Development of a machine learning algorithm to identify cauda equina compression on MRI scans

• Published in: World neurosurgery 2025-01, p.123669
• Main Authors: Biswas, Sayan, Sarkar, Ved, MacArthur, Joshua Ian, Guo, Li, Deng, Xutao, Snowdon, Ella, Ahmed, Hamza, Tetlow, Callum, George, K Joshi
• Format: Article
• Language: English
• ISSN: 1878-8769; 1878-8769
• DOI: 10.1016/j.wneu.2025.123669

Cauda Equina Syndrome (CES) poses significant neurological risks if untreated. Diagnosis relies on clinical and radiological features. As the symptoms are often non specific and common, the diagnosis is usually made after a MRI scan. A huge number of MRI scans are done to exclude CES but nearly 80% of them will not have cauda equina syndrome. This study aimed to develop and validate a machine learning model for automated CES detection from MRI scans to enable faster triage of patients presenting with CES like clinical features.OBJECTIVECauda Equina Syndrome (CES) poses significant neurological risks if untreated. Diagnosis relies on clinical and radiological features. As the symptoms are often non specific and common, the diagnosis is usually made after a MRI scan. A huge number of MRI scans are done to exclude CES but nearly 80% of them will not have cauda equina syndrome. This study aimed to develop and validate a machine learning model for automated CES detection from MRI scans to enable faster triage of patients presenting with CES like clinical features.MRI scans from suspected CES patients (2017-2022) were collected and categorized into normal scans/disc protrusion (0%-50% canal stenosis (CS)) and cauda equina compression (CEC, >50% CS). A convolutional neural network was developed and tested on a total of 715 images (80:20 split) Gradient descent heatmaps were generated to highlight regions crucial for classification.METHODSMRI scans from suspected CES patients (2017-2022) were collected and categorized into normal scans/disc protrusion (0%-50% canal stenosis (CS)) and cauda equina compression (CEC, >50% CS). A convolutional neural network was developed and tested on a total of 715 images (80:20 split) Gradient descent heatmaps were generated to highlight regions crucial for classification.The model achieved an accuracy of 0.950 (0.921-0.971), a sensitivity of 0.969 (0.941-0.987), a specificity of 0.859 (0.742-0.937), a positive predictive value of 0.969 (0.944-0.984) and an area under the curve of 0.915 (0.865-0.958). Gradient descent heatmaps demonstrated accurate identification of any clinically relevant disc herniation into the spinal canal.RESULTSThe model achieved an accuracy of 0.950 (0.921-0.971), a sensitivity of 0.969 (0.941-0.987), a specificity of 0.859 (0.742-0.937), a positive predictive value of 0.969 (0.944-0.984) and an area under the curve of 0.915 (0.865-0.958). Gradient descent heatmaps demonstrated accurate identification of any