Related Experiment Video
Updated: Dec 8, 2025
05:49
Author Spotlight: Advancing CBCT and Digital Dental Image Integration with AI-Assisted Digitization
Published on: February 23, 2024
1.2K
Automated feature detection in dental periapical radiographs by using deep learning
Hassan Aqeel Khan1, Muhammad Ali Haider2, Hassan Ali Ansari3
1Assistant Professor, College of Computer Science and Engineering, University of Jeddah, Kingdom of Saudi Arabia.
Oral Surgery, Oral Medicine, Oral Pathology and Oral Radiology
|September 20, 2020
Summary
Deep learning (DL) shows promise for automatically analyzing periapical radiographs (PRs), with U-Net architectures performing best. Further research with larger datasets and specialized models is recommended for improved accuracy in detecting dental conditions.
Area of Science:
- Dentistry
- Radiology
- Artificial Intelligence
Background:
- Periapical radiographs (PRs) are crucial for diagnosing dental conditions.
- Manual analysis of PRs can be time-consuming and subjective.
- Automated analysis using deep learning (DL) offers potential for improved efficiency and accuracy.
Purpose of the Study:
- To investigate the effectiveness of DL-based computer vision techniques for automated detection, segmentation, and quantification of common findings in PRs.
- To compare the performance of different DL architectures for analyzing PRs.
Main Methods:
- Three specialists labeled caries, alveolar bone recession, and interradicular radiolucencies on 206 digital PRs.
- The dataset was split into training/validation (176 PRs) and test (30 PRs) sets.
- Various DL architectures, including U-Net, Xnet, and SegNet, were trained and evaluated.
Main Results:
- The U-Net architecture and its variants outperformed other tested architectures.
- "U-Net+Densenet121" achieved the highest performance on the validation set (mIoU = 0.501).
- "U-Net" performed best on the test set (mIoU = 0.402), with interradicular radiolucencies being the most challenging to segment.
Conclusions:
- DL holds significant potential for the automated analysis of PRs.
- U-Net and its variants demonstrated superior performance among existing architectures.
- Further advancements require purpose-built architectures and larger, multicentric datasets.
