Automated Grading of Lumbar Disc Degeneration Using a Push-Pull Regularization Network Based on MRI

Area of Science:

• Biomedical Engineering
• Radiology
• Artificial Intelligence

Background:

• Lower back pain is a common health issue linked to intervertebral disc (IVD) degeneration.
• Accurate quantification of IVD degeneration is crucial for biomechanical simulations of spinal conditions.
• Classifying degenerated IVDs from MR images is challenging due to high variability and small differences.

Purpose of the Study:

• To evaluate a computer-assisted method for grading IVD degeneration using a novel push-pull regularization (PPR) strategy.
• To assess the feasibility and performance improvement offered by the PPR strategy in IVD degeneration classification.

Main Methods:

• A retrospective study included 500 subjects with lumbar disorders.
• Deep learning models (VGG-M, VGG-16, GoogleNet, ResNet-34) were trained and tested using T2-weighted spin echo sequences.
• Intervertebral disc (IVD) degeneration was classified into five grades (Pfirrmann system) with and without the PPR strategy.
• Results were compared against classifications by three spinal radiologists using paired t-tests.

Main Results:

• The PPR strategy significantly improved classification accuracy for grades II and III IVD degeneration by over 10% (P < 0.05).
• Overall classification accuracy (grades I-V) was enhanced by more than 8% (P < 0.05) across four CNN models.
• The PPR strategy demonstrated a notable improvement in the performance of convolutional neural network (CNN) models.

Conclusions:

• The push-pull regularization (PPR) strategy effectively enhances the classification performance of deep learning models for intervertebral disc (IVD) degeneration.
• This method improves the representational capability of CNNs, aiding in more accurate grading of spinal degeneration.
• The findings support the utility of PPR for computer-assisted diagnosis and treatment planning in spinal conditions.