Development of AI-Based Predictive Models for Osteoporosis Diagnosis in Postmenopausal Women from Panoramic Radiographs

Objectives: The aim of this study was to develop AI-based predictive models to assess the risk of osteoporosis in postmenopausal women using panoramic radiographs (OPTs). Methods: A total of 301 panoramic radiographs (OPTs) from postmenopausal women were collected and labeled based on DXA-assessed bone mineral density. Of these, 245 OPTs from the Hospital of San Giovanni Rotondo were used for model training and internal testing, while 56 OPTs from the University of Parma served as an external validation set. A mandibular region of interest (ROI) was defined on each image. Predictive models were developed using classical radiomics, deep radiomics, and convolutional neural networks (CNNs), evaluated based on AUC, accuracy, sensitivity, and specificity. Results: Among the tested approaches, classical radiomics showed limited predictive ability (AUC = 0.514), whereas deep radiomics using DenseNet-121 features combined with logistic regression achieved the best performance in this group (AUC = 0.722). For end-to-end CNNs, ResNet-50 using a hybrid feature extraction strategy achieved the highest AUC in external validation (AUC = 0.786), with a sensitivity of 90.5%. While internal testing yielded high performance metrics, external validation revealed reduced generalizability, highlighting the challenges of translating AI models into clinical practice. Conclusions: AI-based models show potential for opportunistic osteoporosis screening from OPT images. Although the results are promising, particularly those obtained with deep radiomics and transfer learning strategies, further refinement and validation in larger and more diverse populations are essential before clinical application. These models could support the early, non-invasive identification of at-risk patients, complementing current diagnostic pathways.