Among these, real-time AI-assisted segmentation and interpretation of key anatomical structures in US imaging represent a major application in regional anesthesia. Computer vision-based techniques enable automated acquisition, processing, and analysis of US data to extract features such as shape, contour, and spatial characteristics (12). Deep convolutional neural network models utilize annotated US images, labeled via freehand segmentation or bounding boxes by experienced anesthesiologists, to facilitate automated recognition of anatomical structures during US-guided procedures (13). These AI-assisted techniques may facilitate more rapid acquisition of procedural skills among anesthesiologists with limited experience (14).

With continued advances in convolutional neural network architectures, various network models optimized for real-time instance segmentation have been developed, with progressive improvements in both accuracy and response time. The You Only Look At CoefficienTs (YOLACT) model is a deep learning-based framework specifically designed for real-time instance segmentation tasks, offering reduced processing times while maintaining image recognition performance (15). To date, the application of YOLACT to US image analysis in the context of ISB has not been reported. Accordingly, the primary objective of this study was to develop a YOLACT-based neural network model for US-guided ISB that reflects the scanning patterns used by anesthesiologists at the study site. The model was designed to enable automatic segmentation of predefined anatomical structures in US images. Model performance was further compared with that of a U-Net-based model to evaluate segmentation accuracy. The findings of this study are intended to provide preliminary methodological evidence to support the development of a larger, multicenter ultrasound image database for ISB. We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0588/rc).

Methods

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This prospective, single-center study was approved by the Institutional Review Board of Beijing Jishuitan Hospital (IRB 202009-07 and IRB 202104-14) and was registered in the Chinese Clinical Trials Registry (ChiCTR2200055591 and ChiCTR2200058470). US images of the supraclavicular-ISB were collected from September 2020 to March 2025. The inclusion criteria comprised patients aged 18–65 years, with a body mass index between 18 and 30 kg/m², who were scheduled for upper limb procedures under ISB. The exclusion criteria included a history of neck trauma or surgery, presence of cervical masses, abnormal cervical anatomical structures, vascular structures traversing the interscalene brachial plexus (such as the transverse cervical artery or dorsal scapular artery), known allergy to ultrasound coupling gel, or refusal to participate in the study. Written informed consent was provided by all participants prior to enrollment. US images used in this study have not been reported in any previous publications, nor are they currently under consideration for publication elsewhere. An overview of the study workflow is presented in Figure 1.

Figure 1 The flow chart of the study. DSC, Dice similarity coefficient; HD, Hausdorff distance; IoU, Intersection over Union.

Establishment of the US image database for ISB

US images were acquired using two US systems: SONIMAGE HS1 (Konica Minolta, Tokyo, Japan) and M9 (Mindray, Shenzhen, China). US videos and static images were obtained by anesthesiologists during clinical procedures and subsequently transferred to a computer via a data cable. The acquired US data were then uploaded to the study database through a wireless network. US images and videos were downloaded by anesthesiologists onto tablet devices equipped with image annotation software, and predefined anatomical structures were manually labeled on the US images. All participating anesthesiologists were attending physicians with more than 5 years of clinical experience and performed over 400 nerve block procedures annually. Prior to enrollment, all anesthesiologists completed a standardized training program and demonstrated proficiency in the acquisition and annotation specific to ISB.

US image acquisition and annotation protocol

All imaging was performed using high-frequency linear array probes (Mindray 12–4 MHz, Konica Minolta 11–3 MHz). During acquisition, the anesthesiologist stood on the same side as the scanning site, while the US system was positioned at the patient’s head. Patients were placed in the supine position with the head slightly rotated to the contralateral side of scanning, and both upper limbs positioned alongside the body. Scanning commenced at the supraclavicular fossa and advanced cranially along the course of the brachial plexus until the C7 vertebral level was identified. US video clips and static images were obtained at the supraclavicular, C8, and C7 levels. Both sides of the interscalene brachial plexus were scanned.

Anesthesiologists downloaded US images from the database onto tablet computers and used dedicated annotation software to label predefined anatomical structures on static US images, including the brachial plexus, ASM, MSM, and subclavian artery (SA). Annotation was performed using a freehand outlining method to trace the boundaries of each anatomical structure and generate closed contours. Distinct color coding was applied to each structure: brachial plexus (yellow), ASM (pink), MSM (green), and SA (red). To improve annotation accuracy, anesthesiologists reviewed 5-second US video clips recorded immediately before and after acquisition of the corresponding static images using the annotation software. If the initially selected static images were of insufficient quality, images were re-extracted from the US video data and the original images were discarded.

Development and validation of neural network models

All annotated US images included in the study database underwent systematic quality control procedures and were pixel-wise normalized prior to model training. Following quality control, the dataset was randomly divided into training, validation, and testing subsets at a ratio of 7:2:1 on a per-patient basis, ensuring that images from the same patient were assigned to the same subset. Deep learning-based segmentation was performed using the YOLACT and U-Net models. The number of extracted features was implicitly defined by the network channels, initializing at 64 channels and doubling per layer (up to 1,024) in U-Net, and standardized to a uniform 256 channels within the Feature Pyramid Network of YOLACT. The model was trained using a computer workstation with the following parameters: AMD Ryzen ThreadRipper Pro 7975WX processor (Advanced Micro Devices Inc., Santa Clara, CA, USA), 256 GB RAM, dual NVIDIA RTX 6000 Ada Generation GPUs (48 GB VRAM each; NVIDIA, Santa Clara, CA, USA), and a 4 TB NVMe solid-state drive. The models were trained using PyTorch in a Python environment managed by Miniconda. TensorRT and OpenCV were employed for inference acceleration and image processing, respectively.

Image annotation and quality control protocol

Two complementary approaches were used to assess the quality and reliability of the annotated US images: expert anesthesiologist review and 10-fold cross-validation.

Expert review protocol

For every batch of 200 annotated images, 20 images (10%) were randomly selected for quality assessment. These images were independently evaluated by three anesthesiologists who were not involved in the annotation process.

• An image was considered to have passed quality control if at least two reviewers agreed that the annotations were accurate.
• If two reviewers independently determined that the annotations were incorrect, the image was classified as failing quality control, and detailed reasons for rejection were documented to guide re-annotation.
• In cases of disagreement, the third anesthesiologist provided the final determination.

If any of the 20 sampled images failed quality control, the entire batch of 200 images was re-annotated, and a new random sample of 20 images was selected for reassessment using the same procedure. This iterative process continued until all 20 randomly selected images passed quality control, at which point the corresponding 200 annotated images were incorporated into the database.

10-fold cross-validation

Prior to model development, all US images in the database were randomly partitioned into 10 subsets. In each iteration, one subset was designated as the validation set, whereas the remaining nine subsets were used for model training. This process was repeated 10 times, ensuring that each subset served as the validation set once. Model performance metrics were averaged across the 10 iterations to provide a comprehensive assessment of model robustness and generalizability.

Model performance evaluation

The performance of the neural network models developed for ISB was comprehensively assessed using five quantitative and qualitative metrics: Intersection over Union (IoU), the proportion of images with IoU >0.5, Dice similarity coefficient (DSC), Hausdorff distance (HD), and anesthesiologist-based accuracy scoring of US image segmentation.

• IoU: IoU quantified the spatial overlap between the predicted segmentation boundary (model-generated target region) and the ground-truth boundary (clinician-annotated target region). It was calculated as the ratio of the intersection to the union of the two regions.
• Proportion of images with IoU >0.5: this metric represented the percentage of ultrasound images in which the IoU was between the predicted and annotated regions exceeded 0.5, a threshold commonly accepted as indicating adequate agreement.
• DSC: the DSC was employed to assess the similarity between predicted and annotated segmentation boundaries. Values ranged from 0 to 1, with higher values indicating greater overlap. Higher DSC values reflect greater concordance between predicted and true boundaries. The DSC is particularly sensitive to small anatomical structures and is widely used in medical image segmentation to evaluate model performance.
• HD: HD measured the maximum distance between predicted and annotated contours. To minimize the influence of outliers, the 95th percentile was calculated, providing a more robust assessment of boundary agreement.
• Accuracy scoring by anesthesiologists: a qualitative evaluation was conducted by two anesthesiologists, independent of image acquisition and quality control. US images from 15 randomly selected patients in the test dataset were reviewed by two anesthesiologists who were not involved in image acquisition or quality control. Machine-generated segmentation boundaries were compared with ground-truth annotations, and accuracy was rated using a 5-point Likert scale (1= very poor, 2= poor, 3= good, 4= very good, 5= excellent).

Statistical analysis

All statistical analyses were conducted using R software (version 4.5.0; R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were summarized as mean ± standard deviation (SD) for normally distributed data or as median with interquartile range (IQR) for non-normally distributed data. Categorical variables were expressed as frequencies and percentages. The Shapiro-Wilk test was used to assess data normality. Comparisons between the YOLACT and U-Net models were performed using the paired t-test for normally distributed continuous variables or the Wilcoxon signed-rank test for non-normally distributed continuous variables, as appropriate. Categorical variables were compared using the chi-squared (χ²) test. A two-sided P value <0.05 was considered to indicate statistical significance.

Results

A total of 1,100 participants were included in the study. Participant demographic characteristics and features of the US image dataset are presented in Table 1. Each US image was assigned exclusively to either the training, validation, or test dataset, with no overlap among subsets.

For segmentation of the interscalene brachial plexus, the YOLACT model achieved a median IoU of 0.6274 (0.5162, 0.7214) and a DSC of 0.7710 (0.6809, 0.8382), both of which were significantly higher than those obtained with the U-Net model [IoU, 0.5208 (0.4131, 0.6166); DSC, 0.6838 (0.5832, 0.7613)]. The HD for the YOLACT model was 6.7 (4.9, 7.5 mm), which was significantly lower than that observed for the U-Net model [6.8 (4.6, 7.6) mm] (P<0.001). These comparative results are presented in Figure 2.

Figure 2 Comparison of ISB ultrasound images segmentation performance between two neural network models. Black solid line: IoU of the YOLACT model for interscalene brachial plexus segmentation in the test set; Black dashed line: IoU of the U-Net model for interscalene brachial plexus segmentation in the test set; Red solid line: DSC of the YOLACT model for interscalene brachial plexus segmentation in the test set; Red dashed line: DSC of the U-Net model for interscalene brachial plexus segmentation in the test set; Gray solid line: HD of the YOLACT model for interscalene brachial plexus segmentation in the test set; Gray dashed line: HD of the U-Net model for interscalene brachial plexus segmentation in the test set. DSC, Dice similarity coefficient; HD, Hausdorff distance; IoU, Intersection over Union; ISB, interscalene brachial plexus block; N, interscalene brachial plexus.

For segmentation of the ASM, MSM, and SA, the YOLACT model demonstrated significantly higher IoU and DSC values and significantly lower HD values compared with the U-Net model (P<0.001) (Table 2). The proportion of US images with an IoU greater than 0.5 for the interscalene brachial plexus was significantly higher in the YOLACT model than in the U-Net model (P<0.001). However, anesthesiologist-based evaluations of the segmented US images did not differ significantly between the two models (P>0.05) (Table 3).

10-fold cross-validation of the YOLACT model yielded a mean IoU of 0.6033 and a mean DSC of 0.7526. The proportion of images with an IoU >0.5 was 77.3%. Representative segmentation outcomes for both neural network models are presented in Figure 3.

Figure 3 Ultrasound images as recognized by different convolutional neural network models. YOLACT model (A,C,E); U-Net model (B,D,F); Supraclavicular level (A,B); C8 level (C,D); C7 (E,F). Green: middle scalene muscle; Pink: anterior scalene muscle; Red: subclavian artery; Yellow: brachial plexus.

Discussion

The YOLACT algorithm was initially introduced by Bolya et al. in 2019 and has significantly advanced the field of instance segmentation by providing an effective framework optimized for real-time performance (16). The YOLACT framework consists of a backbone network, typically based on architectures such as ResNet, for feature extraction, with two parallel branches: one responsible for object classification and bounding box prediction, and the other for generating mask coefficients (17). This bifurcated design enables simultaneous object detection and instance segmentation, achieving efficient processing without compromising segmentation precision.

In recent years, the YOLACT algorithm has undergone further optimization, leading to the development of variants such as YOLACT-Edge and YOLACT++ (16,18). These models have demonstrated promising results in medical imaging contexts, including breast magnetic resonance imaging, retinal surgery, and spinal endoscopic procedures, where accurate localization and delineation of anatomical structures are critical (19-21).

To the authors’ knowledge, the present study represents the first application of the YOLACT algorithm to the segmentation of ISB-related structures in US imaging. Under the conditions of the current dataset and annotation protocol, the YOLACT-based model demonstrated accurate identification and segmentation of ISB-related anatomical structures. In addition, the model outperformed the U-Net architecture, as evidenced by higher IoU and DSC values, lower HD, and a greater proportion of images with an IoU >0.5.

Previous work has explored the use of U-Net-based models for automatic identification of the interscalene brachial plexus, although these efforts were generally limited by small sample sizes (22). In the present study, segmentation models based on both the YOLACT and U-Net architectures were developed using a substantially expanded dataset.

Both models demonstrated improved performance in terms of IoU and DSC compared with the earlier report, which may be attributed to several factors: (I) sufficient data volume is essential for the training and optimization of deep learning models (23). The present study utilized a large and continuously updated US image database, which likely enhanced the model’s capacity for feature representation and generalization. (II) Unlike prior studies in which annotations were limited to the brachial plexus alone (Delvaux et al.) (24), the current protocol incorporated bilateral annotation of the ASM, MSM, and SA, in addition to brachial plexus annotation. This broader anatomical context likely constrained the anatomical region of interest. This strategy reduced overfitting during nerve identification and improved segmentation accuracy. (III) Algorithmic advances in the YOLACT model, including incorporation of feature pyramid networks, may enhance instance segmentation performance for small targets, making the model more suitable for the complex and variable ultrasound appearance of ISB-related anatomical structures (25). (IV) Rigorous quality control represents a critical component in the development of AI neural network models for regional nerve block applications.

In contrast to previous studies that relied primarily on 10-fold cross-validation to ensure database integrity, the present study implemented a dual validation strategy that combined 10-fold cross-validation with independent anesthesiologist-based quality control (26,27). All US images underwent this comprehensive quality assessment prior to model training.

As US image characteristics may vary across different imaging platforms, two distinct US systems were used for image acquisition and model development in this study (28). This strategy aimed to reduce device-dependent variability and enhance the generalizability of AI segmentation model for ISB, facilitating broader use in both educational and clinical settings. Although the number of annotated US images in the present database remains relatively limited, standardized protocols for US image acquisition and annotation were implemented at the study institution, together with a rigorous quality control process. Compared with the study by Yang et al., which utilized a substantially larger training dataset comprising 11,392 images, the present model demonstrated comparable performance (e.g., IoU >50% in 81.6% of images) (29). This finding indicates that standardized acquisition procedures and stringent annotation quality control may partially compensate for smaller dataset size in the development of AI-based US segmentation models.

The results of the present study demonstrated that although the YOLACT algorithm achieved superior performance across objective image segmentation metrics, anesthesiologist-based evaluations did not reveal significant differences between the two algorithms. These findings align with those reported by James et al., who conducted external validations of neural network models and observed that expert human reviewers tended to assign higher segmentation quality scores in human-to-human comparisons than those assigned in human-to-AI assessments, albeit with reduced interrater variability (30,31). The clinical relevance of these observed differences has yet to be fully established.

This study has several limitations. First, it was conducted at a single center. Although the developed model demonstrated favorable performance in assisting US image interpretation, the processes of image acquisition, annotation, and quality control remain time consuming, and the total number of annotated images was relatively limited. In addition, all data were obtained from a single institution, and US scanning techniques may differ among anesthesiologists across institutions, which may restrict the generalizability of the model to other clinical settings. Second, anatomical variations of the interscalene brachial plexus, such as presence of vascular structures including the transverse cervical artery or dorsal scapular artery traversing the brachial plexus, can potentially complicate image interpretation and annotation, adversely affecting model performance. Individuals with such anatomical variations were excluded from the present study. Finally, the interscalene brachial plexus exhibits substantial interindividual morphological variability, in contrast to more anatomically consistent structures such as the paravertebral space, femoral nerve, or sciatic nerve. As a result, effective neural network modeling for ISB requires training on a larger volume of US images. Continued expansion and refinement of the US image database are therefore necessary to further enhance model performance.

Conclusions

US images representing characteristic ISBs, spanning from the supraclavicular region to the C7 vertebral level, were systematically acquired and annotated to include the SA, interscalene brachial plexus, ASM, and MSM. A YOLACT-based neural network model was developed to enable real-time segmentation and interpretation of these predefined anatomical structures.

The YOLACT model demonstrated superior performance compared to the U-Net model across multiple quantitative segmentation metrics. However, the current database remains limited in size, which constrains broader applicability in AI-driven regional nerve block applications. Continued expansion of the US image database is required to further enhance model accuracy and to establish a robust foundation for future clinical validation and implementation.

Acknowledgments

We sincerely appreciate the support of our colleagues from the Department of Orthopedics. We thank Bo Zhang, Shang Gao, Yuhang Yao for the technical support.

Footnote

Reporting Checklist: The authors have completed the CLEAR reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0588/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0588/dss

Funding: This study was supported by the Beijing Hospitals Authority Clinical Medicine Development of Special Funding (No. YGLX202318) and the Beijing JST Research Funding (No. YGQ-202104).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0588/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Institutional Review Board of Beijing Jishuitan Hospital (IRB 202009-07 and IRB 202104-14). A written informed consent was obtained from all participants.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Albrecht E, Chin KJ. Advances in regional anaesthesia and acute pain management: a narrative review. Anaesthesia 2020;75:e101-10. [Crossref] [PubMed]
2. Rawal N. Current issues in postoperative pain management. Eur J Anaesthesiol 2016;33:160-71. [Crossref] [PubMed]
3. Kim JH, Koh HJ, Kim DK, Lee HJ, Kwon KH, Lee KY, Kim YS. Interscalene brachial plexus bolus block versus patient-controlled interscalene indwelling catheter analgesia for the first 48 hours after arthroscopic rotator cuff repair. J Shoulder Elbow Surg 2018;27:1243-50. [Crossref] [PubMed]
4. Park SK, Sung MH, Suh HJ, Choi YS. Ultrasound Guided Low Approach Interscalene Brachial Plexus Block for Upper Limb Surgery. Korean J Pain 2016;29:18-22. [Crossref] [PubMed]
5. Fayed M, Khalil S, Patel N, Hussain A. Unexpected Anatomical Variation While Performing an Ultrasound-Guided Interscalene Block for Shoulder Surgery. Cureus 2022;14:e25079. [Crossref] [PubMed]
6. Feigl GC, Litz RJ, Marhofer P. Anatomy of the brachial plexus and its implications for daily clinical practice: regional anesthesia is applied anatomy. Reg Anesth Pain Med 2020;45:620-7. [Crossref] [PubMed]
7. Kim YD, Yu JY, Shim J, Heo HJ, Kim H. Risk of Encountering Dorsal Scapular and Long Thoracic Nerves during Ultrasound-guided Interscalene Brachial Plexus Block with Nerve Stimulator. Korean J Pain 2016;29:179-84. [Crossref] [PubMed]
8. Saba R, Brovman EY, Kang D, Greenberg P, Kaye AD, Urman RD. A Contemporary Medicolegal Analysis of Injury Related to Peripheral Nerve Blocks. Pain Physician 2019;22:389-400.
9. van Leeuwen KG, Schalekamp S, Rutten MJCM, van Ginneken B, de Rooij M. Artificial intelligence in radiology: 100 commercially available products and their scientific evidence. Eur Radiol 2021;31:3797-804. [Crossref] [PubMed]
10. Niazi MKK, Parwani AV, Gurcan MN. Digital pathology and artificial intelligence. Lancet Oncol 2019;20:e253-61. [Crossref] [PubMed]
11. Lopes S, Rocha G, Guimarães-Pereira L. Artificial intelligence and its clinical application in Anesthesiology: a systematic review. J Clin Monit Comput 2024;38:247-59. [Crossref] [PubMed]
12. Hashimoto DA, Witkowski E, Gao L, Meireles O, Rosman G. Artificial Intelligence in Anesthesiology: Current Techniques, Clinical Applications, and Limitations. Anesthesiology 2020;132:379-94. [Crossref] [PubMed]
13. Smistad E, Johansen KF, Iversen DH, Reinertsen I. Highlighting nerves and blood vessels for ultrasound-guided axillary nerve block procedures using neural networks. J Med Imaging (Bellingham) 2018;5:044004. [Crossref] [PubMed]
14. Cai N, Wang G, Xu L, Zhou Y, Chong H, Zhao Y, Wang J, Yan W, Zhang B, Liu N. Examining the impact perceptual learning artificial-intelligence-based on the incidence of paresthesia when performing the ultrasound-guided popliteal sciatic block: simulation-based randomized study. BMC Anesthesiol 2022;22:392. [Crossref] [PubMed]
15. Bolya D, Zhou C, Xiao F, Lee YJ. YOLACT: real-time instance segmentation. In: Proceedings of the IEEE International Conference on Computer Vision. 2019.
16. Bolya D, Zhou C, Xiao F, Lee YJ. YOLACT++ Better Real-Time Instance Segmentation. IEEE Trans Pattern Anal Mach Intell 2022;44:1108-21. [Crossref] [PubMed]
17. Lee S, Kim JE. Evaluating the Precision of Automatic Segmentation of Teeth, Gingiva and Facial Landmarks for 2D Digital Smile Design Using Real-Time Instance Segmentation Network. J Clin Med 2022;11:852. [Crossref] [PubMed]
18. Liu H, Soto RAR, Xiao F, Lee YJ. YolactEdge: real-time instance segmentation on the edge. IEEE International Conference on Robotics and Automation (ICRA). IEEE; 2021:9579-85.
19. Wang W, Wang Y. Deep Learning-Based Modified YOLACT Algorithm on Magnetic Resonance Imaging Images for Screening Common and Difficult Samples of Breast Cancer. Diagnostics (Basel) 2023;13:1582. [Crossref] [PubMed]
20. Nespolo RG, Yi D, Cole E, Wang D, Warren A, Leiderman YI. Feature Tracking and Segmentation in Real Time via Deep Learning in Vitreoretinal Surgery: A Platform for Artificial Intelligence-Mediated Surgical Guidance. Ophthalmol Retina 2023;7:236-42. [Crossref] [PubMed]
21. Bu J, Lei Y, Wang Y, Zhao J, Huang S, Liang J, Wang Z, Xu L, He B, Dong M, Liu G, Niu R, Ma C, Liu G. A Multi-Element Identification System Based on Deep Learning for the Visual Field of Percutaneous Endoscopic Spine Surgery. Indian J Orthop 2024;58:587-97. [Crossref] [PubMed]
22. Xi Y, Chong H, Zhou Y, Zhu F, Yao Y, Wang G. Convolutional neural network for brachial plexus segmentation at the interscalene level. BMC Anesthesiol 2024;24:17. [Crossref] [PubMed]
23. Hashimoto DA, Rosman G, Rus D, Meireles OR. Artificial Intelligence in Surgery: Promises and Perils. Ann Surg 2018;268:70-6. [Crossref] [PubMed]
24. Delvaux BV, Maupain O, Giral T, Bowness JS, Mercadal L. Evaluation of AI-based nerve segmentation on ultrasound: relevance of standard metrics in the clinical setting. Br J Anaesth 2025;134:1497-502. [Crossref] [PubMed]
25. Min K, Lee GH, Lee SW. Attentional feature pyramid network for small object detection. Neural Netw 2022;155:439-50. [Crossref] [PubMed]
26. Berggreen J, Johansson A, Jahr J, Möller S, Jansson T. Deep Learning on Ultrasound Images Visualizes the Femoral Nerve with Good Precision. Healthcare (Basel) 2023;11:184. [Crossref] [PubMed]
27. Festen RT, Schrier VJMM, Amadio PC. Automated Segmentation of the Median Nerve in the Carpal Tunnel using U-Net. Ultrasound Med Biol 2021;47:1964-9. [Crossref] [PubMed]
28. Ten Cate DF, Luime JJ, van der Ven M, Hazes JM, Kooiman K, de Jong N, Bosch JG. Very different performance of the power Doppler modalities of several ultrasound machines ascertained by a microvessel flow phantom. Arthritis Res Ther 2013;15:R162. [Crossref] [PubMed]
29. Yang XY, Wang LT, Li GD, Yu ZK, Li DL, Guan QL, Zhang QR, Guo T, Wang HL, Wang YW. Artificial intelligence using deep neural network learning for automatic location of the interscalene brachial plexus in ultrasound images. Eur J Anaesthesiol 2022;39:758-65. [Crossref] [PubMed]
30. Bowness JS, Burckett-St Laurent D, Hernandez N, Keane PA, Lobo C, Margetts S, Moka E, Pawa A, Rosenblatt M, Sleep N, Taylor A, Woodworth G, Vasalauskaite A, Noble JA, Higham H. Assistive artificial intelligence for ultrasound image interpretation in regional anaesthesia: an external validation study. Br J Anaesth 2023;130:217-25. [Crossref] [PubMed]
31. Bowness JS, Morse R, Lewis O, Lloyd J, Burckett-St Laurent D, Bellew B, Macfarlane AJR, Pawa A, Taylor A, Noble JA, Higham H. 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; Epub ahead of print. [Crossref]