Ismail, Amelia Ritahani and Azlan, Faris Farhan and Noormaizan, Khairul Akmal and Afiqa, Nurul and Nisa, Syed Qamrun and Ghazali, Ahmad Badruddin and Pranolo, Andri and Saifullah, Shoffan (2025) Privacy-preserving U-Net variants with pseudo-labeling for radiolucent lesion segmentation in dental CBCT. International Journal of Advances in Intelligent Informatics, 11 (2). pp. 275-291. ISSN 2442-6571 E-ISSN 2548-3161
|
PDF
- Published Version
Download (1MB) | Preview |
Abstract
Accurate segmentation of radiolucent lesions in dental Cone-Beam Computed Tomography (CBCT) is vital for enhancing diagnostic reliability and reducing the burden on clinicians. This study proposes a privacy-preserving segmentation framework leveraging multiple U-Net variants—U-Net, DoubleU-Net, U2-Net, and Spatial Attention U-Net (SA-UNet)—to address challenges posed by limited labeled data and patient confidentiality concerns. To safeguard sensitive information, Differential Privacy Stochastic Gradient Descent (DP-SGD) is integrated using TensorFlow-Privacy, achieving a privacy budget of ε ≈ 1.5 with minimal performance degradation. Among the evaluated architectures, U2-Net demonstrates superior segmentation performance with a Dice coefficient of 0.833 and an Intersection over Union (IoU) of 0.881, showing less than 2% reduction under privacy constraints. To mitigate data annotation scarcity, a pseudo-labeling approach is implemented within an MLOps pipeline, enabling semi-supervised learning from unlabeled CBCT images. Over three iterative refinements, the pseudo-labeling strategy reduces validation loss by 14.4% and improves Dice score by 2.6%, demonstrating its effectiveness. Additionally, comparative evaluations reveal that SA-UNet offers competitive accuracy with faster inference time (22 ms per slice), making it suitable for low-resource deployments. The proposed approach presents a scalable and privacy-compliant framework for radiolucent lesion segmentation, supporting clinical decision-making in real-world dental imaging scenarios.
Actions (login required)
 |
View Item |
Download Statistics
Download Statistics