Artificial intelligence based intervertebral discs grading: From unsupervised detection to explainable Pfirrmann classification

Intervertebral disc (IVD) degeneration is a leading cause of chronic low back pain, necessitating early and accurate diagnosis to improve patient outcomes. This study presents an artificial intelligence-based framework for automated grading of IVD degeneration, integrating unsupervised disc detection with an explainable Pfirrmann classification model. The proposed pipeline eliminates the need for manual segmentation by leveraging image processing techniques to detect IVDs from sagittal T2-weighted MRI scans. Furthermore, the Pfirrmann classification task was conducted through a convolutional neural network trained on the extended 8-level grading system, providing a more refined assessment than conventional 5-level approaches. To enhance interpretability, Grad-CAM saliency maps were used to highlight the key image features influencing predictions. The model achieved a classification accuracy of 93.2%, with predictions deviating by at most one grade from expert assessments. Furthermore, the introduction of a confidence-based rejection mechanism, proved the possibility to reach accuracy rates of 96.8% on the majority of samples. The explainability analysis revealed distinct patterns of activation for each predicted Pfirrmann grade, reflecting the model's ability to capture specific degenerative features associated with different levels of IVD degradation. This study demonstrates that AI-driven methods can help to standardize the 8-level IVD Pfirrmann grading, reducing inter-observer variability while maintaining high diagnostic reliability.