YOLOv11 Models: Enhancing CT-Based Diagnosis of Cervical Foraminal Stenosis Through Deep Learning

This study aims to construct a deep learning (DL) system from the YOLOv11 family by using computed tomography (CT) to overcome the scarcity of magnetic resonance imaging (MRI) availability and the standard imaging assessments to diagnose cervical foramina stenosis (CFS). This retrospective study involved 1,437 cervical spine CT scans from three centers (Center A, B, and C) of a local hospital. The training and validation were done using scans from Centers A and B, with Center C used as the external test set. A stenosis detection model, based on the YOLOv11 architecture with Multiplanar Reconstruction (MPR) technology, was employed. The model was trained on 5-fold cross-validated images (640×640 pixels) in five variants (n, s, m, l, x) that were labeled. Five radiologists analyzed 215 images from Center C, and performance metrics were determined. Statistical analysis was conducted using paired t-tests and McNemar tests, with Bonferroni and FDR corrections. In external testing (n = 215), the YOLOv11-x model had higher diagnostic performance at detecting cervical foraminal stenosis: sensitivity was 95.6% (p = 0.005), specificity was 97.4% (p < 0.001), and accuracy was 96.7% (p < 0.001). YOLOv11-x has demonstrated high performance in standardized CT-based diagnosis of CFS, and lightweight versions could be used in real-time clinical applications. This technique presents a deep learning modality, which is effective as an alternative diagnostic tool in settings with limited MRI resources. Future validation and correlation with clinical outcomes should verify the generalizability of treatment responses and symptom severity results.