Osteoporosis screening: Leveraging EfficientNet with complete and cropped facial panoramic radiography imaging

Abstract

This paper introduces a novel approach for detecting osteoporosis through the analysis of dental panoramic radiographs (PR) using convolutional neural networks (CNN) based on the EfficientNet architecture. A dataset of PR images from 750 patients was curated, with 579 images showing no osteoporosis (Class C1) and 171 indicating signs of osteoporosis (Class C3). EfficientNet B5, B6, and B7 models were trained to identify osteoporosis in both full PR images and cropped sections focusing on the mandibular cortical region. The models achieved accuracy rates of 95% to 99% using k-fold cross-validation. To validate the models, a new dataset of 60 adult patients who underwent DXA at the lumbar spine and hip was used, with 37 diagnosed with osteoporosis, 22 with normal bone density, and 2 with osteopenia. These patients were invited for PR acquisition, and their PRs were assessed by a general dentist, a radiologist, and the EfficientNet models. Using DXA results as a reference, the CNN models outperformed the dentist in both full and cropped PR evaluations, although the radiologist showed slightly higher accuracy. The study advances computer-aided dental diagnostics by providing an efficient, non-invasive method for osteoporosis screening, with the training and testing codes made publicly available.