AcuSim: a Synthetic Dataset for Cervicocranial acupuncture Points Localisation

This innovative study presents the development of AcuSim, a revolutionary synthetic dataset to automate the localization of acupuncture points in the cervicocranial region (head and neck). The research addresses a critical limitation in traditional Chinese medicine: dependence on experts to accurately locate acupuncture points, a process that is expensive, time-consuming, and subject to individual variation. The researchers created 504 photorealistic synthetic anatomical models, representing different body types, genders, hairstyles, and facial features. Using advanced computer graphics techniques in software such as Blender and DAZ Studio, they generated 63,936 high-resolution RGB-D images (1024×1024 pixels), each with precise annotations of 174 volumetric acupuncture points.

The methodology used domain randomization to capture human anatomical diversity, including variations in height, weight, and body fat proportions. The automatic annotation system used the traditional finger-cun method based on World Health Organization guidelines, positioning three-dimensional points as volumetric spheres rather than simple 2D points. For validation, they developed a convolutional neural network based on the VGG19 architecture with fully connected layers for multitask prediction of point coordinates and names. The model was trained for 3,480 epochs, converging after approximately 800 epochs for the training set and 1,250 epochs for validation.

The results demonstrated exceptional accuracy of 99.73% in point localization, with mean error of only 0.12-0.17 pixels in the final phases of training. When compared with annotations by traditional Chinese medicine experts, 92.86% of predictions were within the clinically acceptable error margin of 5 mm. The system incorporated automatic occlusion filters to remove points not visible due to hair or other obstructions, ensuring superior data quality. Two-way ANOVA statistical analysis confirmed that variations between anatomical models did not significantly affect accuracy, while different point locations showed expected variations due to anatomical complexity.

The dataset substantially surpasses previous work in scale and accuracy, offering more than 11 million annotations compared with the maximum of 28,000 in earlier studies. The clinical implications are significant: the system can facilitate acupuncturist training, reduce educational costs, and improve treatment standardization. Automation of point localization can make acupuncture more accessible and reliable, especially in regions with a shortage of specialists. Limitations include the need for additional validation with real images and expansion to other anatomical regions.

Future developments plan to incorporate transfer learning and domain adaptation techniques to improve generalization to real-world scenarios. This work represents a milestone in the digitalization of traditional Chinese medicine, demonstrating how artificial intelligence can preserve and modernize ancient practices, maintaining their therapeutic efficacy while improving accuracy and accessibility.