An explainable deep learning method for diagnosing lumbar spine disorders from medical images

Lower back pain (LBP) is a global medical condition that is suffered by up to 80% of people at least once in their lifetime. Various abnormal conditions could cause LBP, and their diagnostic importance can range from Mild, to Severe level of clinical importance. One of these conditions is Lumbar Spinal Stenosis (LSS), which is usually a Serious or Severe condition, hence the need for its prompt diagnosis. The manual (visual) assessment process of magnetic resonance imaging (MRI) is time-consuming, labour-intensive, and prone to delays. This makes the integration of an AI-enabled diagnostic system an important opportunity for improving clinical workflow efficiency. In this paper, two convolutional neural network (CNN) models have been built to diagnose T2-weighted (T2W) sagittal MRI scans. The first model classifies the patient's sagittal view MR images into five grades of clinical importance, while the second model detects spinal stenosis in those images. Both models achieved strong test performance, with accuracies and recalls between 97% and 98%, and the stenosis model achieved 88% accuracy and recall on an independent external validation dataset. Four explainable AI (XAI) methods - LIME, Grad-CAM, SHAP, and Integrated Gradients (IG) - were applied to both models to get explainable insights. Quantitative faithfulness testing demonstrated that SHAP consistently provided the most reliable explanations, achieving the lowest deletion mean AUCs (0.2164 and 0.4315) and highest insertion mean AUCs (0.9194 and 0.8874) across both models. These explainable diagnoses help ensure the models are not only accurate but also become reliable for potential clinical use.