Deep Convolutional Neural Networks Using SegNet for Automatic Spinal Canal Segmentation in Axial MRI

Spinal canal stenosis is a disease in which the spinal cord and nerve roots in the spinal canal are compressed. This condition can cause back or neck pain, numbness, tingling, balance problems, muscle weakness and negatively affect people’s daily lives. Magnetic resonance imaging (MRI) is often used to diagnose spinal canal stenosis. Diagnosis using MRI, together with the increasing number of patients, increase the workload of doctors and, thus, human-based errors. In addition, the time taken for diagnosis is also prolonged due to the workload. However, an early and accurate diagnosis of the disease is critical in accelerating the healing process by determining appropriate patient treatment protocols. In this study, SegNet, a deep learning algorithm for segmentation of the spinal canal in T2-weighted axial images, is proposed to reduce the workload of physicians, shorten the diagnosis time and increase accuracy. Moreover, experimental analyses were performed at different epoch values to obtain the best performance of the method. Pixel accuracy and intersection over union (IoU) metrics were used to determine the method’s success. The best result was 99.87% for pixel accuracy and 0.8169 for IoU at 200 epochs.