Development and Validation of a Deep Learning Based Automated Lumbar Spinal Canal Segmentation and Measurement Tool

Poster Presentation at the European Congress of Radiology, Vienna, 2019


We describe a novel use-case of Convolutional Neural Networks to automatically measure spinal canal diameter and detect stenosis on lumbo-sacral T2-weighted MRI images by automated segmentation of the spinal canal.

Methods and Materials

Axial T2 weighted MRI scans of the lumbo-sacral spine of 80 patients (60 normal, 20 with disc herniation) were extracted from PACS and anonymised using HIPAA compliant methodologies. The lumbar spinal canal was manually segmented by a radiologist having 8 years’ experience of MRI – in 10 scans all T2-W axial slices were segmented and in 70 scans segmentation was performed only at the level of the intervertebral discs. The manual segmentations, along with the T2 images were fed into a deep CNN based architecture, having 101 layers, built from scratch with a 80-20 training-validation split. DICE score was used as the loss function. The output was the segmentation of the lumbar spinal canal, its area and anteroposterior diameter. Student’s t-test was performed to compare the AP diameter reported by the model against that reported by the radiologist.


We obtained a final DICE score of 98.8 at the level of normal inter-vertebral discs and 98.1 in those with pathology. The difference in predicted and radiologist-reported AP diameters was insignificant (p<0.01) in 16 cases. The model was also able to successfully detect all axial slices with disc herniation.


The model gave near radiologist-level performance in detecting stenosis and calculating the spinal canal diameter.