Endolymphatic-to-Vestibular Volume Ratios (ELR)

• ELR is a quantitative biomarker defined as the percentage of endolymph volume relative to vestibular volume, crucial for assessing inner ear disorders.
• Automated ELR quantification via deep learning modules (EarGate, AuriBox, EHMasker) improves reproducibility and reduces operator dependency.
• Performance metrics show high segmentation accuracy and physiologically realistic ELR values, aiding in precise diagnosis of Ménière's disease.

The endolymphatic-to-vestibular volume ratio (ELR) is a volumetric biomarker quantifying the proportion of endolymphatic space within the vestibular cavity of the inner ear. It is operationalized as the percentage ratio of endolymph volume to total vestibular volume, obtained through MRI-based 3D segmentation. ELR serves as a direct, quantitative measure for diagnosing and grading endolymphatic hydrops (EH), a pathological hallmark of disorders such as Ménière's disease. Recent advancements in deep learning have enabled automated, reproducible ELR estimation that is less operator-dependent and more physiologically plausible.

1. Mathematical Definition and Computation

ELR is defined by the following equation: $$\text{ELR}~[\%] = \frac{V_{\mathrm{endolymph}}}{V_{\mathrm{vestibule}}} \times 100$$ where $V_{\mathrm{endolymph}}$ and $V_{\mathrm{vestibule}}$ are the physical volumes of the endolymphatic space and the vestibule, respectively. These volumes are calculated by:

• Generating 3D binary masks from MRI, with modality-specific segmentation (SPACE-MRC for vestibule; REAL-IR for endolymph).
• Counting foreground voxels in each mask.
• Multiplying by the corresponding voxel size (SPACE-MRC: $0.125\,\mathrm{mm}^3$; REAL-IR: $0.200\,\mathrm{mm}^3$).
• Converting sum of voxel volumes to cm³ or mL ($1~\mathrm{cm}^3 = 1000~\mathrm{mm}^3$).

This operationalization requires precise, sequence-matched segmentation and rigorous voxel accounting to ensure accuracy and reproducibility (Fuster-Barceló et al., 26 Jan 2026).

2. Automated Pipeline for ELR Quantification

The OREHAS pipeline represents the first fully automated workflow for ELR quantification from routine 3D-SPACE-MRC and 3D-REAL-IR MRI. It integrates three deep learning modules:

1. EarGate (Slice Classification): Discards slices lacking inner-ear anatomy using a custom 5-layer CNN or an ImageNet-fine-tuned ResNet-50; post-processing enforces block continuity and corrects isolated errors.
2. AuriBox (Inner-Ear Localization): Detects and crops 2D regions around left/right ears using a YOLOv5su model trained with centroid-annotated bounding boxes.
3. EHMasker (Sequence-Specific Segmentation): Applies a 2D U-Net to generate binary masks for the vestibular cavity and endolymphatic space, leveraging hybrid BCE + Dice loss and minimal expert annotation (only 3–6 slices per ear).

Volumes are then reconstructed from stacked, post-processed slice masks; ELR is computed as above. The overall system eliminates the need for manual slice delineation and operator intervention while ensuring methodological consistency (Fuster-Barceló et al., 26 Jan 2026).

3. Training Regimen and Data Requirements

OREHAS is trained on a cohort of 90 subjects (83 with Ménière's disease, 7 controls), each with both SPACE-MRC and REAL-IR sequences. Annotation requirements are minimal: only 3–6 axial slices per ear are needed, resulting in 437 SPACE-MRC and 409 REAL-IR annotated slices in total. The pipeline generalizes from sparse slice-level annotations to volumetric reconstructions across 10–15 slices per volume, drastically reducing expert effort compared to labor-intensive manual workflows, which typically require contiguous slice-by-slice delineation and proprietary spatial interpolation (Fuster-Barceló et al., 26 Jan 2026).

4. Performance Metrics and Validation

Quantitative assessment of OREHAS demonstrates:

• Segmentation Accuracy (2D Patches):
  • Vestibular (SPACE-MRC): Dice = 0.91, IoU = 0.83, Recall = 0.92.
  • Endolymph (REAL-IR): Dice = 0.75, IoU = 0.56, Recall = 0.74.
• Volumetric Similarity Index (VSI) vs. Manual Ground Truth (External N = 5):
  • Vestibule: OREHAS VSI = 86.1% ± 4%; Syngo.via VSI = 88.8% ± 4%.
  • Endolymph: OREHAS VSI = 74.3% ± 22%; Syngo.via VSI = 42.5% ± 17%.
• End-to-End ELR on Held-Out Test Set (N = 19):
  • Vestibular volumes (cm³): Syngo.via 0.083 ± 0.013; OREHAS 0.084 ± 0.010.
  • Endolymph volumes (cm³): Syngo.via 0.043 ± 0.020; OREHAS 0.019 ± 0.008.
  • ELR (%): Right ear—Syngo.via 46.0 ± 22.2, OREHAS 20.2 ± 9.0; Left ear—Syngo.via 57.6 ± 26.1, OREHAS 25.7 ± 11.6.

OREHAS achieves segmentation and volumetric agreement with expert ground truth and delivers more physiologically realistic ELR values, correcting overestimations inherent in manual and interpolative workflows (Fuster-Barceló et al., 26 Jan 2026).

5. Implementation Considerations and Volumetric Workflow

Volume computation is executed by stacking sequence-specific slice masks, counting foreground voxels, and applying modality-dependent scaling. The following workflow is used:

Step	SPACE-MRC (Vestibular)	REAL-IR (Endolymph)
Voxel size	$0.5 \times 0.5 \times 0.5$ mm$^3$	$0.5 \times 0.5 \times 0.8$ mm$^3$
Voxel volume	$0.125$ mm$^3$	$0.200$ mm$^3$
Binary mask	$M_\mathrm{SPACE}$	$M_\mathrm{REAL}$
Slice stacking	Axial order	Axial order
Post-processing	Largest component, hole fill, threshold	Largest component, hole fill, thresh

This rigorous, pipeline-based approach yields consistent, volumetric ELR measurements compatible with established imaging protocols and suitable for large-scale studies (Fuster-Barceló et al., 26 Jan 2026).

6. Clinical Interpretation, Thresholds, and Physiological Realism

Established diagnostic thresholds, based on syngo.via-derived ELR:

• ELR < 30%: Non-significant hydrops
• ELR > 60%: Significant hydrops

OREHAS measurements are systematically lower than those obtained via syngo.via, implying that clinical cutoffs may require recalibration to prevent overdiagnosis of EH that arises from interpolation artifacts or volume overestimation. Physiological plausibility is improved with OREHAS, yielding ELR values strictly within $[0, 100)\%$ for all cases, whereas syngo.via sometimes reports biologically impossible ELR > 100% (Fuster-Barceló et al., 26 Jan 2026). The approach supports the refinement of clinical criteria and more accurate phenotyping in research and practice.

7. Reproducibility, Scalability, and Open Science

By requiring only a small number of slice-level annotations and automating all subsequent steps, OREHAS ensures reproducible, transparent ELR quantification. The open-source codebase facilitates rapid, batch processing (<2 min for 10 patients on a single GPU) and supports integration within research and clinical workflows. This robust foundation enables standardized data collection across cohorts, supports recalibration of diagnostic thresholds, and advances large-scale studies of inner ear pathology (Fuster-Barceló et al., 26 Jan 2026).