An Explainable End-to-End Artificial Intelligence Framework for Lumbar Spondylolisthesis Diagnosis from Radiography Images Using Anatomical Feature Engineering

Spondylolisthesis is a prevalent spinal condition whose diagnosis can be subjective and challenging, creating a need for pre•cise, automated solutions in radiology. This study presents a comprehensive pipeline combining deep learning and machine learning techniques to accurately detect, locate, and classify spondylolisthesis from lumbar spine X-rays. Using the BUU-LSPINE curated dataset of 7200 Lateral/Anteroposterior view X-Ray images of lower spine from 3600 individuals, eight object detection models were evaluated, and the U-Net model was identified as the best performing model for robust vertebral landmark segmentation, to avoid data loss from missed detections. From these landmarks, an initial set of 720 biomechanical features were engineered per patient, which were then restructured to a per-vertebra level and refined to the 85 most significant features to enhance analytical efficiency and reduce model complexity. Considering the limited number of data instances identified with various spondylolisthesis abnormalities, the Augmentation class rebalancing technique was applied to generate a more balanced data set. Nine machine learning models were trained using the balanced dataset to identify the presence of spondylolisthesis in the individual and the grade of abnormality based on Meyerding classification. Among the evaluated models, XGBoost demonstrated the highest classification accuracy and stability (86.7% Test, 86.3% CV). Finally, the interpretability of the AI models is evaluated using the Explainable AI (XAI) technique, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), to interpret the model’s decisions. XAI analysis confirmed that the models learned clinically relevant patterns, enabling transparent, per-case predictions and building confidence in the models for their real-world adoption.