Neurosurgery 97:242–249, 2025
Automated ventricular volume analysis using convolutional neural networks and radiomics, combined with clinical features, improves accuracy in diagnosing hydrocephalus and shunt malfunction from CT scans, outperforming traditional 2D methods and enabling earlier, more reliable detection without requiring prior imaging.
• Automated ventricular volume analysis and radiomics were used to predict shunt malfunction and diagnose hydrocephalus from CT scans.
• Traditional 2D measurements are less sensitive than 3D volumetric analysis for detecting ventricular changes.
• A convolutional neural network (CNN) segmented ventricles, and volumes were compared to age- and sex-matched normative data.
• Volumetric analysis alone achieved up to 73% accuracy (AUC 0.772) for detecting pathology; combining radiomics and clinical data improved accuracy significantly.
• The best model (support vector machine) using selected radiomics features and clinical data reached an AUC of 0.92–0.93 and F1-score of 0.848.
• Automated segmentation is time-efficient, reduces observer variability, and may improve early and accurate diagnosis.
• Limitations include single-institution data, scanner variability, and need for external validation.
• Future work should expand datasets and improve model generalizability, including to MRI and other manufacturers.
You must be logged in to post a comment.