Application of deep neural networks in automatized ventriculometry and segmentation of the aqueduct in pediatric hydrocephalus patients

TL;DR

This paper compares two deep learning models for automatically measuring brain ventricles in children with hydrocephalus, finding that one model is more reliable and easier to use clinically.

Contribution

The study evaluates and compares VParNet and nnU-Net for pediatric ventricular segmentation, highlighting nnU-Net's advantages in handling complex cases without preprocessing.

Findings

Both VParNet and nnU-Net achieved high segmentation accuracy (DSC: 0.87–0.95) for pediatric ventricles.

nnU-Net outperformed VParNet in handling challenging cases due to no preprocessing requirements.

Aqueduct segmentation remained difficult, with nnU-Net achieving a DSC of 0.68.

Abstract

This study validated VParNet and nnU-Net for ventricular segmentation in pediatric hydrocephalus, a condition characterized by irregular and asymmetric ventricular shapes.

Manual segmentation of 139 MRI scans (ages range 2.6–20.3 years) was performed for the four ventricles and the aqueduct. A five-fold cross-validation was conducted for both models. VParNet was tested with its original weights and after retraining on pediatric data. nnU-Net was extended to also segment the aqueduct. Performance was evaluated using the Dice Similarity Coefficient (DSC), Intraclass Correlation Coefficient (ICC), and Minimal Detectable Change (MDC).

VParNet preprocessing failed in 20.9% of cases, requiring subject exclusion. Both models showed good to excellent segmentation accuracy and reliability (DSC: 0.87–0.95; ICC: 0.81–1.0). Retraining VParNet improved DSC scores. MDC values (0.05–3.0) indicated high sensitivity for the lateral and third ventricles and acceptable sensitivity for the fourth ventricle. Aqueduct segmentation remained challenging (nnU-Net: DSC = 0.68; ICC = 0.81; MDC = 0.04).

All tested models performed well in pediatric hydrocephalus segmentation, with no fundamental differences in overall performance. However, nnU-Net demonstrated key advantages due to its lack of preprocessing requirements, which allow the successful handling of even the most challenging subjects. These features make it easily implementable for clinical applications, providing fast and reliable ventricular segmentation and quantification.

The online version contains supplementary material available at 10.1007/s00234-025-03848-y.

Figures