IE-Vnet: Deep Learning-Based Segmentation of the Inner Ear's Total Fluid Space

Area of Science:

• Medical Imaging
• Deep Learning
• Neuroscience

Background:

• Accurate quantification of endolymphatic hydrops (ELH) via in-vivo MR imaging requires reliable segmentation of the inner ear's total fluid space (TFS).
• Current methods are often limited in accuracy and speed, hindering large-scale research.

Purpose of the Study:

• To develop and validate a novel, open-source deep learning (DL) model for segmenting the inner ear's TFS.
• To improve the accuracy and efficiency of ELH quantification.

Main Methods:

• A V-Net based deep learning model (IE-Vnet) was trained on multivariate MR scans (T1, T2, FLAIR, SPACE) from 179 patients.
• Ground-truth TFS masks were generated using a semi-manual, atlas-assisted approach.
• The model's performance was evaluated on four independent datasets (80 ears total) for accuracy, generalizability, and robustness.

Main Results:

• IE-Vnet achieved high segmentation accuracy (Dice: 0.9 ± 0.02) and surface distance metrics (Hausdorff: 0.93 ± 0.71 mm).
• Performance was consistent across different datasets and ear sides (p>0.05).
• Prediction time was significantly reduced (0.2 s), being 2,000 times faster than atlas-based methods.

Conclusions:

• The IE-Vnet model provides accurate, robust, and rapid TFS segmentation for ELH studies.
• It integrates seamlessly with existing open-source pipelines for automatic ELH quantification.
• This tool facilitates high-volume, multi-institutional inner ear research, with code freely available.