Deep Learning-Based Segmentation and Quantification of Endolymphatic Hydrops in Inner Ear MRI

Deep Learning-Based Segmentation and Quantification of Endolymphatic Hydrops in Inner Ear MRI

🎓 Program: Master in Machine Learning for Health
🏫 Institution: Universidad Carlos III de Madrid
🤝 In collaboration with: Clínica Universidad de Navarra
👩‍🎓 Student: Laura Rodrigo Muñoz
📆 Academic Year: 2024/2025
📝 Grade: 10 / 10
🔗 Related thesis: Claudia Castrillón Álvarez – Classification and Localization of Inner Ear Structures

This Master Thesis presents a deep learning-based segmentation pipeline for the automated quantification of Endolymphatic Hydrops (EH) in patients with Ménière’s Disease, using high-resolution MRI scans. The goal was to compute the Endolymphatic Ratio (ELR) reliably, minimizing the need for manual segmentation by radiologists.

Using optimized 3D U-Net architectures, the system successfully segmented:

• Vestibular cavities
• Vestibular endolymph regions

These volumes were then used to automatically calculate the ELR. Various architectural enhancements and tuning strategies were explored to maximize generalization and performance.

🧪 Highlights

• Evaluation metrics included Dice Similarity Coefficient and Sensitivity
• Achieved high segmentation accuracy with reproducible ELR estimation
• Built for full automation of current manual workflows in clinical diagnosis
• This thesis complements a related project focused on classification and localization

Together, both theses create a complete deep learning pipeline for EH diagnosis support, contributing to faster and more consistent patient evaluation.