To refine a method of assessing lumbar stability by applying artificial intelligence and incorporating the use of a walking aid during imaging procedures. A software application was developed to evaluate lumbar stability parameters using machine learning and neural network models (Swin-Resnet). The intra-class correlation coefficient (ICC) was used to assess the agreement between Swin-Resnet and the evaluations conducted by three spine surgeons, each with over ten years of experience. Subsequently, traditional flexion-extension radiographs (FET) and flexion-extension radiographs with a walking aid (FEW) were performed on the enrolled patients with lumbar spondylolisthesis. The developed software was used to measure parameters such as sagittal translation, segmental angulation, and detection rate of lumbar instability. There was no significant statistical difference in the average error between the Swin-Resnet model and physicians in terms of sagittal translation, segmental angulation, and agreement of lumbar instability assessment. The average values of sagittal translation and segmental angulation were significantly higher in FEW than in FET (2.40 (1.09, 3.59) vs. 0.56 (0.16, 1.40); 8.00 (5.00, 10.25) vs. 2.00 (1.00, 5.25), P < 0.05). Among 50 patients, FEW detected 19 cases (38%) of lumbar instability, while FET detected only 3 cases (6%). Additionally, five cases (10%) exhibited a posterior opening angle ≥ 5° in FEW, whereas none met this threshold in FET. The incorporation of a walking aid in lumbar flexion-extension radiographs and the development of automated parameter measurement software facilitates a more accurate and consistent evaluation of lumbar stability in patients with lumbar spondylolisthesis.
Lin et al. (2026) studied this question.