Deep learning framework for automated classification of thoracolumbar fractures using spinal CT images

Objective Thoracolumbar fractures (T11–L2) are common spinal injuries, and accurate classification is essential for treatment planning. However, diagnosis remains challenging because of complex fracture morphology and variability in clinical experience. This study evaluated the feasibility of a deep learning framework for automated classification of thoracolumbar fractures using CT images. Methods A retrospective study was conducted using CT images from 596 patients with thoracolumbar fractures. A convolutional neural network based on a Multiple Instance Learning (MIL) framework was developed, using AlexNet with a gated attention mechanism for feature aggregation. Separate models were trained for coronal (Z-type) and axial (S-type) views, and an ensemble strategy combined the two predictions. Model performance was compared with that of interns and senior physicians. Results The S-type model achieved a classification accuracy of 75.2%, while the Z-type model achieved an accuracy of 88.0%. The ensemble model reached an overall accuracy of 83.8%. With assistance from the deep learning model, the diagnostic accuracy of interns improved from 74% to 95%, approaching the performance of senior physicians (98%). Conclusion The proposed deep learning framework showed promising performance for automated classification of thoracolumbar fractures and may serve as a decision-support tool to improve diagnostic efficiency.