A CNN Autoencoder for Learning Latent Disc Geometry from Segmented Lumbar Spine MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Spine Biomechanics

Background:

• Low back pain is a leading cause of disability globally.
• Lumbar intervertebral disc pathology is a frequent pain driver.
• Disc geometry is crucial for understanding mechanical behavior and pathology.

Purpose of the Study:

• To develop a convolutional neural network (CNN) autoencoder for extracting latent features from segmented lumbar disc MRI.
• To interpret these latent features and assess their utility in identifying disc pathology.
• To complement standard geometric measures with novel feature insights.

Main Methods:

• Examined 195 sagittal T1-weighted lumbar spine MRIs.
• Developed a pipeline involving MRI segmentation and CNN autoencoder training.
• Extracted latent geometric features and compared them with standard geometric measures for predicting disc narrowing.

Main Results:

• Achieved high segmentation accuracy (IoU 0.82, DSC 0.90).
• The CNN autoencoder effectively extracted latent features, with optimal convergence at a 4x1 bottleneck.
• Combined latent and geometric features significantly improved disc narrowing prediction compared to using either set alone.
• Latent features captured disc shape and angular orientation.

Conclusions:

• A CNN autoencoder successfully extracts interpretable latent features from lumbar disc MRI.
• This approach enhances the prediction of disc narrowing.
• Future research will incorporate voxel intensity for compositional analysis.