PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluation Principles for MRI-MRA Registration in Trigeminal Neuralgia: An ROI-Centered Neurovascular Benchmark

Xupeng Zhang, Xihang Wang, Michael Xie, Haoyuan Liang, Hau Ern Lien, Oishika Das, James Feghali, Risheng Xu, Peirong Liu

arXiv:2609.08805v1cs.CV

TL;DR

TN MRI–MRA fusion is poorly captured by generic whole-brain benchmarks despite its relevance to local neurovascular visualization. This paper benchmarks six registration pipelines with ROI-centered measures and stratified analyses. The results show that evaluation metrics can mislead in different ways, supporting a local, vessel-aware, contrast- and FOV-aware approach.

  • Problem

    TN MRI–MRA registration remains underexplored, and generic whole-brain benchmarks do not account for the local ROI, partial vessel annotations, variable MRA contrast, and FOV mismatch.

    Source: §1.3 · p. 2

  • Method

    The study benchmarks six registration pipelines in 149 patients with clinician-annotated bilateral trigeminal ROIs using image-based, vessel-localization, prediction-volume, contrast- and FOV-stratified, and reader-based analyses.

    Source: §5 · p. 10

  • Results

    Metrics did not co-rank methods: one-sided distances were affected by predicted vessel extent, vessel separability depended on contrast, and locally favorable distances could coexist with global misalignment; deformable refinement offered only a small, FOV-dependent benefit.

    Source: §4 · pp. 8–9, §5 · p. 10

  • Takeaways & Limitations

    TN MRI–MRA registration should be evaluated as a local, vessel-aware, contrast-sensitive, and FOV-aware visualization task rather than generic whole-brain registration.

    Source: §4 · p. 8, §5 · p. 10

    The retrospective, single-center cohort was predominantly Siemens 3T, and its clinically focused partial vessel annotations complicate geometric evaluation.

    Source: §4.1 · p. 10

Abstract

from arXiv · show

Preoperative evaluation of trigeminal neuralgia (TN) often requires joint interpretation of structural MRI, which depicts the trigeminal nerve and surrounding cisternal anatomy, and time-of-flight MRA, which highlights vascular structures. Although MRI-MRA fusion is clinically attractive for visualizing neurovascular compression, this task is poorly captured by conventional whole-brain registration evaluation because the clinically relevant target is a small trigeminal ROI, vessel annotations are partial and clinically focused, local TOF-MRA contrast is variable, and field-of-view mismatch can limit deformable alignment. We formulate TN MRI-MRA fusion as an ROI-centered neurovascular registration-evaluation problem and construct a benchmark from 149 patients with clinician-annotated bilateral trigeminal ROIs. Six representative registration pipelines were evaluated using local image-based metrics, segmentation-derived vessel-localization metrics, prediction-volume analysis, and contrast- and FOV-stratified comparisons. Conventional evaluation summaries were often misleading: local image similarity, vessel-background separability, and downstream vessel localization did not co-rank methods; one-sided vessel distances were strongly affected by predicted vessel extent under partial annotations; and local MRA contrast determined when vessel-separability metrics were informative. Deformable refinement provided only a small, FOV-dependent benefit over affine alignment, while reader review showed that locally favorable vessel distances could coexist with globally implausible registrations. These findings indicate that TN MRI-MRA registration should be evaluated as a local, vessel-aware, contrast-sensitive, and FOV-aware visualization task rather than as generic multimodal brain registration. Our code is publicly available at https://github.com/jhuldr/TN-Reg-Benchmark.

1. Introduction

TN MRI–MRA fusion addresses a local nerve–vessel visualization need that generic whole-brain registration benchmarks do not capture. The paper frames this as an ROI-centered evaluation problem and identifies the clinical, annotation, contrast, and field-of-view factors that complicate it.

1. Introduction

  • Preoperative TN assessment combines structural MRI, which depicts nerve and cisternal anatomy, with TOF-MRA, which highlights vessels; clinicians often review them side-by-side.The paper situates fusion within neurovascular compression assessment, where a vessel contacts or compresses the cisternal trigeminal nerve.Source: §1 · p. 1

1.2. Clinical Use Case

  • Structural MRI depicts the trigeminal nerve and cisternal anatomy, while TOF-MRA emphasizes vessels; their fusion may support assessment of nerve–vessel contact morphology.The technical goal is anatomically plausible multimodal visualization, not vessel-type determination or outcome prediction.Source: §1.1 · p. 1, §1.2 · p. 2

1.1. MRI–MRA Fusion as a Local Neurovascular Task

  • Whole-brain alignment and local vessel metrics can each mislead: global acceptability may conceal ROI failure, while favorable local distances may coexist with implausible global alignment.This motivates evaluating local neurovascular interpretability rather than whole-volume similarity alone.Source: §1.1 · p. 1

1.3. Why Conventional Registration Evaluation Is Insufficient

  • TN MRI–MRA co-registration remains underexplored, and conventional benchmarks do not address its small trigeminal ROI, partial vessel labels, variable MRA conspicuity, and FOV mismatch.These factors make one-sided distances sensitive to predicted vessel volume and can limit deformable refinement.Source: §1.3 · p. 2

1.4. Related Work

  • Prior work describes registration methods and TN fusion, but reliability for local TN visualization with FOV mismatch, variable conspicuity, and partial annotations remains unclear.Earlier fusion studies motivate multimodal visualization but do not define evaluation for local vessel interpretability in clinician-defined ROIs.Source: §1.4 · p. 2

1.5. Contributions

  • The paper formulates TN MRI–MRA fusion as ROI-centered evaluation and benchmarks six pipelines using complementary image-based and vessel-localization metrics.Predicted-volume, contrast- and FOV-stratified, and reader-based analyses probe confounds in conventional evaluation.Source: §1.5 · p. 2

2. Materials and Methods

The methods use a retrospective cohort with bilateral clinician-annotated trigeminal ROIs to compare six MRI–MRA registration pipelines. Evaluation combines local image and segmentation-based measures with contrast and FOV stratification and blinded reader review.

2.1. Clinical Cohort

  • The retrospective single-institution cohort included 149 TN patients with paired preoperative structural MRI and TOF-MRA and 298 bilateral trigeminal ROIs.Patients subsequently underwent microvascular decompression; inclusion required clinician-provided trigeminal annotations.Source: §2.1 · p. 2

2.2. Imaging Data and ROI Annotations

  • Structural MRI and TOF-MRA were paired with bilateral ROI annotations drawn on MRI alone, marking the clinically relevant vessel segment rather than an exhaustive vascular tree.The 48×48×48 ROI crops were sampled at 0.47 mm isotropic resolution.Source: §2.2 · pp. 2–3, Figure 1 · p. 3

2.3. Preprocessing and Registration Methods

  • MRI served as the fixed image and MRA as the moving image; six representative registration methods resampled MRA into MRI space.ANTs SyN, EasyReg, and SynthMorph included affine alignment, while ConvexAdam and FireANTs received ANTs Affine output as external pre-alignment.Source: §2.3 · p. 3

2.4. ROI-Centered Segmentation-Based Evaluation

  • After registration, warped MRA was sampled in the matched trigeminal ROI and VesselFM generated vessel predictions for comparison with clinician labels.Predicted vessel volume was tracked because prediction extent varied substantially and could affect distance metrics.Source: §2.4 · p. 3

2.5. Stratified Analyses

  • ROIs were stratified by local vessel-to-background contrast and worst-axis FOV coverage, with paired Affine-versus-SyN comparisons and patient-level bootstrap intervals.The contrast tiers were Low (≤1.1), Mid (>1.1 to ≤1.3), and High (>1.3); 12 ROIs lacked valid SyN-based contrast measurements.Source: §2.5 · p. 4

2.6. Blinded Reader Study

  • Two attending neurosurgeons jointly reviewed 100 stratified ROIs for clinical evaluability, vessel type, and confidence using MRI, original MRA, warped MRA, and ROI labels.Evaluability required an anatomically consistent, usable warped overlay; vessel type was assessed by tracing candidate vessels across slices and checking continuity.Source: §2.6 · p. 4

3. Experimental Results

Across 149 patients and 298 trigeminal ROIs, registration methods were ranked differently by local image and vessel-localization metrics. Their interpretation also depended on local contrast and field of view, with deformable refinement offering only a small, FOV-dependent benefit over affine alignment.

3.1. Evaluation Metrics

  • Within the clinician-defined ROI, Vessel AUC measures vessel-background separation, ROI NMI local MRI–MRA correspondence, and directional distances distinguish annotation coverage from prediction overreach under partial labels.The annotated vessel segment is clinically focused and partial, rather than an exhaustive vascular tree.Source: §3.1 · pp. 4–5

3.2. Cohort and Analysis Subsets

  • The cohort comprised 149 TN patients and 298 bilateral ROIs; image-based metrics used all 298, while vessel-distance analysis excluded ROIs with empty predictions.The FOV-stratified Affine-versus-SyN analysis included 223 ROIs after excluding empty predictions or out-of-coverage regions.Source: §3.2 · p. 5

3.3. Conventional Metrics Reveal Discordant Rankings

  • Vessel AUC, ROI NMI, and directional vessel distances produced discordant method rankings, showing that these metrics capture non-interchangeable aspects of registration quality.Fig. 2 shows that local vessel separability and intensity correspondence do not co-rank methods; Fig. 3 shows directional distances rank them differently alongside differences in predicted vessel volume.Source: Figure 2 · p. 5, Figure 3 · p. 5, §3.3 · p. 5

3.4. Predicted Volume Confounds One-Sided Distance Metrics

  • ρ = −0.53 for dann→Pred versus predicted volume, compared with ρ = +0.08 for dPred→ann; larger masks may improve annotation coverage without bringing predicted voxels closer.The symmetric-distance association was intermediate (ρ = −0.32), and remained a segmentation-derived localization metric rather than an independent registration-error measure.Source: §3.4 · pp. 5–6, Figure 4 · p. 6

3.6. Local Contrast Determines When Vessel AUC Is Informative

  • In Low contrast, Vessel AUC converged toward chance; in High contrast, SyN median AUC was approximately 0.76, while Mid-tier AUC was not significantly associated with localization (ρ = −0.12, p = 0.35).The contrast-stratified analysis included 181 Low, 66 Mid, and 39 High contrast ROIs; in the Mid tier, 36/65 valid SyN predictions missed the annotation under a 2-mm criterion.Source: §3.6 · pp. 6, 8

3.5. Visibility Does Not Guarantee Downstream Localization

  • Across the illustrated cases, high-contrast conspicuity aligned with vessel localization, but Mid-tier visibility could coexist with a displaced prediction and Low-contrast AUC was near chance.The examples distinguish whole-brain alignment, image-level conspicuity, and downstream localization as related but non-interchangeable evaluation axes.Source: Figure 5 · p. 7

3.7. FOV Mismatch Limits the Benefit of Deformable Refinement

  • In Good-FOV ROIs, SyN reduced dann→Pred by a median paired ∆= −0.18mm; this benefit was not observed in Bad-FOV ROIs (∆= −0.07mm, p = 0.32).The paired analysis included 177 Good-FOV and 46 Bad-FOV ROIs; the Bad-FOV interval was wide and case trajectories moved in both directions.Source: §3.7 · p. 8, Figure 7 · p. 8

4. Discussion

The discussion argues that TN MRI–MRA registration should be evaluated as a local, vessel-aware task, because metrics can disagree and favorable ROI distances may coexist with globally implausible alignment. Contrast, field-of-view coverage, and partial annotations qualify method comparisons and clinical interpretation.

4. Discussion

  • Under partial annotations, Vessel AUC, ROI NMI, and dsym rank methods differently, while one-sided distances can favor larger predicted vessel volumes.The discussion recommends reporting these metrics jointly with |VPred| rather than ranking methods by a single one-sided criterion.Source: §4 · p. 8
  • Among Mid-tier ROIs with AUCSyN > 0.60, 48.5% still missed the annotated vessel segment under the 2-mm criterion.Vessel AUC was not significantly associated with dann→Pred (ρ = −0.12, p = 0.35); downstream distances reflect the registration–segmentation pipeline.Source: §4 · p. 8
  • The Low-contrast tier comprised 63% of ROIs and drove Vessel AUC toward chance for all methods, making contrast stratification necessary to reveal method differences.The discussion treats contrast stratification as identifying when Vessel AUC is informative, not as evidence that contrast causally outweighs method choice.Source: §4 · p. 8
  • Deformable refinement gave a statistically significant but small benefit over affine alignment in Good-FOV ROIs (p < 0.001), with no significant benefit in Bad-FOV ROIs (p = 0.32).The bootstrap confidence interval nearly touched zero, and the discussion treats robust affine initialization as a prerequisite under partial modality overlap.Source: §4 · pp. 8–9
  • Reader review found that SynthMorph could have deceptively low dann→Pred despite global distortion, because predictions remained locally close to annotations.SynthMorph showed global shifts or failures in 71 of 100 reviewed ROIs; Fig. 8 contrasts this with ANTs SyN’s anatomically plausible alignment.Source: §4 · p. 9, Figure 8 · p. 9
  • ANTs SyN had the most balanced six-method profile, with the highest High-contrast Vessel AUC (∼0.76) and median |VPred| of 1493.It was not the lowest performer on any metric, though it required approximately 20 minutes per MRI–MRA pair.Source: §4 · p. 9, Table 4 · p. 9
  • When local contrast and FOV coverage are adequate, ANTs SyN is a reasonable offline baseline for local overlays, while MRI–MRA integration may support NVC morphology and prognosis assessment.TOF-MRA may help generate hypotheses about vessel identity, but venous interpretation remains non-definitive without complementary imaging or intraoperative confirmation.Source: §4 · pp. 9–10

4.1. Limitations

  • The retrospective, single-center cohort is predominantly Siemens 3T, and external validation in more heterogeneous TN cohorts would further assess generalizability.Partial vessel annotations also complicate geometric evaluation because predicted trees may extend beyond labeled segments.Source: §4.1 · p. 10

5. Conclusion

The benchmark finds that TN MRI–MRA fusion requires ROI-centered evaluation: image and vessel-localization signals do not align, and deformable gains are small and FOV-dependent. It recommends combining vessel-aware metrics with contrast, FOV, prediction-volume, and registration-quality checks.

  • The benchmark frames TN MRI–MRA fusion as ROI-centered neurovascular evaluation rather than generic whole-brain alignment.Source: §5 · p. 10
  • Across 149 patients and 298 ROIs, six pipelines showed non-interchangeable evaluation signals, extent-confounded one-sided distances, contrast-dependent AUC informativeness, and small FOV-dependent deformable gains.Source: §5 · p. 10
  • Evaluation should jointly report image and segmentation metrics, prediction volume, contrast, FOV compatibility, and registration quality control.Source: §5 · p. 10

Supplementary Material

The supplement details cohort selection, imaging acquisition, ROI annotation, registration and vessel-inference settings, and stratification and statistical-analysis procedures.

S1. Cohort Selection and Demographics

  • Paired same-encounter MRI–TOF-MRA and clinician annotations were required for this retrospective single-institution MVD cohort; 149 patients yielded 298 bilateral ROIs after exclusions.Source: S1. Cohort Selection and Demographics · p. 13

S2. Imaging Acquisition Parameters

  • CISS-type MRI depicted nerve and soft tissue, while 3D TOF-MRA depicted vessels; median in-plane spacing was 0.597 mm and 0.260 mm, respectively.Median slice thickness was 0.60 mm for MRI and 0.50 mm for MRA.Source: S2. Imaging Acquisition Parameters · p. 13

S3. ROI Annotation Protocol

  • Two-pass MRI-only labels on 48 × 48 × 48 nerve-centered crops mark clinically relevant contact or nearby vessels, not exhaustive trees; consensus review limits direct inter-annotator variability estimates.Source: S3. ROI Annotation Protocol · p. 13

S4. Registration Settings and Computing Resources

S5. VesselFM Inference and Postprocessing

  • VesselFM v1.0 segmented each warped MRA without retraining; probability maps were thresholded at 0.5, components under 500 voxels removed, and predictions compared within trigeminal ROIs.Source: S5. VesselFM Inference and… · p. 14

S6. Contrast Stratification and Exclusions

S7. Descriptive Thresholds and Statistical Analysis

  • AUC >0.60, distance >2 mm, and rmin ≤0.76 thresholds were descriptive; paired tests used patient-level bootstrap confidence intervals without multiplicity adjustment.Source: S7. Descriptive Thresholds and… · p. 15
Loading 2609.08805v1…