International Congress on Human-Computer Interaction, Optimization and Robotic Applications · May 21, 2026
Osteoporosis Risk Classification from Lumbar Spine DXA Images
By Mirza AKM, Hafiz WA, Rajab AS et al.
Osteoporosis is a major global health problem, particularly affecting older adults. If left untreated, it can lead to fragility fractures, chronic pain, disability, reduced quality of life, and increased mortality. Therefore, early identification and timely preventive interventions are essential. Dual-energy X-ray absorptiometry (DXA) is widely used in the clinical assessment of osteoporosis. In this study, we perform three-class osteoporosis risk classification (normal, osteopenia, and osteoporosis) from anteroposterior lumbar DXA images using contemporary deep learning models. Multiple backbones were trained and evaluated under a unified experimental protocol to enable fair comparison across model families. On the held-out test set, ConvNeXtV2-Base achieved the best overall performance, reaching an accuracy of 88.89% with macroprecision, macro-recall, macro-specificity, and macro-F1 of$\mathbf{8 9. 1 8 \%, ~} \mathbf{8 9. 2 0 \%, ~} \mathbf{9 4. 2 4 \%}$, and$\mathbf{8 9. 1 3 \%}$, respectively. The bestperforming model achieved a macro-averaged ROC-AUC of 0.971. Explainable AI was used to visualize the image regions contributing most to the predictions. The resulting overlays generally highlighted anatomically plausible lumbar spine structures, supporting transparent interpretation of the decision cues. Overall, the findings indicate that deep learning can provide accurate and interpretable osteoporosis risk stratification from DXA images and may serve as a supportive second-reader tool to assist clinicians in DXA-based evaluation.