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

Abstract

Background: Orthopantomograms (OPGs) are essential diagnostic tools in dental and maxillofacial care, providing a panoramic view of the jaws, teeth, and surrounding bone structures. Detecting bone loss, which indicates periodontal disease and systemic conditions like osteoporosis, is crucial for early diagnosis and treatment planning. Periodontists use OPGs to identify subtle radiographic features that signify different stages of bone loss. Automated systems integrating radiographic imaging with textual notes can enhance diagnostic accuracy and minimize interobserver variability. Radiographic notes, which summarize clinical observations and preliminary interpretations, can be utilized for classification through natural language processing techniques, including Transformer-based models. This study will classify bone loss severity (normal, mild, or severe) from OPG notes using a novel dual-embedding few-shot learning framework.

Methods: This study used a dataset of radiographic notes from OPGs gathered at Saveetha Dental College and Hospital in Chennai. Bone loss was classified according to Glickman's Classification system. The proposed DualFit model architecture consists of two main branches: a Text Processing Branch for converting textual data into dense vectors and a Feature Processing Branch for analyzing numerical and categorical data. Key techniques such as batch normalization and dropout layers were implemented to improve learning stability and reduce overfitting. A Fusion Layer was utilized to merge outputs from both branches, optimizing classification performance.

Results: The DualFit model outperformed leading models like BioBERT, ClinicalBERT, and PubMedBERT. It attained an accuracy of 98.98%, precision of 98.71%, recall of 99.14%, and an F1-score of 98.92%, marking a 5.53% accuracy increase over PubMedBERT. Additionally, the model excelled in multi-class classification tasks, ensuring class balance and achieving near-perfect values for precision, recall, and area under both the ROC and precision-recall curves.

Conclusions: The DualFit model significantly advances the automated classification of OPG radiographic notes related to periodontal bone loss. Outperforming existing Transformer-based models streamlines the diagnostic workflow, reduces the workload of radiologists, and enables timely interventions for improved patient outcomes. Future work should explore external validation and integration with multimodal diagnostic systems.