Feasibility of Fully Automatic Assessment of Cervical Canal Stenosis at MRI Using Deep Learning

Motivation: Artificial intelligence can provide a stable and objective cervical spine MRI diagnostic tool, thereby reducing doctors' labor. Goal(s): Evaluating the feasibility of fully automatic diagnosis of cervical spine MR imaging assisted by artificial intelligence. Approach: 890 patients participated in this multicenter retrospective study, which integrated a convolutional neural network (CNN) and Transformer for automated segmentation and classification of cervical spine MRI images, constructing a deep learning model, and comparing its diagnostic capabilities with those of human readers. Results: The model and readers achieved similar AUCs of 0.931 and 0.936 respectively. Correlation analysis show that the model, the readers had a moderate correlation. Impact: The deep learning (DL) model could fully automatic and reliably assess cervical canal stenosis at MRI and provide a stable and objective cervical spine MRI diagnostic tool, thereby reducing doctors' labor.