Multiview hybrid 2D-3D CNN with transformer-based feature fusion for spine lesion classification in MR

As osteoporosis and degenerative spine conditions become increasingly prevalent, the demand for automated and accurate diagnostic methods in medical imaging also grows. Deep learning-based computer-aided diagnosis (CAD) methods have shown promise in identifying spine lesions using X-ray, MRI, and CT scans. However, precise lesion identification requires capturing spatial and volumetric features, which traditional 2D CNN models lack. This study presents a multi-view hybrid deep learning model that integrates 2D and 3D convolutional neural networks (CNNs) with Transformer-based feature fusion to enhance spine lesion diagnosis from MRI scans. The proposed model is evaluated on the RSNA 2024 Lumbar Spine Degenerative Classification dataset, which consists of MRI scans of spine degeneration images. Model performance is assessed using accuracy, recall, F1 score, and AUC-ROC, achieving 0.9019, 0.8984, 0.9019, 0.8998, and 0.97 on the class average, respectively. Our results demonstrate that integrating multi-view learning and volumetric feature extraction significantly enhances diagnostic accuracy for spine lesions, providing a robust and clinically relevant advancement in computer-aided diagnosis.