Purpose
To develop and validate a deep learning–based artificial intelligence (AI) model for automated measurement of lumbar lordosis (LL) angles from whole spine lateral radiographs.
Materials and Methods
A total of 888 lateral spine X-rays (2019–2021) were retrospectively collected and annotated with four anatomical keypoints (L1 and S1 vertebral landmarks). An AI model using Detectron2 with a Keypoint R-CNN and ResNeXt-101 backbone was trained with data augmentation. Performance was evaluated on 50 test images, comparing AI results to manual annotations by two orthopedic surgeons using intraclass correlation coefficient (ICC), Pearson’s correlation, and Bland–Altman analysis.
Results
The model achieved an average precision of 71.63 for bounding boxes and 86.61 for keypoints. ICCs between AI and human raters ranged from 0.918 to 0.962. Pearson correlation coefficients were r=0.849 and r=0.903. Bland–Altman analysis showed minor underestimation biases (–3.42° and –4.28°) with acceptable agreement.
Conclusions
The AI model showed excellent agreement with expert measurements and high reliability in LL angle assessment. Despite a slight underestimation, it offers a scalable, consistent tool for clinical use. Further studies should evaluate generalizability and interpretability in broader settings.

Citations

• Efficacy of Biportal Endoscopic Decompression for Lumbar Spinal Stenosis: A Meta-Analysis With Single-Arm Analysis and Comparative Analysis With Microscopic Decompression and Uniportal Endoscopic Decompression
  Shuangwen Lv, Haiwen Lv, Yupeng He, Xiansheng Xia
  Operative Neurosurgery.2024; 27(2): 158.     CrossRef

Objective
To investigate the utility of a deep learning model in diagnosing traumatic lumbar fractures on computed tomography (CT) images.
Summary of Background
Data: CT scans are widely used as the first choice for detecting spinal fractures in patients with severe trauma. Although CT scans have high diagnostic accuracy, fractures can occasionally be missed. Recently, deep learning has been applied in various fields of medical imaging.
Methods
CT images from 480 patients (3695 vertebrae) who visited a level-one trauma center with lumbar fractures were retrospectively analyzed. The diagnostic results were confirmed by two experienced musculoskeletal radiologists and one experienced spine surgeon using magnetic resonance imaging (MRI). Deep learning networks were employed for diagnosis, with 425 cases used for training and 55 cases for testing. Sensitivity, specificity, accuracy, and the area under the receiver operating characteristic curve (AUROC) were calculated to evaluate diagnostic performance.
Results
The model successfully identified 107 out of 129 vertebrae with fractures, achieving a sensitivity of 82.95%, a specificity of 93.24%, an AUROC of 0.936, and an overall accuracy of 88.45%.
Conclusions
This study demonstrated that the deep learning model showed high accuracy in diagnosing traumatic lumbar fractures. This approach has the potential to assist spine specialists, radiologists, and trauma care experts. Further validation is needed to determine its effectiveness in clinical settings.