Attention-Driven Deep Learning for Hydrocephalus: Preliminary Benchmarking of Multiscale 3D Attention Against Conventional 3D CNN Models Using Ventricular Enlargement on CT.

Hydrocephalus is characterized by abnormal cerebrospinal fluid accumulation causing ventricular enlargement, which, if undetected, can lead to neurological damage. Conventional two-dimensional CT indices often underestimate volumetric deformation. Deep learning offers potential for automated three-dimensional (3D) assessment. This study aimed to develop and benchmark a multiscale attention 3D convolutional neural network (MSA3DNet) against a conventional 3D-CNN for CT-based hydrocephalus classification based solely on ventricular enlargement. A total of 98 anonymized pediatric head CT examinations (0-15 years; 48 hydrocephalus, 50 normal) were retrospectively analyzed. DICOM data were converted to 3D NIfTI format and preprocessed through Hounsfield-unit clamping, brain-windowing, Z-score normalization and isotropic resampling (1.0 × 1.0 × 2.0 mm3). Both models were trained under identical augmentation and optimization settings using fivefold stratified group cross-validation. Quantitative evaluation included accuracy, sensitivity, specificity, F1-score, Matthews correlation coefficient (MCC) and area under the ROC curve (AUC). MSA3DNet achieved superior discriminative performance compared with the baseline 3D CNN (AUC = 0.971 ± 0.034 vs 0.952 ± 0.057; F1 = 0.948 ± 0.064 vs. 0.936 ± 0.068; accuracy = 0.948 ± 0.064 vs. 0.939 ± 0.066). Sensitivity improved by 4.2 pp (0.958 vs. 0.916) with only a marginal reduction in specificity (0.940 vs 0.960). Interfold variability was low (CV ≤ 10%), confirming robust reliability and generalization. Integrating multiscale encoding and attention mechanisms enhances volumetric CT classification of ventricular enlargement, yielding improved sensitivity and stability without loss of specificity. The proposed MSA3DNet framework shows potential as a reliable tool for automated pediatric hydrocephalus screening.