Deep Learning in Medical Imaging: Classification of Lumbar Spine Degeneration from MRI Scans

The bachelor thesis focuses on the classification of lumbar spine degeneration from MRI scans using deep learning methods. The theoretical background covers neural networks for image processing, including convolutional neural networks, transfer learning, vision transformers, and medical image segmentation architectures. The practical part is based on the RSNA 2024 Lumbar Spine Degenerative Classification competition. After exploratory analysis of the MRI dataset, several direct classification benchmarks are trained and evaluated. Since each diagnosis depends on a specific condition, imaging plane, spinal level and side, the solution is further extended into a localization-guided workflow. The final pipeline combines anatomical localization, slice selection, ROI-based classification and model ensembling for making the predictions. The results show that this task cannot be treated as simple classification framework and that incorporating anatomical structure is essential.