Comparison of dual-energy X-ray absorptiometry and lumbar radiographs using deep learning for osteoporosis diagnosis

Osteoporosis and osteopenia, characterized by reduced bone mineral density (BMD), increase fracture risk, particularly in older adults. Early detection is critical for preventing fractures and preserving quality of life. Variability in radiographic interpretation highlights the need for automated diagnostic approaches. This study evaluates whether deep learning (DL) analysis of lumbar radiographs can serve as an alternative to dual-energy X-ray absorptiometry (DEXA) for detecting osteopenia and osteoporosis.

In this retrospective study, 1039 patients (aged 20-95 years, mean 64.99 ± 11.61) who underwent both DEXA and lumbar spine radiography within three months were included. BMD was classified using T-scores for adults over 50, including postmenopausal women and men, and Z-scores for younger participants. Lateral lumbar radiographs (L1-L4) were categorized as normal, osteopenic, or osteoporotic according to DEXA results. Five pretrained transfer learning models (DenseNet-169, EfficientNet-B7, Inception-ResNetV2, ResNet-152V2, Xception) were used for feature extraction, with hyperparameters optimized using the Grey Wolf Optimization (GWO) algorithm. Models were evaluated for both multiclass and binary classification tasks.

In multiclass classification, DenseNet169 achieved the highest test AUC of 0.81. In binary classification, Xception (AUC: 0.90) was most effective in distinguishing normal from osteopenia, while InceptionResNetV2 (accuracy: 90.08%) performed best in separating normal from osteoporosis.

These results suggest that DL analysis of lumbar radiographs, combined with transfer learning and metaheuristic optimization, may support BMD category classification as a complementary screening tool. This approach may provide a reliable, non-invasive, and accessible alternative to DEXA for osteoporosis screening and diagnosis.