A novel dual embedding few-shot learning approach for classifying bone loss using orthopantomogram radiographic...

Area of Science:

• Artificial Intelligence in Dentistry
• Natural Language Processing for Medical Imaging
• Machine Learning for Periodontal Disease Detection

Background:

• Orthopantomograms (OPGs) are vital for diagnosing dental and maxillofacial conditions, including bone loss indicative of periodontal disease and osteoporosis.
• Automated analysis of OPG radiographic notes can enhance diagnostic accuracy and reduce interobserver variability.
• Transformer-based models offer potential for classifying bone loss severity from clinical text.

Purpose of the Study:

• To classify bone loss severity (normal, mild, severe) from Orthopantomogram (OPG) radiographic notes.
• To develop and evaluate a novel dual-embedding few-shot learning framework for this classification task.

Main Methods:

• Utilized a dataset of radiographic notes from OPGs, classifying bone loss using Glickman's Classification system.
• Developed the DualFit model with separate Text and Feature Processing Branches, incorporating batch normalization and dropout.
• Employed a Fusion Layer to integrate outputs from both branches for optimized classification performance.

Main Results:

• The DualFit model achieved superior performance compared to BioBERT, ClinicalBERT, and PubMedBERT.
• Attained high accuracy (98.98%), precision (98.71%), recall (99.14%), and F1-score (98.92%), a 5.53% accuracy increase over PubMedBERT.
• Demonstrated excellent multi-class classification with near-perfect precision, recall, and AUC values.

Conclusions:

• The DualFit model represents a significant advancement in automated classification of OPG notes for periodontal bone loss.
• This model streamlines diagnostics, reduces radiologist workload, and facilitates timely patient interventions.
• Future research should focus on external validation and integration with multimodal diagnostic systems.