Deep learning for predicting lumbar segmental instability using neutral lateral lumbar radiographs: a retrospective study

To develop a deep learning model for predicting lumbar segmental instability (LSI) using neutral lateral lumbar radiographs and to identify key radiographic features associated with LSI.

A DenseNet121-based stacking ensemble model was integrated with Support Vector Machine, Random Forest, and Softmax classifiers. Model validation employed 10-fold cross-validation, with performance assessed using AUC, accuracy, sensitivity, specificity, and F1-score. Sensitivity analyses evaluated robustness across spinal/non-spinal regions, age/gender subgroups, and feature interactions. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to localize critical anatomical regions, which were further validated in machine learning frameworks.

The DenseNet121-stacking model achieved an AUC of 0.838, accuracy of 77.5%, sensitivity of 75.9%, and specificity of 78.4%. Calibration curves confirmed strong alignment with clinical outcomes. Grad-CAM identified facet joints (34.93%), intervertebral discs (25.99%), and osteophytes (22.66%) as the predominant contributors. The integration of these features into machine learning models yielded an AUC of 0.749. Subgroup analyses demonstrated consistent performance across age and gender groups. Decision curve analysis confirmed the clinical utility of the model in all cohorts.

The stacking ensemble model developed in this study effectively predicts LSI based on neutral lateral lumbar radiographs and identifies key imaging biomarkers, including facet joint hypertrophy, disc degeneration, and osteophyte formation. The model demonstrated stable performance across different age and gender groups, indicating strong generalizability and providing a reliable tool for precise clinical screening and individualized decision-making.