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

Comprehensive Summary

In this study, Rhee et al. used a variety of deep learning (DL) architectures to predict CCS (cervical canal stenosis) solely from sagittal T2-weighted MRI images, as well as gradient-weighted class activation mapping (Grad-CAM) techniques to evaluate the explainability of each architecture. ResNet50, VGG16, MobileNetV3, EfficientNetV2, and an ensemble were trained with a cohort of 5 to 10 times the typical size compared to preliminary studies. Image pre-processing was done with the OpenCV-Python library, and reference standards were obtained from a spine specialist with >5 years of experience in spine surgery, using Kang’s grading classification. Nearly all models performed similarly, with the highest AUC (area under the receiver operating characteristic curve) value being 0.96 for the ensemble model. No improvement was noted when demographic data, such as Age and Sex were integrated into the data provided for each model. Grad-CAM analysis showed incongruity in explainability between models, but there was a general understanding of critical locations on MRI to base decision-making on. Overall, all models show generally even capability based on AUC values, with a difference in explainability as seen by Grad-CAM analysis. In comparison to previous studies with redundant data and significantly smaller cohorts, this study used an appropriately large cohort dataset to train and test models on.

Outcomes and Implications

Degenerative cervical myelopathy is an age-related disorder that typically results in neurological impairment that can impact daily activities of living and require surgical intervention. The common method of diagnosis is through the identification of cervical canal stenosis on MRI. The Kang grading system for CCS using sagittal T2-weighted images shows high inter- and intra-observer reliability and high correlation with neurological symptoms. Previous studies have examined the potential for deep learning AI models to be used for the radiological diagnostic process, but have been restricted by their low generalizability as a result of small sample size and the single use of models. This study improves on previous studies by using T2-weighted images from a significantly larger cohort of patients alongside a variety of models and an ensemble model, paired with an assessment of generalizability, to assess the capability of deep learning models. The significance of this study is in its generalizability and proof of concept for deep learning in radiological assessment, and it demonstrates the potential for future implementation and further research into deep learning AI in radiographic diagnostics.