Efficient adaptive distribution priors with multi-exposure perturbation for cephalometric landmark localization
Biomedical Signal Processing and Control · 2026-07 · 120 · 110120
DOI: 10.1016/j.bspc.2026.110120 ↗Kurz gefasst
Fachartikel in der Zeitschrift Biomedical Signal Processing and Control, 2026-07. Verfasst von Jingyu Gao und 5 weiteren Autor:innen. Die Arbeit nennt 58 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Accurate and efficient localization of cephalometric landmarks is critical for craniofacial analysis, morphological assessment, and orthodontic treatment planning. Deep learning based methods have achieved remarkable progress by leveraging manual annotations as supervision, yet the inherent ambiguity of annotations and the complex illumination conditions in lateral cephalograms pose persistent challenges for anatomical feature extraction. To address this, we propose a lightweight enhancement framework that integrates perturbation-based augmentation with adaptive landmark distribution modeling, which can be detached during inference. Our framework mainly includes two components: (i) a multi-shape exposure perturbation strategy that guides the network toward critical anatomical structures while enhancing robustness to diverse non-ideal imaging conditions; and (ii) a dynamic differentiable Gaussian mixture module that adaptively captures landmark distribution priors with minimal complexity, enabling efficient and flexible modeling of landmark annotation ambiguity. Experiments show that representing each landmark with at most two Gaussian components suffices to capture intrinsic ambiguity patterns, enabling a simple backbone (e.g., ResNet18) to achieve state-of-the-art localization accuracy (1.12 mm mean radial error). Further analysis indicates that landmark uncertainty predominantly appears as anisotropic unimodal variations or directionally constrained multimodal patterns aligned with local anatomical textures, providing interpretable insights into annotation variability.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Gao, J., Zeng, H., Xu, H., Xue, C., Mezei, J., & Chen, Y. (2026). Efficient adaptive distribution priors with multi-exposure perturbation for cephalometric landmark localization. Biomedical Signal Processing and Control, 120, 110120. https://doi.org/10.1016/j.bspc.2026.110120
Vancouver
Gao J, Zeng H, Xu H, Xue C, Mezei J, Chen Y. Efficient adaptive distribution priors with multi-exposure perturbation for cephalometric landmark localization. Biomedical Signal Processing and Control. 2026;120:110120. doi:10.1016/j.bspc.2026.110120
BibTeX
@article{Gao2026,
author = {Gao, Jingyu and Zeng, Heng and Xu, Hui and Xue, Chaoran and Mezei, Jozsef and Chen, Yuanyuan},
title = {Efficient adaptive distribution priors with multi-exposure perturbation for cephalometric landmark localization},
journal = {Biomedical Signal Processing and Control},
year = {2026},
volume = {120},
pages = {110120},
doi = {10.1016/j.bspc.2026.110120}
}
EUVIMED