Deep Learning Model for Automated Detection and Classification of Central Canal and Neural Foraminal Stenosis at Cervical Spine MRI

Abstract

Abstract Objectives: Developing a Deep learning (DL) model to automatically detect and classify cervical canal and neural foraminal stenosis on cervical spine MRI can improve the accuracy and efficiency of its diagnostic. Methods: A method for cervical spinal stenosis was proposed based on the DL model, consisting of region of interest (ROI) detection and cascade prediction. First, three part-specific convolutional neural networks were used to detect the ROIs in different parts of cervical MRI images. Then, the cascade prediction of stenosis categories was performed to obtain the results of stenosis level and position on each slice of the patients. Finally, in the testing, the results were fused to obtain a patient-level diagnostic report. The performance was evaluated with the metrics of accuracy (ACC), area under curve (AUC), sensitivity, specificity, F1 Score, and diagnosis time of the DL model, as well as recall rate for ROI detection localization. Results: The average recall rate of ROIs localization reached 89.3% (neural foramen), and 99.7% (central canal) under the five-fold cross-validation of our DL model. In dichotomous classification (normal or mild vs moderate or severe), ACC and AUC of the DL model were very close to the level of radiologists, and the F1 score (84.8%) of the DL model was slightly higher than that of radiologists (83.8%) at central canal. Conclusion: The DL model showed comparable performance with subspecialist radiologists for detection and classification of the central canal and neural foraminal stenosis at cervical spine MRI with significant time-saving.