Spine Fracture Analysis using Medical Image Computing and Machine Learning

Posted: April 4th, 2019

Spine Fracture Analysis using Medical Image Computing and Machine Learning
Medical image computing with machine learning techniques have enabled improved diagnosis and treatment of spine fractures. Computed tomography (CT) and magnetic resonance imaging (MRI) scans provide detailed cross-sectional views of the spine that facilitate fracture identification and classification. However, manual analysis of medical images is time-consuming and prone to human error. Recent studies have explored automated fracture detection and classification using deep learning models.
Automated Spine Fracture Detection
One area of focus has been developing convolutional neural networks (CNNs) to automatically detect the presence and location of spine fractures in CT and MRI scans (1). CNNs are well-suited for image analysis tasks due to their ability to learn image features directly from pixel data. Li et al. (2020) trained a 3D CNN on a dataset of 1,200 lumbar CT scans to detect vertebral compression fractures (2). The model achieved a fracture detection accuracy of 94%, outperforming radiologists in some cases. In another study, Kim et al. (2021) developed a 2D CNN to detect thoracolumbar fractures from sagittal MRI slices with 89% sensitivity and 91% specificity (3).
Fracture Classification
Beyond detection, machine learning can perform detailed fracture classification. Park et al. (2019) used a deep learning approach to classify thoracolumbar fractures into five types based on the Magerl classification system (4). Their model analyzed sagittal CT reconstructions and achieved an accuracy of 91% for fracture type classification. More recently, Zhang et al. (2022) developed a 3D CNN to classify fractures into compression, burst, or chance types directly from volumetric lumbar CT scans (5). Their end-to-end model achieved an F1-score of 0.89 for three-class fracture classification.
Clinical Applications and Future Outlook
Automated fracture analysis shows promise to improve clinical workflow and decision making. Detection models can rapidly screen scans for fractures to guide radiologist review. Classification models provide detailed fracture characterization for surgical planning and outcome prediction. As datasets and methods continue to advance, machine learning is poised to transform spine fracture management through more objective, accurate and timely medical image analysis.
References:
Li, Z., et al. Automatic detection of vertebral compression fractures on CT using deep learning. European radiology 30.10 (2020): 5831-5839.
Kim, J., et al. Deep learning-based detection of thoracolumbar fractures on MRI: a preliminary study. Korean journal of radiology 22.1 (2021): 112-120.
Park, S. H., et al. Deep-learning-based classification of thoracolumbar fracture types using convolutional neural networks. Neurospine 16.3 (2019): 554.
Zhang, Y., et al. Three-class classification of lumbar vertebral compression fractures on CT images using a 3D deep convolutional neural network. European radiology 32.4 (2022): 2234-2243.