Deep learning models for MRI-based clinical decision support in cervical spine degenerative diseases

Purpose: The purpose of our study is to develop a deep learning (DL) model based on MRI and analyze its consistency with the treatment recommendations for degenerative cervical spine disorders provided by the spine surgeons at our hospital.

Methods: In this study, MRI of patients who were hospitalized for cervical spine degenerative disorders at our hospital from July 2023 to July 2024 were primarily collected. The dataset was divided into a training set, a validation set, and an external validation set. Four versions of the DL model were constructed. The external validation set was used to assess the consistency between the DL model and spine surgeons' recommendations about indication of cervical spine surgery regarding the dataset.

Results: This study collected a total of 756 MR images from 189 patients. The external validation set included 30 patients and a total of 120 MR images, consisting of 43 images for grade 0, 20 images for grade 1, and 57 images for grade 2. The region of interest (ROI) detection model completed the ROI detection task perfectly. For the binary classification (grades 0 and 1, 2), DL version 1 showed the best consistency with the spine surgeons, achieving a Cohen's Kappa value of 0.874. DL version 4 also achieved nearly perfect consistency, with a Cohen's Kappa value of 0.811. For the three-class classification, DL version 1 demonstrated the best consistency with the spine surgeons, achieving a Cohen's Kappa value of 0.743, while DL version 2 and DL version 4 also showed substantial consistency, with Cohen's Kappa values of 0.615 and 0.664, respectively.

Conclusion: We initially developed deep learning algorithms that can provide clinical recommendations based on cervical spine MRI. The algorithm shows substantial consistency with experienced spine surgeons.