A hybrid MMBERT framework for classifying periodontal bone loss: Integrating visual and textual information

Introduction: Periodontitis, a chronic inflammatory disease, causes bone loss and tooth instability. Early diagnosis, aided by clinical and radiographic assessments, is crucial. mmBERT, a multimodal deep learning approach, integrates image and text data to enhance bone loss prediction but faces challenges such as image variability and limited datasets. This study proposes a Hybrid mmBERT Framework to improve diagnostic accuracy in intraoral periapical (IOPA) radiographs. Methods: A dataset of 150 IOPA images from Saveetha Dias includes expert-annotated interpretations of periodontal bone loss. Images are split into training and test sets. Clinical notes are preprocessed, normalized, and feature-extracted using ClinicalBERT. The mmBERT model fuses image and text embeddings into a 1024-dimensional space with a 0.3 dropout rate. Its architecture integrates a vision encoder (ResNet50), text encoder (ClinicalBERT), and a cross-modal transformer with 16 attention heads. Results: The study shows strong overlap between predicted and ground-truth masks, good intersection-over-union performance, and acceptable boundary accuracy. The MMBERT model achieved final training and test accuracies of 99.11 % and 100 %, respectively, demonstrating its robustness and reliability in clinical applications. Conclusion: The study presents a classification method for periodontal bone loss based on MMBERT architecture. It demonstrates promising radiographic interpretation performance while necessitating further clinical application research.