Automated Detection of Lumbosacral Transitional Vertebrae on Plain Lumbar Radiographs Using a Deep Learning Model.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Lumbosacral transitional vertebra (LSTV) is a common variant with inconsistent radiographic identification.
• Inconsistent LSTV detection can lead to clinical issues like back pain and surgical errors.

Purpose of the Study:

• Develop and validate a deep learning AI model for automated LSTV detection on lumbar radiographs.
• Improve the accuracy and reliability of LSTV identification in clinical practice.

Main Methods:

• Retrospective analysis of 3116 lumbar lateral radiographs.
• Evaluation of deep learning models (DINOv2, CLIP, ResNet-50) for LSTV classification.
• ResNet-50 model selected and validated using fivefold cross-validation.

Main Results:

• The AI model achieved 76.4% accuracy, 85.1% sensitivity, and 61.9% specificity on an independent test set.
• The model demonstrated an AUC of 0.84, correctly identifying most LSTV and normal cases.
• Gradient-weighted class activation mapping (Grad-CAM) was used for model interpretability.

Conclusions:

• An AI-based system provides accurate automated detection of LSTV on plain radiographs.
• This tool has potential to reduce diagnostic errors and enhance pre-operative planning.
• The AI model can improve patient safety by aiding in correct LSTV identification.