Use of neural networks to predict a potential osteoporotic metabolic condition based on the panoramic mandibular index.

OBJECTIVE

This study investigates the use of neural networks to predict potential osteoporotic metabolic conditions using the Panoramic Mandibular Index (PMI) derived from panoramic radiographs.

STUDY

DESIGN

A total of n = 750 radiographs were collected and 1,500 regions of interest (ROIs) were analysed. Convolutional neural network (CNN), specifically ResNet-50, were used with transfer learning, achieving the highest accuracy. Three classification approaches were tested: binary classification (PMI > 0.25 vs. PMI < 0.25), 3-class classification (PMI > 0.25, 0.2-0.25, PMI < 0.2), and a modified 2-class classification (PMI > 0.25 vs. PMI < 0.2).

RESULTS

The best classification, the modified 2-class classification, achieved an accuracy of 89.5%. Data augmentation techniques including rotation, reflection, and scaling, improved model robustness.

CONCLUSIONS

This study demonstrates that a neural network can be trained to effectively classify PMI values in panoramic radiographs, providing an easy applicable and noninvasive method to identify patients at risk of osteoporosis.