Transforming Osteoporosis Detection: Leveraging Vision Transformer using Radiographic Analysis of Mandibular Indices

Abstract

Background Osteoporosis is a prevalent bone disease characterized by decreased bone density and structural deterioration, leading to increased fracture risk. Osteoporosis affects 200 million people globally, with one in three women and one in five men over 50 experiencing fractures. Early detection and intervention are crucial for reducing morbidity and mortality. Dental panoramic radiographs (DPRs) can be valuable in identifying osteoporosis by analyzing mandibular indices such as the Mental Index, Panoramic Mandibular Index, Gonial Index, Antegonial Index, and Antegonial Depth. These indices reflect specific anatomical features of the mandible that may correlate with bone density changes indicative of osteoporosis. This study introduces a novel approach to osteoporosis detection using Vision Transformer architecture, focusing on long-range dependencies and complex spatial relationships in medical images, aiming for early clinical application. Methods The study will include 600 digital panoramic radiographs from female patients aged 20-30, 30-40, 40-50, 50- 60, 60-70, and above 70 years, for routine dental checkups and examinations. The data will be saved in DICOM format and morphometric measurements will be performed by two oral radiologists. Quantitative indices such as the Mental Index (MI), Panoramic Mandibular Index (PMI), Gonial Index (G.I.), Antegonial Index (A.I.), and Antegonial Depth (A.D.) will be measured. The initial phase of the methodology involves meticulous acquisition and processing of digital panoramic radiographs, which were divided into six age groups. Each radiograph undergoes comprehensive quality assessment, evaluating technical parameters including brightness, contrast, and positioning accuracy. The preprocessing pipeline uses a multi-stage approach, including histogram equalization, Gaussian filtering, CLAHE, and unsharp masking techniques, to enhance contrast and reduce noise. The annotation and the labeling process uses a rigorous multi-reader approach to ensure data quality and reliability, providing a structured summary of key indices and clinical observations and subjected to transformers architecture. Results The Vision Transformer (ViT) model is highly accurate for osteoporosis detection, identifying 96.5% of cases. However, its lower sensitivity raises concerns about its effectiveness. DenseNet-169 and EfficientNet-B4 models are reliable options, with DenseNet-169 promoting feature reuse and EfficientNet-B4 balancing computational efficiency and performance. ResNet-152 needs improvement for accurate patient identification. The “ViT (Best Tuned)” model is the superior choice for osteoporosis detection in dental panoramic radiographs. Conclusion The study explores transformer models for osteoporosis detection using dental panoramic radiographs, highlighting the potential of A.I. in early diagnosis and timely intervention. Future research should focus on creating diverse datasets and integrating multi-modal data like medical history, genetic predispositions, and imaging techniques for better accuracy. This could enhance predictive capability and make machine learning a crucial component of proactive osteoporosis management and patient care.

Bangladesh Journal of Medical Science Vol. 25. Supplementary Issue 2026, Page : S159-S169