Deep learning-based prediction of cervical canal stenosis from mid-sagittal T2-weighted MRI.

Abstract

Objective: This study aims to establish a large degenerative cervical myelopathy cohort and develop deep learning models for predicting cervical canal stenosis from sagittal T2-weighted MRI.

Materials and methods: Data was collected retrospectively from patients who underwent a cervical spine MRI from January 2007 to December 2022 at a single institution. Ground truth labels for cervical canal stenosis were obtained from sagittal T2-weighted MRI using Kang's grade, a four-level scoring system that classifies stenosis with the degree of subarachnoid space obliteration and cord indentation. ResNet50, VGG16, MobileNetV3, and EfficientNetV2 were trained using threefold cross-validation, and the models exhibiting the largest area under the receiver operating characteristic curve (AUC) were selected to produce the ensemble model. Gradient-weighted class activation mapping was adopted for qualitative assessment. Models that incorporate demographic features were trained, and their corresponding AUCs on the test set were evaluated.

Results: Of 8676 patients, 7645 were eligible for developing deep learning models, where 6880 (mean age, 56.0 ± 14.3 years, 3480 men) were used for training while 765 (mean age, 56.5 ± 14.4 years, 386 men) were set aside for testing. The ensemble model exhibited the largest AUC of 0.95 (0.94-0.97). Accuracy was 0.875 (0.851-0.898), sensitivity was 0.885 (0.855-0.915), and specificity was 0.861 (0.824-0.898). Qualitative analyses demonstrated that the models accurately pinpoint radiologic findings suggestive of cervical canal stenosis and myelopathy. Incorporation of demographic features did not result in a gain of AUC.

Conclusion: We have developed deep learning models from a large degenerative cervical myelopathy cohort and thoroughly explored their robustness and explainability.