IgeaNet: a deep learning model for opportunistic osteoporosis screening from chest X-ray images
Fabio Mattiussi1, Chiara Camponovo2, Francesco Magoga3
1Integrated Diagnostics Institute of Southern Switzerland (IDISI), Ente Ospedaliero Cantonale (EOC), Via Tesserete 46, 6900, Lugano, Switzerland. fabio.mattiussi@eoc.ch.
Purpose:
To develop and validate a multi-modal deep learning model for opportunistic screening of osteoporosis by classifying bone mineral density (BMD) from chest radiographs integrated with clinical metadata.
Materials And Methods:
This single-centre retrospective study included 833 adult patients who underwent chest radiography and dual-energy X-ray absorptiometry (DEXA) within 12 months (January 2018-December 2024). A convolutional neural network (CNN) model was developed using a convolutional branch to process chest radiographs, while a dense branch processed patient demographics (age, sex, height, and weight). Clinical data available in the DEXA report were also included. The two representations were merged into a fully connected classifier optimized end-to-end.
Results:
The dataset comprised training (n = 589), validation (n = 124), and testing (n = 120) sets. On the internal test set, the model achieved 85.8% accuracy (95% CI: 78.5-90.9%), F1-macro 0.86, and AUC-macro 0.95. Class-specific performance showed sensitivity/specificity of 84.2%/91.4% for normal, 90.2%/77.1% for osteopenia, and 82.3%/91.9% for osteoporosis, with 14.2% misclassification rate. Analysis of DEXA anatomical site distribution revealed wrist as the determining site in 302 patients (36.3%), hip in 384 patients (46.1%), and lumbar spine in 141 patients (16.9%). Inclusion of wrist measurements changed diagnostic classification in 115 patients (13.8%), with 34 patients reclassified from normal to osteopenia, 14 from normal to osteoporosis, and 67 from osteopenia to osteoporosis. This differential diagnostic yield reflects distinct bone composition differences, as wrist cortical bone may reveal mineral density reductions undetected at predominantly trabecular sites.
Conclusions:
Integration of chest radiographs and clinical data by residual network with attention represents a feasible approach for the preliminary opportunistic classification of BMD without additional radiation exposure. However, these findings are hypothesis-generating and derive from internal validation; external prospective studies are required before clinical implementation.