Background: Lumbar spondylolisthesis (LS) is a common spinal disorder characterized by the forward displacement of the vertebra. Early detection is challenging due to asymptomatic presentation in the early stages. This study develops and validates an AI-based deep learning model for the early, high-precision diagnosis of LS using lumbar X-ray images. Methods: A total of 3300 lateral lumbar X-ray images were collected from Beijing Tongren Hospital, and an external dataset of 1100 images was used for validation. The images were randomly divided into the training, validation, and test sets. The model uses semantic segmentation to precisely segment vertebral bodies and calculate distances between vertebrae to identify and grade LS using the Meyerding classification. Model performance was compared to other algorithms and clinical experts. Results: The model achieved F1 Scores of 0.92 and 0.91 on the hospital and external datasets, respectively, outperforming other methods. It showed diagnostic accuracies of 96.1% and 94.4%, exceeding the performance of physicians (90.6% and 89.3%). These results highlight the potential of AI in improving diagnostic accuracy and clinical decision-making. Conclusions: Our deep learning model demonstrates high accuracy and reliability in diagnosing LS, providing a valuable tool for early detection and better patient outcomes. Future work will involve expanding the dataset and validating the model in clinical settings.

The biomechanical differences between stand-alone interbody bone grafting and cage implantation in unilateral biportal endoscopic lumbar interbody fusion (ULIF) remain unclear. This study compared the two strategies in a patient-specific finite element model and assessed model credibility using geometric reconstruction, published range-of-motion (ROM) data, and mesh convergence analysis. A subject-specific L4-L5 finite element model with residual Grade I spondylolisthesis was used to compare bilateral pedicle-screw fixation with stand-alone interbody impaction bone grafting versus porous titanium cage implantation. A 500-N axial preload and a 7.5-N m pure moment were applied independently in six directions. Segmental ROM, model-reported stiffness, and von Mises stresses in the interbody constructs, screw-rod system, and endplates were evaluated. ROM was calculated from the relative rotation of L4 with respect to L5. In addition to comparison with published cadaveric ROM, five characteristic mesh sizes (4.0, 3.5, 3.0, 2.5, and 2.0 mm) were assessed under identical model settings. Cage implantation reduced ROM by 30.4%-37.3% across the six loading directions and produced higher model-reported stiffness than stand-alone bone grafting. The cage model also generated higher localized stresses in the interbody construct and screw-rod system; during right lateral bending, screw and rod stresses increased by 556.6% and 361.2%, respectively. Endplate effects were direction- and level-dependent, including a 326.7% increase at the L4 inferior endplate during right lateral bending. In the mesh study, the monitored von Mises stresses for 4.0-, 3.5-, 3.0-, 2.5-, and 2.0-mm meshes were 6.820, 11.499, 9.849, 9.865, and 9.840 MPa, respectively; the maximum pairwise difference among meshes of 3.0 mm or finer was 0.25%. Within the prescribed loading, material, contact, and mesh assumptions of this single patient-specific model, cage implantation provided greater immediate segmental restraint but generated a less favorable localized stress environment in selected implant and endplate regions. Stand-alone bone grafting produced greater motion but a more diffuse load-transfer pattern. The mesh convergence results support numerical stability of the reported stress solution but do not constitute independent biological validation. These computational findings are hypothesis-generating and should not be interpreted as proof of clinical superiority, long-term safety, or a direct basis for patient selection.