Automatic Segmentation of the Spinal Canal in MR Images with Deep Learning Method

Spinal stenosis is the narrowing and compression of the spinal cord that can be caused by many different reasons. This can also cause compression of the nerves. It can lead to consequences such as difficulty walking, nerve damage, and numbness in the legs, and may affect human life negatively and limit mobility. Data obtained by magnetic resonance imaging (MRI) are often used to diagnose this disease. The frequent use of MRI images increases radiologists' workload and human error. Mistakes made in the diagnosis negatively affect the patients' lives and cause financial losses. To reduce the workload of radiologists and increase the accuracy in diagnosis, a high-accuracy segmentation method using U-net, a deep learning algorithm, on T2-weighted axial MR images is proposed in this study. Pixel accuracy and intersection over union (IoU) metrics are used to calculate performance. Experimental analyses were made at different epoch values to increase the success of the method. According to the experimental analysis, the best results were obtained for pixel accuracy of 99,91% in 100 epochs and 0,8853 for IoU.