Real-time location of acupuncture points based on anatomical landmarks and pose estimation models

Acupuncture is one of the oldest medical practices in the world, with thousands of years of history. This therapy uses fine needles inserted at specific body points known as acupoints, located along meridians or energy channels. Although these points have traditionally been identified through palpation and anatomical knowledge passed down through generations, manual localization can be imprecise, especially when performed by individuals without adequate training. Accuracy in identifying these points is fundamental to treatment efficacy, creating a need for more precise and consistent localization methods.

This study, conducted by researchers at Pukyong National University in South Korea and published in November 2024, investigated how artificial intelligence can be applied to automate and improve accuracy in identifying acupuncture points. The researchers developed and compared two different approaches using computer vision: one using real-time anatomical landmark detection and another based on deep neural networks specialized in pose estimation. The goal was to create systems that could automatically identify acupuncture points on the face and hands with high precision, making this ancient practice more accessible and accurate for both practitioners and beginners.

The study methodology involved two distinct strategies. In the first approach, the researchers used Google's MediaPipe framework, a computer vision tool that detects anatomical landmarks in real time. This technique identified 468 reference points on the face and 21 points on the hands, using these landmarks as a basis to mathematically calculate the location of 38 specific acupuncture points through formulas based on traditional proportional measurements from Chinese medicine. The second approach used a convolutional neural network called YOLOv8-pose, specifically trained with a custom database containing nearly 6,000 images annotated by experts.

This model was developed to directly detect five specific acupuncture points on the arm and hand. To validate both methods, the researchers compared the locations predicted by the systems with those manually marked by experts in oriental medicine.

The results demonstrated that both approaches achieved high accuracy in locating acupuncture points. The anatomical landmark-based system was able to map multiple points on both the face and hands with mean localization error below 5 millimeters when compared with expert annotations. This method proved particularly effective for facial points, likely due to the larger number of anatomical landmarks available in this region. The deep neural network approach also demonstrated excellent performance, achieving 99% mean average precision in detecting the five specific points studied.

The model was able to locate points near the fingertips with greater precision than those located in the middle of the hand, suggesting that certain anatomical features are more easily identifiable by the system. Both methods were able to operate in real time, processing video images instantaneously.

For patients and acupuncture practitioners, these findings represent significant advances with important practical implications. For experienced practitioners, these systems can serve as support tools, increasing confidence and precision during treatments, especially in situations where manual localization can be more challenging. For students and beginners in acupuncture, the technology offers a valuable educational tool, allowing more efficient and standardized learning of point localization. The researchers developed a demonstration application that allows real-time visualization of acupuncture points through a webcam, demonstrating the practical potential of this technology.

For patients, this can mean more precise and consistent treatments, regardless of the therapist's experience level, potentially improving therapeutic outcomes.

The study has some important limitations that should be considered. The database used to train the neural network was relatively small and collected in a controlled environment, which may limit the system's ability to function adequately in real-world situations with greater variability. The images were captured against a white background and with relatively uniform poses, conditions that may not reflect the diversity encountered in daily clinical practice. In addition, quantitative evaluation focused mainly on points on the hands and arms, with less data available for validation of facial points.

The lack of large public databases with expert-annotated acupuncture points also represents a challenge for the future development of this technology. The researchers acknowledge that additional work is needed to expand these systems to other parts of the body, such as legs and trunk, and to test their efficacy in a wider variety of environmental conditions and patient types. Despite these limitations, this work represents an important step toward modernizing acupuncture through integration with artificial intelligence technologies, offering a solid foundation for future developments in this promising area.