IE-Vnet

IE-Vnet segments and quantifies endolymphatic hydrops (ELH) in the inner ear from in-vivo MR images using a V-Net deep learning model for volumetric total fluid space (TFS) analysis.

Key Features:

• Deep Learning Architecture: Implements a V-Net architecture optimized for high-resolution TFS segmentation.
• Multivariate Training Dataset: Trained on dataset D1 comprising 179 consecutive patients with peripheral vestibulocochlear syndromes using MR sequences T1, T2, FLAIR, and SPACE.
• Ground-Truth Generation: Uses semi-manual, atlas-assisted generation of ground-truth TFS masks for training and evaluation.
• Performance Metrics: Achieved Dice overlap coefficient 0.9 ± 0.02, Hausdorff maximum surface distance 0.93 ± 0.71 mm, and mean surface distance 0.022 ± 0.005 mm on heterogeneous test datasets D2–D5 (20 ears each).
• Robustness and Generalizability: Showed no significant differences across sides or datasets (p > 0.05), indicating resistance to domain shift.
• Efficiency: Produces segmentation predictions in 0.2 seconds per prediction, approximately 2,000 times faster than atlas-based methods.
• Integration Capability: Outputs integrate with an existing open-source pipeline for automatic endolymphatic space (ELS) segmentation.

Scientific Applications:

• Large-scale ELH studies: Enables high-volume, trans-institutional quantitative studies of inner ear fluid spaces for research on ELH and vestibulocochlear disorders.

Methodology:

Uses a V-Net segmentation model trained on D1 (179 patients; T1, T2, FLAIR, SPACE MR sequences) with semi-manual atlas-assisted ground-truth TFS masks, evaluated on test datasets D2–D5 (20 ears each) using Dice, Hausdorff maximum surface distance, mean surface distance, and statistical testing (p > 0.05), with inference time measured at 0.2 s per prediction.