Spine-GraphX: A Graph Neural Network Model for Analyzing Structural Relationships in Lumbar Intervertebral Discs

Magnetic resonance imaging of the lumbar spine is the key to the diagnosis of intervertebral disc degeneration and related pathology. Automated analysis is reliable and can assist clinical decision making while also reducing reader variability. Convolutional networks, including CNNs and U-Net, emphasize local patterns of pixels, which is limited to represent dependencies between vertebrae, discs, and the spinal canal. To bridge this gap, we present Spine-GraphX, a framework that combines GCNs with convolutional features to encode explicit anatomical associations. Experiments are conducted on the SPIDER MRI Spine T2 PNG dataset, which has about 1,550 sagittal T2-weighted slices of 210 subjects. Spine-GraphX was able to achieve an accuracy of 93.5%, a sensitivity of 0.91, Dice score of 0.902, and IoU of 0.829. These results were even better than ResNet-50 U-Net (accuracy 88.7% Dice 0.861) and DenseNet U-Net (accuracy 89.6% Dice 0.868). Group comparisons showed p-values of less than 0.05 which shows statistically reliable increases. The results indicate that the structural relationship modeling offers greater accuracy under noise and small sample sizes and computational efficiency for automated analysis of the lumbar spine.