Since the 1970s, WHO has been actively promoting acupuncture to the world, setting up acupuncture training institutions in many countries, supporting the creation of the World Federation of Acupuncture Societies (WFA), announcing many times the appropriate conditions for acupuncture treatment, encouraging global patients to choose acupuncture therapy, and striving to promote the internationalization and standardization of acupuncture (Lim, 2010). Acupuncture and massage are important parts of Chinese medicine, which is based on the theory of Chinese medicine's internal organs and meridians, with acupuncture and massage as the main treatment, and is an important means used to prevent and treat diseases and eliminate fatigue. By stimulating specific points on the human body, acupuncture and massage can regulate the body's qi and blood circulation and the balance of yin and yang, so as to achieve the purpose of treating diseases and relieving fatigue. Acupuncture and massage have been widely used in clinical practice and are gradually gaining attention and recognition worldwide. Acupoint localization serves as a pivotal component in the modalities of acupuncture and tuina therapy, where it's precision holds a direct correlation with the resultant therapeutic efficacy (Zheng, 2022). The imprecision in acupoint targeting may culminate in the ineffectuality of acupuncture interventions, potentially precipitating severe repercussions including, but not limited to, localized trauma, neural impairment, aneurysmal formations, ocular injuries, and incidences of needle fracture (Godson and Wardle, 2019). Pertaining to the human facial region, which is characterized by a plethora of acupoints each with distinct functionalities and in close proximity to one another, the task of exact acupoint identification is further complicated owing to the inter-individual anatomical variability, thereby presenting substantial challenges in accurate acupoint delineation (Lee et al., 2020). In clinical practice, the commonly used methods of taking acupoints can be divided into the method of anatomical marking on the body surface, the method of bone measurement and the method of body measurement (Zheng et al., 2005; Lin and Yi, 2019), all of which are artificial positioning of acupoints from the visual level, and the degree of accuracy is highly dependent on the professional skills and experience of physicians. Due to the complexity of human anatomy and individual differences, artificial positioning methods inevitably have certain subjectivity and errors, which affect the accuracy of acupoint positioning. Therefore, it is necessary to develop a new method to improve the objectivity, efficiency and accuracy of acupoint positioning to assist practitioners in prevention and treatment.

Artificial intelligence techniques may help to build intelligent, efficient and accurate models for point detection and localization. In recent years, AI techniques have been widely used in the fields of human behavior recognition, facial feature recognition and keypoint detection (Berretti et al., 2011; Maji et al., 2022; Pranavan et al., 2023; Zhang et al., 2023). However, as the available research deepens, the existing acupoint detection models have problems such as over-reliance on external devices such as infrared, insufficient mining of feature representations, and low accuracy and robustness of acupoint detection (Zhang et al., 2023). These issues seriously affect the generality of the models, and researchers are gradually recognizing the advantages of high-precision, lightweight detection models for better migration to mobile or embedded devices, and can achieve high-precision performance comparable to that of larger models.

Due to the problems of oversized models, low accuracy and insufficient robustness in existing studies, this study innovatively adopts the YOLOv8-ACU algorithm for facial acupoint recognition. Since acupoints do not have clear physical features or geometric shapes on the human surface, and locating facial acupoints by Chinese medicine practitioners is actually a process of locating facial regions or points, this study combines the facial acupoint detection task with face keypoint detection, and divides the task of facial acupoint detection into two aspects: identification of the type of acupoints and localization of the acupoints. Deep learning and computer vision techniques are used to recognize human facial features and acupoint location features, so as to achieve intelligent acupoint detection and assist doctors in acupoint treatment. The main contributions of this research are summarized as follows:

The algorithm used in this research has the advantages of high efficiency, multi-task processing, high accuracy, and robustness improvement, which better meets the needs of keypoint detection of facial acupoints. It possesses extremely important practical significance and clinical application potential. The specific implementation tools utilized in this algorithm are illustrated in Table 1.

In recent years, facial image processing, as a pivotal technology, has propelled technological advancements in domains such as facial recognition and image analysis. Deng et al. (2019) introduced a novel loss function, ArcFace, aimed at enhancing the capability of deep convolutional neural networks in feature learning. This innovation not only optimized the efficiency of feature extraction but also laid a foundational framework for the subsequent evolution of facial image processing technologies. Utilizing the efficient feature extraction methodology of ArcFace, Jin et al. (2022) applied deep learning models, originally developed for facial recognition tasks, to facial diagnostics. This cross-disciplinary application underscores the versatility and generalizability of deep learning models, offering new perspectives for medical image analysis and diagnostics. To further advance facial image processing, innovative techniques such as the generation of pseudo-depth information from traditional 2D RGB images using Generative Adversarial Networks (GANs) have been proposed (Jin et al., 2020). These advancements not only enhance the accuracy of facial recognition but also enable more complex facial recognition capabilities in resource-constrained settings. Concurrent with the rapid development of facial image processing, facial keypoint detection has emerged as a research focus, categorically divided into traditional facial keypoint detection methods and those based on deep learning. Among them, the traditional methods require manual design, feature extraction, and construction of subsequent classification or regression models, which can be further divided into parametric shape model-based methods and cascade shape regression-based methods. Among them, two representative algorithms based on parametric shape model approach are Active Shape Model (ASM) (Cootes et al., 1995), Active Appearance Models (AAM) (Cootes et al., 1998). ASM is a face keypoint detection algorithm proposed by Cootes et al. (1995), it is a point distribution model based algorithm, which firstly obtains the training set by manual calibration, and then obtains the shape model after training, and then abstracts the target object through the shape model, and then achieves the face shape matching through the keypoint matching in the testing stage. Cootes et al. (1998) further improved the ASM algorithm and obtained the AAM algorithm, which takes the texture features of the face region into account while adopting the shape constraints, and establishes the texture model while building the shape model, and combines the two models to obtain the active epistemic model. Dollár et al. (2010) put forward a classic work called the Cascaded Pose Regression (CPR), which is a method for matching the shape of the face. Cascaded Pose Regression model, which adopts a multi-stage cascade from coarse to fine learning idea to gradually improve the accuracy of keypoint locations through multiple iterations. This idea still influences many deep learning-based computer vision algorithms, such as Convolutional Pose Machines (CPM) (Wei et al., 2016) and Stacked Hourglass Network (Newell et al., 2016) in the field of human pose estimation, Cascaded R-CNN (Cai and Vasconcelos, 2018) in target detection and so on. Although the above traditional methods based on face keypoint detection can achieve better detection results to a certain extent, they inevitably have the limitations of complex data processing, poor anti-interference ability and low computational efficiency.

To overcome these problems, deep learning-based face keypoint detection methods have emerged in recent years and have substantially outperformed traditional methods in face keypoint detection tasks. Sun et al. (2013) first proposed and employed a deep convolutional neural network, DCNN (Deep convolutional network) (Ren et al., 2017), for face keypoint detection, proposing a three-stage coarse-to-fine cascade regression network, which achieved then-optimal results on several publicly available datasets. Currently, some scholars have applied deep learning-based keypoint detection techniques to face acupoint detection, and Zhang et al. (2023) constructed the face point dataset FAcupoint and proposed the FADbR facial point detection algorithm framework, which adequately extracts face features through the reconstruction task and reduces the model's dependence on labeled data. Berretti et al. (2011) conducted experiments on the BU-3DFE dataset to extract the local features of the face using the SIFT (Lowe, 2004) model for keypoint detection and used a multi-class SVM (Platt, 1998) for classification, and the experimental results showed that the method has a good performance in recognizing facial expressions. YOLO series algorithms, which use one-stage detection method, have many applications in keypoint detection by virtue of its simple network model and high accuracy rate. Maji et al. (2022) proposed the YOLO-pose algorithm for 2D multi-person pose keypoint estimation in images, achieving impressive results. On the MS COCO validation and test sets, the model achieved mAP@0.5 values of 90.2% and 90.3%, respectively. Pranavan et al. (2023) applied YOLOv7-pose to keypoint detection for assessing falls in elderly individuals, achieving an accuracy of 89.6% and precision of 91.2%, demonstrating good detection performance. However, the aforementioned models still have drawbacks such as slow speed, poor performance in detecting small faces, difficulties in training, and relatively low model accuracy. Therefore, this research proposes three improvements to the original YOLOv8-pose model, aiming to both reduce model size and enhance the accuracy of acupoint recognition.

ECA-net (Wang et al., 2020) is an improvement upon SE-net (Hu et al., 2018), introducing a non-dimensional reduction local cross-channel interaction strategy. This strategy effectively utilizes one-dimensional convolution and employs ResNets (He et al., 2016) and MobileNetV2 (Sandler et al., 2018) as backbones. The proposed ECA module has been extensively evaluated in image classification, object detection, and instance segmentation. Experimental results demonstrate that the module outperforms other modules in terms of performance while maintaining high efficiency. In this research, the ECA attention mechanism is employed to enhance the capture of acupoint features, with a focus on improving these features.

Slim-neck (Li et al., 2022) is used to balance model accuracy and speed. Originally applied to the SODA10M autonomous driving dataset, the Slim-neck detector shows significant improvements in speed and accuracy compared to the original detector. In this work, the Slim-neck module is substituted for the original neck module, enabling lightweight modeling while maintaining good detection performance.

GIoU (Rezatofighi et al., 2019) is an improvement upon IoU (Yu et al., 2016) and is utilized to optimize the non-overlapping regions. It has been verified that replacing IoU with GIoU significantly improves detection accuracy and is applicable to YOLO (Redmon et al., 2016) algorithms and the Faster R-CNN (Ren et al., 2015) series. In this research, GIoU is used as a loss function to enhance facial object detection capability and facial acupoint keypoint detection.

This section firstly introduces the improvement process of YOLOv8-pose, and then gives a brief introduction to the YOLOv8 keypoint detection algorithm in subsection 3.1, explaining the role and structure of each module. In subsection 3.2, it focuses on the improvement strategy of this research and gives a detailed description of each improved module in subsections 3.2.1, 3.2.2, and 3.2.3.

The YOLOv8-ACU critical point detection model and the improved experimental flow proposed in this research are shown in Figure 1, including the following ten steps.

Acupuncture points are a key concept in Chinese medicine, different points have different effects, and the precise identification of points is also the most important part of the acupuncture therapy process (Li et al., 2015). There are fewer studies applying keypoint detection models to acupoint recognition, Zhang et al. (2023) used the FADbR model to detect the acupoints on the face, and most of the acupoints are detected well, but the detection effect is poor in the sparse position of the acupoints, and the model only has the ability of keypoint detection, while YOLOv8-ACU has the ability of both the target detection and the keypoint detection, which is able to have a better interactivity, and the applicable scenarios are more abundant. In comparison with YOLO-pose (Maji et al., 2022) and YOLOv7-pose (Pranavan et al., 2023) YOLOv8-ACU can also be found to have better accuracy and lighter model size.

By comparing YOLOv8-ACU, YOLOv8-pose, YOLOv7-pose, YOLO-pose, it can be found that YOLOv8-ACU not only shows better results in accuracy, but also is lighter in model size, and has a good generalization ability in the external test set. It shows that the recognition effect and performance of the model can be effectively improved by adding the ECA attention mechanism, using the Slim-neck module and changing to the GIoU loss function.

However, the model also has some defects. There are fewer categories of acupoints, which do not cover all the acupoints on the face, and there is no classification experiment for data under various light and occlusion conditions, which may lead to errors in the detection and localization of acupoints under low light, dark light or occlusion conditions. In the future, we will further expand the categories of facial acupoints and improve the performance of detecting acupoints under different occlusions and different light conditions, and deploy them in embedded devices for real clinical practice.

To enhance the accuracy of facial acupoint recognition for better application in clinical practice, this study utilized the YOLOv8-pose model, which has been recently introduced with high detection and localization accuracy, as the base model. It was applied to the task of facial acupoint recognition. Building upon this base model, the study incorporated the ECA attention mechanism, introduced the Slim-neck module, and replaced its loss function. These enhancements further improved the precise recognition of acupoints while reducing the model's complexity and computational load. Compared to the base YOLOv8-pose model, YOLOv8-ACU demonstrated improvements in Precision (+0.2%), Recall (+2.0%), mAP@0.5 (+1.8%), and mAP@0.5–0.95 (+3.1%). Additionally, the Parameters, Model Size and GFLOPs decreased by 0.44M, 0.82MB, and 9.3% respectively. Comparing with other keypoint models, YOLOv8-ACU emerged as the most suitable model for acupoint recognition.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

ZY: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. PS: Formal analysis, Funding acquisition, Methodology, Writing – original draft, Writing – review & editing, Project administration, Resources. JL: Conceptualization, Data curation, Formal analysis, Funding acquisition, Writing – original draft. YW: Investigation, Methodology, Project administration, Supervision, Writing – original draft, Writing – review & editing. ZZ: Data curation, Writing – original draft. WQ: Data curation, Writing – original draft. BC: Data curation, Writing – original draft. YT: Resources, Software, Supervision, Validation, Writing – original draft, Writing – review & editing. AH: Conceptualization, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing.