Classification of Lumbar Spine Degenerative Diseases Using Deep Learning Techniques

This study investigates the use of deep learning (DL) techniques for the classification of lumbar spine degenerative diseases. In particular, Magnetic Resonance (MR) images used for the detection of spinal canal stenosis are evaluated. The potential of deep learning models to accelerate diagnostic processes through their capability for automatic analysis of radiological images is demonstrated. Various deep learning models were employed in the study; however, the lowest loss value was achieved with the EfficientNetV2-Large architecture. Advanced data augmentation techniques, especially targeted approaches for rare cases, and the use of high-resolution (512x512) images significantly improved the model's performance. As a result of architectural updates and data processing strategies, the test log-loss value was reduced to as low as 0.69. Additionally, the results obtained by combining the predictions of different models through ensemble learning with a soft voting method are also presented. This approach yielded a low log-loss value of 0.604510 on the public test dataset. The results demonstrate that the model is capable of distinguishing clinically critical "severe" cases and maintains its generalization ability even in an expanded class structure.