High-resolution ultrasound images and videos targeting specific nerves were collected during clinical procedures. Because ultrasound images are collected from various devices, standardized protocols, such as consistent probe orientation, depth settings, and gain controls, help minimize artifacts and variations. Expert regional anesthesiologists manually delineated the nerve boundaries to create ground-truth bounding boxes or segmentation masks. To address potential differences in annotations among various experts, consensus labeling and multirater validation strategies have been implemented to enhance reliability and reduce subjectivity [15].

Raw ultrasound images often contain speckle noise, shadows, and variable contrast. The preprocessing pipeline aims to eliminate these artifacts by applying normalization, filtering, and resizing [16]. To compensate for data scarcity and variations in sonographic appearance, data augmentation strategies have been employed to expand the size and diversity of training datasets [17,18]. These include geometric transformations (such as rotation, scaling, and elastic deformations) and photometric adjustments (such as brightness, gamma correction, and color mapping) that stimulate typical noise patterns in ultrasound [19].

Modern ultrasound AI models use spatiotemporal architectures to integrate detailed anatomical information with the temporal dynamics of ultrasound imaging sequences. Traditional approaches involve using two-dimensional convolutional backbones (e.g., U-Net) to encode the spatial patterns of each ultrasound image, which are then decoded into segmentation masks (Fig. 2) [20]. These backbones can be integrated with temporal modules, such as convolutional Long Short-Term Memory or Recurrent Neural Network layers, which process sequences of ultrasound frames to capture temporal information across consecutive frames [21]. These models facilitated the extraction of both spatial anatomical features and temporal patterns of nerve structures, thereby enhancing their applicability in real-time clinical environments.

Recent advancements in ultrasound foundation models (USFMs) have demonstrated improved performance and generalizability [22-24]. The USFM framework is pretrained on millions of multiorgan ultrasound images to learn universal anatomical patterns, which can then be fine-tuned for nerve segmentation with minimal additional labeled data [24]. This approach reduces the need for extensive annotations while maintaining high accuracy. To deploy these high-capacity foundation models on resource-constrained ultrasound devices, knowledge distillation compresses them into lightweight student networks without compromising performance. The development of foundation models and associated technologies may enhance the efficiency and versatility of AI models in medical imaging [25].

Models are typically trained using Dice loss to optimize both overlap with ground-truth masks and pixel-level classification accuracy. Performance is then evaluated using metrics such as the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), which quantify segmentation overlap and pixel-wise accuracy. In addition to these quantitative measures, qualitative expert assessments are used to ensure anatomical plausibility [26]. For real-time clinical integration, inference speed and computational efficiency are optimized using lightweight model variants and quantization, facilitating deployment on portable or embedded ultrasound systems.

ScanNav^TM is an AI software approved in Europe and the United States (US) that provides real-time anatomical overlays to support PNB procedures. The system overlays color-segmented labels on B-mode ultrasound images to assist in the identification of anatomical structures relevant to the 11 common block regions [27]. Developed with a U-Net convolutional architecture trained on over 800,000 annotated images, ScanNav^TM aligns with Plan A block recommendations from major regional anesthesia societies [6,28].

Intelligent Ultrasound has also been used to develop NeedleTrainer, an augmented-reality-based training system designed to simulate needle insertion during ultrasound-guided procedures. By projecting a virtual needle onto live ultrasound images without actual skin penetration, NeedleTrainer allows trainees to practice needle-probe coordination and improve hand–eye skills in safe and repeatable environments [29]. In October 2024, GE HealthCare acquired the clinical AI division of Intelligent Ultrasound, including the ScanNav^TM product suite. Following this acquisition, the standalone version of ScanNav^TM was withdrawn from commercial distribution, and its AI functionalities were integrated into GE’s proprietary ultrasound platforms.

Nerveblox^® is a real-time AI guidance software developed by SmartAlpha Teknoloji and approved for use in Europe, distributed globally by PAJUNK Medical. It assists clinicians during ultrasound-guided PNBs by providing dynamic, color-coded overlays on B-mode images to identify nerves, vessels, muscles, and pleura using deep-learning-based multistructure segmentation [30].

The system currently supports 12 block types: interscalene, supraclavicular, infraclavicular, axillary, superficial cervical plexus, PECS I, PECS II, transversus abdominis plane (TAP), rectus sheath, femoral, saphenous (adductor canal), popliteal sciatic, and erector spinae plane (ESP) blocks. A proprietary scan-success gauge provides real-time feedback on probe positioning, and the system is compatible with most ultrasound machines via high-definition multimedia interface (HDMI) or digital visual interface (DVI) output. In addition to its clinical use, Nerveblox^® offers educational modules with anatomical references and probe guidance, making it an ideal tool for training purposes.

Although not yet Food and Drug Administration (FDA)-approved, Nerveblox^® is used in the US as a training tool and is widely adopted in clinical settings across Europe and other Conformité Européenne (CE)-recognized regions. Nerveblox^® was recently evaluated in a prospective clinical study in the US (NCT06878911) involving 40 adults. This study assessed anatomical structure recognition across 12 PNB sites, providing key evidence to support FDA clearance and demonstrating its potential applicability across diverse body types and clinical settings.

cNerve^TM is a proprietary assistive AI feature integrated into GE Healthcare’s Venue^TM family of ultrasound systems, designed to enhance anatomical recognition during PNBs. Utilizing a deep-learning-based segmentation approach, reportedly based on a U-Net architecture, cNerve provides real-time, yellow-colored overlays that highlight peripheral nerve structures, including the brachial plexus (supraclavicular and interscalene) and the femoral and popliteal sciatic nerves, to support accurate visualization and procedural planning during UGRA.

The system was trained on a large heterogeneous dataset comprising ultrasound images from diverse geographic populations (for example, the US, Israel, Japan), to enhance generalizability across anatomical variations. According to internal retrospective validation studies reported by the manufacturer, cNerve demonstrated > 99% accuracy in peripheral nerve identification during both live scanning and retrospective image review. The tool is seamlessly integrated into the scanning workflow, allowing region-specific guidance to be activated with a single button and enabling dynamic, real-time tracking of target structures without disrupting the procedural flow. Although initial reports suggested promising performance, prospective trials have only recently begun to assess their clinical impact [31]. Nevertheless, details regarding model architecture, dataset composition, and external validation metrics remain largely undisclosed, underscoring the need for transparent and independent studies to evaluate the generalizability and real-world effectiveness. Following GE HealthCare’s acquisition of ScanNav^TM, future updates to cNerve may incorporate ScanNav^TM’s advanced segmentation capabilities and broader anatomical coverage, potentially expanding its clinical utility and alignment with expert practice.

Smart Nerve is Mindray’s embedded AI tool that uses segmentation to display real-time, color-coded overlays of nerves and surrounding anatomy. The system currently supports two main regions for brachial plexus block: the interscalene and supraclavicular regions.

Similar to most commercially available AI tools, Smart Nerve uses pixel-level segmentation, which may enhance anatomical precision for procedures requiring detailed structural differentiation. It is primarily marketed in Asian and European healthcare settings. Although Smart Nerve offers real-time anatomical segmentation for brachial plexus blocks, peer-reviewed clinical evaluations remain limited. Further studies are warranted to validate diagnostic accuracy, usability, and impact on procedural outcomes.

NerveTrack^TM, developed with Intel, is Samsung Medison’s embedded AI tool that uses object detection to dynamically track nerves. Unlike segmentation-based systems such as ScanNav^TM or cNerve^TM, NerveTrack^TM uses an object detection paradigm, applying bounding boxes to track nerve structures in real time. Initially developed for median and ulnar nerve tracking in the forearm, its capabilities have since expanded to include identifying neural structures in the interscalene brachial plexus and terminal nerves in the axillary region.

Building on Intel’s Edge AI platform, Geti software, NerveTrack provide optimized performance for embedded computing environments. NerveTrack^TM’s object detection approach offers an alternative paradigm that emphasizes rapid target localization, which may be advantageous in mobile or fast-paced procedural settings. Although promising, its performance across a broader range of block types and clinical conditions requires further investigation to confirm generalizability and clinical utility.

A key limitation of this field is the lack of standardized evaluation frameworks that integrate technical performance metrics and clinical relevance. Studies often employ segmentation metrics such as the DSC, IoU, sensitivity, and precision. However, these metrics are defined inconsistently and rarely contextualized within real-world clinical decision-making [15]. Although Dice scores provide a quantitative measure of overlap, their interpretation requires caution, because high scores do not always indicate clinically meaningful segmentation, particularly in small or variably shaped nerve structures. Using task-specific thresholds that balance false positives and false negatives may improve the clinical applicability of such metrics [39]. Furthermore, heterogeneity in the target anatomy, ultrasound views, and block types complicates comparisons across models. Meaningful evaluation of AI tools in UGRA requires both objective accuracy and expert-informed clinical usability. Developing consensus-driven, clinically meaningful composite metrics that potentially combine technical accuracy, procedural efficiency, and patient-centered outcomes is essential. In parallel, standardized reporting checklists for clinical AI studies can facilitate transparency, reproducibility, and meaningful comparisons across models and studies [40].

Nevertheless, the generalizability of the AI model remains a critical limitation. Many systems are trained on datasets composed predominantly of healthy volunteers, leading to an underrepresentation of anatomically diverse populations, such as persons with obesity, pediatric or geriatric populations, or individuals with comorbidities [32]. In practice, this may cause models to perform reliably in healthy adults but fail to recognize target structures in patients with altered anatomy, such as difficulty in identifying nerves in patients with obesity or distorted fascial planes in trauma cases. Differences in ultrasound devices, imaging presets, and acquisition protocols further introduce biases related to specific devices that may reduce the model’s robustness in various clinical environments. Most current models have been developed using data from a single center, raising concerns regarding reproducibility in different clinical environments. In addition, accurate annotation of peripheral nerve structures is inherently challenging, with substantial interobserver variability that can propagate bias into training labels. Although open datasets have driven significant progress in perioperative medicine, including vital sign monitoring and intraoperative event prediction, these resources remain scarce in UGRA [41]. The lack of publicly available and well-annotated datasets specific to PNBs impedes the development, benchmarking, and external validation of AI tools for UGRA, where task-specific imaging and complex anatomical structures require domain-specific data [42].

Most commercially available AI tools are designed for specific anatomical regions or block types [43]. For example, systems such as cNerve^TM, Smart Nerve, and NerveTrack^TM primarily focus on PNBs in the upper or lower extremities. Although these systems have demonstrated reliable performance within their intended scope, their narrow anatomical focus limits their applicability in comprehensive UGRA practice. Blocks that are technically more demanding or performed less frequently remain largely unsupported. This gap restricts its adoption in institutions where anesthesiologists routinely perform a wide spectrum of blocks across diverse surgical populations. Recent advances in USFMs have aimed to address this limitation by pretraining on large-scale, multi-organ, multicenter, and multi-device databases that exceed two million ultrasound images, enabling broader task adaptability with minimal additional annotation [24]. However, despite these advances, the practical implementation of UGRA remains challenging because of computational demands, the need for task-specific fine-tuning, and unresolved issues regarding real-time integration. Broader anatomical coverage and support for multiple block types are essential to maximize clinical utility and flexibility.

Although AI-guided systems show promise in improving ultrasound image acquisition and anatomical recognition, there is limited evidence linking their use with meaningful patient outcomes [13,37]. Most existing studies have been conducted in simulation environments or single-center observational contexts, often without direct evaluation of clinically relevant endpoints, such as block success rates, complication rates, procedure time reduction, or improvements in diagnostic accuracy. Even when technical performance improves, it remains unclear whether it translates into measurable benefits for patients, such as less postoperative pain, reduced opioid use, and shorter hospital stays. Furthermore, multicenter external validation and prospective trials assessing the generalizability across diverse patient populations and practice settings are lacking. One service evaluation reported that the introduction of an AI guidance tool increased the proportion of patients with trauma receiving UGRA from 34.3% to 48.2%, indicating improved procedural uptake [37]. However, outcomes such as procedural efficiency, block success rates, and patient satisfaction were comparable to those in standard practice. Importantly, no studies to date have assessed long-term endpoints, such as reduced complications, functional recovery, or cost-effectiveness, which are critical for establishing clinical value. Moreover, no randomized controlled trials have demonstrated the clinical superiority of AI-assisted UGRA, underscoring the need for rigorous, prospective studies across multiple centers to confirm its diagnostic and therapeutic benefits.

The adoption of AI technology entails significant financial and regulatory hurdles. The initial costs include hardware acquisition, software licensing, system integration, clinician training, and long-term maintenance. In addition, the regulatory approval processes for AI medical devices vary across regions and can delay their clinical deployment. Although some AI tools in UGRA have achieved regulatory clearance—ScanNav^TM, for example, received FDA approval in 2022. Its clinical AI division was later acquired by GE Healthcare, and the product has since been withdrawn from commercial sale [44]. This underscores the volatility of the AI medical device market and the uncertainty surrounding long-term product availability and support. Moreover, manufacturers are often not obligated to publish performance data or make training datasets available, a practice that limits transparency, external validation, and reproducibility. These barriers are consistent with the broader challenges in surgical AI implementation, in which regulatory, economic, and translational hurdles often delay clinical deployment [45].

Seamless integration of AI tools into UGRA workflows is essential for maximizing clinical benefits without introducing unintended risks. Poorly designed systems can disrupt procedural flow, distract attention from critical anatomical cues, or increase the cognitive load by adding unnecessary visual information. Although AI assistance has been shown to improve block-view acquisition rates and anatomical identification accuracy among non-expert users [35], excessive reliance on such tools may hinder the development of independent ultrasound interpretation skills and reduce procedural autonomy. Therefore, effective integration requires clinician-centered design principles that prioritize usability, intuitive interfaces, and transparency regarding algorithmic limitations, risk of overreliance, and potential errors [42,46]. Features such as customizable overlays, user override options, and clear visualization layouts are critical for ensuring that AI enhances, rather than obstructs, procedural performance. Furthermore, rigorously evaluating AI within realistic clinical workflows rather than isolated simulation environments is necessary to confirm that AI systems support decision-making, improve safety, and facilitate skill development without diminishing core clinical competencies.