ABSTRACT The rapid and accurate diagnosis of spinal disorders, particularly lumbar disc herniation (LDH), through deep learning applied to magnetic resonance imaging (MRI) scans has significant potential to enhance clinical decision‐making and reduce physicians' workload. However, many existing approaches do not simultaneously provide high diagnostic accuracy and sufficient clinical interpretability. To address this shortcoming, this study proposes Lumbar‐DAFSNet (Densely‐Attended Feature Selection Network), a hybrid deep and machine learning framework designed to improve both performance and explainability in MRI‐based LDH detection. The framework is built on a DenseNet backbone and is enhanced with the Convolutional Block Attention Module (CBAM) and the Squeeze‐and‐Excitation Network (SENet). In this way, both channel‐wise and spatial feature representations are strengthened. To determine the most informative feature subset for classification, multiple feature selection strategies were systematically evaluated. Accordingly, the extracted deep features were refined using Principal Component Analysis (PCA), SHapley Additive exPlanations (SHAP), Recursive Feature Elimination (RFE), and SelectKBest, a univariate feature selection method based on statistical scoring functions. In addition, Gradient‐weighted Class Activation Mapping (Grad‐CAM) was employed to visualize the regions that most strongly influenced the model's predictions, thereby improving interpretability. As a consequence of this comprehensive design, Lumbar‐DAFSNet achieved outstanding classification performance. Among the evaluated classifiers, Extreme Gradient Boosting (XGBoost) yielded the highest accuracy (98.41%) and F1 score (98.40%), while Logistic Regression reached an exceptional ROC AUC of 99.73%. Furthermore, Grad‐CAM visualizations highlighted clinically relevant spinal regions, supporting the reliability and clinical relevance of the model. Overall, Lumbar‐DAFSNet provides an accurate and interpretable framework for MRI‐based LDH detection and offers explainable outputs that can assist clinical decision‐making and surgical planning.
Yildirim et al. (Mon,) studied this question.