acuSimNet: Multi-View Self-Occlusion-Awared Visibility Learning for Cranio-Cervical Acupuncture Points
Organizations: College of Computing, Georgia Institute of Technology · University of Derby, United Kingdom
Abstract
The localization of acupuncture points (acupoints) in Traditional Chinese Medicine (TCM) presents unique challenges since they are defined by abstract principles rather than distinct anatomical landmarks. Existing approaches are typically constrained to single view or rely on indirect calculations of relative positions with respect to other landmarks, thereby overlooking human anatomical variations. Furthermore, acupoint visibility assessment, which determins whether points are occluded by human itself, hair, clothing, or other objects, has received limited attention in practical applications. Acupoint localization itself does not require 3D reconstruction, but inferring their occlusion relationships with anatomical surfaces does, which adds computational cost and limits real-time inference. In this work, we introduce acuSimNet, an efficient hierarchical multi-task learning architecture for multi-view, self-occlusion-aware visibility prediction of acupoints, achieving 99.97% accuracy on the validation set. Our approach also addresses the challenges of high-dimensional classification (174 acupoints in cervicocranial region), negative convergence issues for visible acupoints, and inter-task scheduling optimization, resulting in substantially accelerated convergence. We improved the training efficiency from the exisiting methods of 3000 epochs to achieve 99% validation accuracy, to our optimized framework achieving 90% accuracy in 39 epochs and 99% accuracy in 86 epochs. This architecture overcomes the limitations of existing methods, could enable practical applications in acupoints detection and visualization, advancing the automation of TCM.
Figures & tables
| Backbone NN | Feature Dimention | Params (M) | Batch Size | Batch/s | Total Loss | Top Vis Acc. (%) | Regression Loss (pixel) | |
| 16 | 5.64 | 8.49 | 94.39 | 0.89 | ||||
| ResNet 50 | 167.09 | 36 | 2.63 | 511.73 | 76.49 | 1.74 | ||
| 16 | 3.23 | 9.46 | 93.69 | 0.91 | ||||
| ResNet 101 | 224.20 | 36 | 1.90 | 514.74 | 76.58 | 1.71 | ||
| 16 | 3.26 | 8.47 | 94.47 | 0.89 | ||||
| ResNet He et al. (2016) | ResNet 152 | 2048 | 271.27 | 36 | 1.42 | 512.37 | 77.09 | 1.70 |
| Experimental Configuration | Module Combination | Top Val Vis Acc (%) | Min Val Coord Loss (Pixels) | Coord Loss (mm) | False Positive Rate ( ‰ ) | e 90 | e 95 |
|---|---|---|---|---|---|---|---|
| A0: Maintain benchmark configuration | A0 B0 C0 | 95.30 | 0.8759 | 0.4735 | 6.14 | 92 | 118 |
| A1: Random Assignment (Structure Preserved) | A1 B0 C0 | 94.89 | 0.7826 | 0.4230 | 7.04 | 98 | - |
| A2: Random Assignment (Balanced) | A2 B0 C0 | 94.02 | 0.8306 | 0.4490 | 6.72 | 98 | - |
| A3: Unified prediction for all 174 acupoints Sun et al. (2025) | - | 99.73 (3000 epoch) | 0.7560 | 0.4086 | - | 1305 | - |
| B0: Maintain benchmark configuration | A0 B0 C0 | 95.30 | 0.8759 | 0.4735 | 6.14 | 92 | 118 |
| B1: Hard binary thresholding (p_vis 0.5) | A0 B1 C0 | 85.25 | 0.9402 | 0.5082 | 19.61 | - | - |
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Methodology | Target Area | Subjects | Acupoints | Image | Total Annotations | Image Resolution | Multi-view | Visibility Info |
|---|---|---|---|---|---|---|---|---|
| Sun et al. (2025) | Cervicocranial | 504 | 174 | 63,936 | 11,126,952 | 1024*1024 | Yes | Yes |
| Ji & Zhou (2023) | Back | - | 4 | - | - | - | No | No |
| Masood & Qi (2022) | Human Hands | - | 9 | 1,200 | 28,800 | 848*480 | Yes | No |
| Sun et al. (2020) | Forearm | - | 2 | 719 | 2,876 | - | No | No |
| Zhang et al. (2023) | Facial | 4 | 43 | 654 | 28,122 | - | No | No |