Explainable 3D VGG-style convolutional neural network for pediatric hydrocephalus detection on computed tomography: A segmentation-free and fully volumetric deep learning framework.

Pediatric hydrocephalus is commonly assessed on computed tomography (CT) using manual two-dimensional indices that incompletely reflect the ventricular system. We developed and evaluated an explainable, segmentation-free three-dimensional convolutional neural network (3D CNN) for automated hydrocephalus detection on pediatric CT. Ninety-eight pediatric CT scans were included (48 hydrocephalus, 50 normal), distributed across four age groups: <1 year (n = 24), 1-5 years (n = 24), 5-10 years (n = 19), and 10-15 years (n = 31). Diagnoses were established by experienced radiologists using Evans Index ≥ 0.30 with consensus resolution. A lightweight VGG-style 3D CNN was trained from scratch using five-fold StratifiedGroupKFold cross-validation. Performance was assessed using AUC-ROC, accuracy, sensitivity, specificity, F1-score, Matthews correlation coefficient (MCC), Cohen's κ, and Brier score. Explainability was evaluated using 3D Grad-CAM and four quantitative attention metrics. The model achieved high internal performance, with mean AUC-ROC 0.95 ± 0.06, accuracy 0.94 ± 0.06, F1-score 0.94 ± 0.06, and MCC 0.88 ± 0.11; the pooled fold-averaged ROC yielded AUC 0.944. Grad-CAM analyses showed progressively stronger, anatomically coherent ventricular activation from normal to severe cases. In preliminary external validation on 118 independent pediatric CT scans (56 hydrocephalus, 62 normal), AUC-ROC was 0.88 and AUC-PR was 0.85; accuracy, sensitivity, specificity, precision, and F1-score were 0.83, 0.84, 0.82, 0.81, and 0.82, respectively. This segmentation-free 3D CNN provides accurate, calibrated, and anatomically interpretable hydrocephalus detection directly from routine pediatric CT, although larger multicenter validation remains necessary before clinical deployment.