Hybrid attention and MLP-based deep learning for MRI classification of lumbar spine degeneration

Abstract Lumbar spine degeneration (LSD) is a major contributor to lower back pain (LBP) worldwide, typically diagnosed using magnetic resonance imaging (MRI). Manual interpretation of MRI scans is time-consuming and prone to variability, creating an urgent need for automated solutions. This study introduces a deep learning (DL) framework for classifying three key degenerative conditions, neural foraminal narrowing, subarticular stenosis, and spinal canal stenosis, across five lumbar disc levels. We evaluated multiple architectures, including VGG-16, EfficientNetV2-S, and YOLOv5, and enhanced feature extraction using the Convolutional Block Attention Module (CBAM) while improving classification through a Multilayer Perceptron (MLP). The dataset comprises 48,657 multi-centre MRI images and was partitioned using a stratified 80:20 training–testing split. Class imbalance was addressed using Random Oversampling (ROS), Synthetic Minority Over-sampling Technique (SMOTE), and Generative Adversarial Network (GAN)-based augmentation. Model development incorporated validation monitoring and early stopping during training, while final performance was assessed on an independent testing set using accuracy, precision, recall, F1-score, confusion matrices, and receiver operating characteristic (ROC) analysis. The proposed hybrid model, VGG‑16+CBAM+MLP, achieved 94% accuracy, with precision, recall, and F1‑score exceeding 91%. Paired statistical tests confirmed these improvements were significant ( p < 0.05), reinforcing the robustness of the approach. These results demonstrate the potential of Artificial Intelligence (AI)-driven tools to support lumbar spine MRI assessment. Further external validation is required before clinical deployment.