Comparative Model Evaluation of Lightweight Transformer-Based and CNN-Based Architechture for Degenerative Lumbar Spine Classification

Low back pain is a leading cause of disability globally, is often associated with degenerative lumbar spine conditions. Accurate diagnosis of these conditions is critical but challenging due to the subjective nature of MRI interpretation and the weak correlation between imaging findings and symptoms. Thus, the purpose of this study is to evaluate the performance of CNN-based architectures (VGG-16, EfficientNetB0, EfficientNetV2) and transformer-based architecture specifically Vision Transformers (ViT) in classifying lumbar spine conditions as normal, moderate, or severe. Using the RSNA 2024 Lumbar Spine Degenerative Classification dataset. The models are evaluated using the ROC-AUC and PRAUC as the performance metrics due to the dataset imbalance. Results indicate that EfficientNetB0 achieved the highest overall performance, with an average ROC-AUC of 0.784 and PR-AUC of 0.528, demonstrating strong adaptability to imbalanced datasets. EfficientNetV2 also performed competitively, while VGG-16 showed moderate effectiveness. The Vision Transformer (ViT), however, underperformed due to its reliance on larger datasets and challenges in capturing fine-grained spatial features. The findings highlight the potential of EfficientNet-based models for accurate and efficient lumbar spine diagnostics. This study underscores the potential of advanced deep learning approaches in improving diagnostic workflows for degenerative lumbar spine conditions.