AI-based prediction models for non-recognizable thoracolumbar compression fractures by X-ray inspection

Abstract

Objective:The study evaluated the application value of an artificial intelligence-based classification model of vertebral fractures in lumbar X-ray images. Methods: Patients who received lateral lumbar radiographs and MRI in our unit from 2021 to 2022 were retrospectively selected. According to the MRI results, the included vertebrae were divided into three categories: fresh fracture, cold fracture, and normal vertebrae. A ResNet-18 classification model was constructed using delineated ROIs on the MRI images and the performance of the model was evaluated. Results: A total of 662 vertebrae from 272 patients were included in the study. The vertebrae were randomly divided into a training set (N=529) and an independent validation set (N=133). The performance of the model in discerning normal vertebrae, recent fractures, and chronic fractures was evaluated, showing accuracies of 0.91%, 0.42%, and 0.75%, respectively. The sensitivity measurements for these categories were 0.91%, 0.408%, and 0.72%, while the specificities were 0.796%, 0.892%, and 0.796%, respectively. Conclusion: This study demonstrated the feasibility of applying artificial intelligence to develop a tripartite classification model for lumbar X-ray images; however, additional refinements are required to enhance its efficacy..