A deep learning model for detection of cervical spinal cord compression in MRI scans

Abstract

Magnetic Resonance Imaging (MRI) evidence of spinal cord compression plays a central role in the diagnosis of degenerative cervical myelopathy (DCM). There is growing recognition that deep learning models may assist in addressing the increasing volume of medical imaging data and provide initial interpretation of images gathered in a primary-care setting. We aimed to develop and validate a deep learning model for detection of cervical spinal cord compression in MRI scans. Patients undergoing surgery for DCM as a part of the AO Spine CSM-NA or CSM-I prospective cohort studies were included in our study. Patients were divided into a training/validation or holdout dataset. Images were labelled by two specialist physicians. We trained a deep convolutional neural network using images from the training/validation dataset and assessed model performance on the holdout dataset. The training/validation cohort included 201 patients with 6588 images and the holdout dataset included 88 patients with 2991 images. On the holdout dataset the deep learning model achieved an overall AUC of 0.94, sensitivity of 0.88, specificity of 0.89, and f1-score of 0.82. This model could improve the efficiency and objectivity of the interpretation of cervical spine MRI scans.

Conflict of interest statement

The authors declare no competing interests.

Figures

Figure 1

Consort diagram showing the process of data acquisition and partitioning into training/validation and holdout datasets. CSM-NA—Cervical Spondylotic Myelopathy North American Clinical Trial. CSM-I—Cervical Spondylotic Myelopathy International Trial.

Figure 2

Representative axial T2-weighted MRI images showing (A) no spinal cord compression, (B) partial spinal cord compression, and (C) circumferential spinal cord compression.

Figure 3

Overview of the convolutional neural network model architecture. The convolutional layers (orange) were derived from the Resnet-50 model, while the fully connected layers were modified for our classification task. Seven separate model configurations were tested as shown.

Figure 4

An area under the receiver operating characteristic curve plot showing model performance on each patient in the holdout dataset. The green curve represents the average ROC curve for all patients, while the light blue lines represent ROC curves for each individual patient. The gray region represents one standard deviation above and below the average curve.

Figure 5

Class activation maps for example images that were classified correctly (true positives) or incorrectly (false negatives). The first column represents the axial T2-weighted MRI image. The second column shows the MRI image with the class activation map overlaid. In the class activation maps blue represents no activation while red represents maximal activation. On these example images of true positive classifications, the maximal activation tended to be over the spinal canal and spinal cord (Images A,B,C,D,E,F). In the false negative classifications, the activation was sometimes over irrelevant areas of the image, such as the paraspinal muscles or vascular structures (Images G,I,K,L). In other examples of false negative classifications there was activation over the spinal cord and spinal canal (Images H and J), but in these cases there was also activation over other seemingly irrelevant areas of the image.