Segmentation of the Area Between Anterior and Posterior Vertebral Elements in Axial MR Images Using U-Net

Lumbar spinal stenosis (LSS) is a disease that appears due to narrowing of the nerve root and intervertebral canals in the lumbar spine and generally affects patients over the age of 50. Common symptoms of LSS include leg weakness, cramps, pain, and sensation loss in the lower back and leg. Early diagnosis is crucial as the severity of LSS can significantly reduce a patient's quality of life. The area between the anterior and posterior (AAP) vertebral elements in Axial MR images is a critical diagnostic region for LSS, which causes compression in this area. This compression affects both the spinal canal and nerve roots. In this study, U-Net, a CNN-based deep learning algorithm, is proposed and developed for segmenting the AAP region from axial T2-weighted lumbar spine MR images. The performance of the proposed method is evaluated by pixel accuracy and intersection over union (IoU) metrics. It is observed that as the batch size varies from 8 to 128, there are no significant fluctuations in performance metrics; however, beyond this range, the results decrease. The U-Net model achieves 0.99435 pixel accuracy and 0.72276 IoU with a batch size of 8, outperforming existing studies and highlighting its potential to improve early diagnosis and treatment strategies of LSS disease.