Developing Deep Learning-Based Cerebral Ventricle Auto-Segmentation System and Clinical Application for the Evaluation of Ventriculomegaly.

Objective: Current methods for evaluating ventriculomegaly, particularly Evans' Index (EI), fail to accurately assess three-dimensional ventricular changes. We developed and validated an automated multiclass segmentation system for precise volumetric assessment, simultaneously segmenting 5 anatomical classes (ventricles, parenchyma, skull, skin, and hemorrhage) to support future augmented reality-guided external ventricular drainage (EVD) systems.

Methods: Using the nnUNet architecture, we trained our model on 288 brain computed tomography scans with diverse pathological conditions and validated it using internal (n = 10), external (n = 43), and public (n = 192) datasets. Clinical validation involved 227 patients who underwent cerebrospinal fluid (CSF) drainage procedures. We compared automated volumetric measurements against traditional EI measurements and actual CSF drainage volumes in surgical cases.

Results: The model achieved exceptional performance with a mean Dice similarity coefficient of 93.0% across all 5 classes, demonstrating consistent performance across institutional and public datasets, with particularly robust ventricle segmentation (92.5%). Clinical validation revealed EI was the strongest single predictor of ventricular volume (adjusted R = 0.430; P < 0.001), though influenced by age, sex, and diagnosis type. Most significantly, in EVD cases, automated volume differences showed remarkable correlation with actual CSF drainage amounts (β = 0.956; adjusted R = 0.936; P < 0.001), validating the system's accuracy in measuring real CSF volume changes.

Conclusions: Our comprehensive multiclass segmentation system offers a superior alternative to traditional measurements with potential for noninvasive CSF dynamics monitoring and augmented reality-guided EVD placement.