Using Octuplet Siamese Network For Osteoporosis Analysis On Dental Panoramic Radiographs.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Dental Panoramic Radiography (DPR) offers a cost-effective method for assessing bone density changes via trabecular bone structure analysis.
• Challenges in DPR analysis include dense bone structure overlap and limited labeled data, hindering accurate osteoporosis prediction.
• Existing methods struggle with the complexity of mapping DPR image features to osteoporosis conditions.

Purpose of the Study:

• To develop a deep learning model for accurate osteoporosis condition prediction using multiple DPR patches.
• To address the challenges of feature extraction and data scarcity in analyzing trabecular bone structure from DPR images.

Main Methods:

• Proposed a deep Octuplet Siamese Network (OSN) to learn and fuse discriminative features from eight different DPR patch locations.
• Employed a two-stage fine-tuning strategy, utilizing an augmented texture analysis dataset to prevent overfitting with limited samples (108).
• Implemented feature fusion techniques considering spatial symmetry for enhanced feature representation.

Main Results:

• The OSN model successfully learned and fused discriminative features for osteoporosis prediction.
• Leave-one-out testing demonstrated superior performance of the proposed OSN compared to state-of-the-art methods.
• Achieved high accuracy in osteoporosis category classification using DPR images.

Conclusions:

• The developed deep Octuplet Siamese Network (OSN) is a promising tool for non-invasive osteoporosis assessment using DPR.
• The OSN effectively handles feature complexity and data limitations inherent in dental panoramic radiography.
• This approach offers a significant advancement in leveraging readily available radiographic data for early osteoporosis detection.