Feasibility of Deep Learning Algorithms for Reporting in Routine Spine Magnetic Resonance Imaging

Area of Science:

• Spine imaging and diagnostics
• Artificial intelligence in medicine
• Deep learning for medical image analysis

Background:

• Magnetic resonance imaging (MRI) analysis for spine pathologies requires accurate predictors of clinical outcomes.
• Developing AI-driven MRI analysis can improve surgical indications and patient outcomes.
• Automated analysis can lead to cost savings by avoiding invasive procedures.

Purpose of the Study:

• To assess deep learning neural network models' ability to identify spinal pathologies in MRI data.
• To evaluate the feasibility of generating automated verbal MRI reports comparable to radiologist reports.
• To correlate MRI findings with varying severities of common spinal injuries.

Main Methods:

• A 3D anatomical model was fitted to patient MRIs for segmentation model training.
• Convolutional neural networks were trained using MRI series (T1, T2, sagittal, axial, transverse).
• Natural language processing was used to generate reports from extracted radiologist findings.

Main Results:

• Deep learning models achieved high accuracy in detecting spinal pathologies.
• Foraminal stenosis detector accuracy: 81% (sensitivity 72.4%, specificity 83.1%).
• Central stenosis detector accuracy: 86.2% (sensitivity 91.1%, specificity 82.5%).
• Disc herniation detector accuracy: 85.2% (sensitivity 81.8%, specificity 87.4%).

Conclusions:

• Deep learning algorithms show potential for routine spine MRI reporting.
• Minimal disparity in accuracy, sensitivity, and specificity suggests models are not overfitted.
• Variability in training data helps deep neural networks focus on common pathologies.