Automated Segmentation of Digital Artifacts in Intraoral Photostimulable Phosphor Radiographs

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Dental Diagnostics

Background:

• Intraoral radiographs from photostimulable phosphor (PSP) plates often contain artifacts compromising diagnostic accuracy.
• Existing artifact taxonomies are extensive, but automated, pixel-level detection of multiple artifact types remains limited.
• This study evaluates a deep learning framework for fine-grained, multi-class PSP artifact segmentation.

Purpose of the Study:

• To establish a realistic baseline for automated, multi-class segmentation of artifacts in intraoral PSP radiographs.
• To systematically evaluate a deep learning-based framework for detecting and localizing various PSP artifact types at the pixel level.

Main Methods:

• A multi-center dataset of 1497 intraoral PSP radiographs was analyzed.
• Expert radiologists provided pixel-level annotations for 29 artifact classes and a background class.
• A 2D nnU-Net v2 architecture was used, with performance evaluated using Dice coefficient, IoU, Precision, and Recall.

Main Results:

• The deep learning model achieved a mean Dice score of 0.0952 on an independent test set, indicating task complexity.
• Performance varied significantly by artifact class, with better results for larger, distinct artifacts.
• Severe class imbalance, small object sizes, and limited training data constrained model generalization.

Conclusions:

• Multi-class pixel-level segmentation of PSP artifacts is inherently challenging due to data imbalance and artifact heterogeneity.
• The developed framework demonstrates feasibility for automated artifact localization, suitable for quality control or screening support.
• Future research requires imbalance-aware learning, hierarchical modeling, and data-centric approaches to improve performance.