Related Experiment Video
Updated: Jul 2, 2026
10:23
Three-Dimensional Cephalometric Landmark Annotation Demonstration on Human Cone Beam Computed Tomography Scans
Published on: September 8, 2023
Synthetic data augmentation for improving performance in deep learning models for anatomical landmark localization on
Prathiksha Padmanabha1, Kasunika Guruge1, H M K K M B Herath1
1Industry 4.0 Convergence Bionics Engineering, Pukyong National University, Busan, Republic of Korea.
Frontiers in Bioengineering and Biotechnology
|July 1, 2026
Summary
Deep learning models for Traditional Eastern Medicine landmark localization improved with synthetic data. Mixed real and synthetic images enhanced model accuracy, achieving expert-level consistency and addressing data scarcity challenges in anatomical AI.
Area of Science:
- Medical Imaging and Artificial Intelligence
- Computational Anatomy
- Traditional Eastern Medicine
Background:
- Accurate anatomical landmark localization is crucial for Traditional Eastern Medicine (TEM) efficacy.
- Manual localization methods exhibit significant inter-practitioner variability (exceeding 5 mm).
- Deep learning offers automation potential but is hindered by limited real-world data.
Purpose of the Study:
- To investigate the utility of MetaHuman-rendered synthetic images for training deep learning models for TEM anatomical landmark localization.
- To assess the texture fidelity of synthetic images and their impact on model performance.
- To compare the performance of deep learning models trained on real-only versus mixed real-synthetic datasets.
Main Methods:
- Trained six landmark localization models (HRNet-W32/W48, YOLO26-pose variants) on distal upper limb anatomical landmarks.
- Utilized two training configurations: real-only dataset (2,963 images) and a mixed dataset (3,863 images) including synthetic data.
- Assessed synthetic image quality using Local Binary Pattern (LBP) and Gray-Level Co-occurrence Matrix (GLCM) features; evaluated models on 50 held-out real images.
Main Results:
- Synthetic image texture fidelity was confirmed, with LBP chi-squared distances falling within natural real-image variation.
- Mixed-data training significantly reduced mean localization error across all tested models compared to real-only training.
- Statistically significant improvements in localization accuracy were observed for YOLO26x-pose (17.8%), YOLO26s-pose (14.5%), and HRNet-W48 (6.3%) with mixed data.
Conclusions:
- Integrating MetaHuman-rendered synthetic images into deep learning pipelines can effectively augment limited real data for TEM applications.
- Mixed-data training approaches achieved sub-3.1 mm accuracy, approaching expert-level consistency in anatomical landmark localization.
- The study provides practical insights for leveraging synthetic data to overcome data scarcity in specialized medical AI development.
