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

AI summary

Overview

Research area: Medical image analysis — multimodal image registration and its evaluation, applied to neurosurgical planning for trigeminal neuralgia.

Technical level: Intermediate. The methods are established (ANTs, SynthMorph, ConvexAdam, VesselFM), but the paper's contribution is conceptual and evaluative rather than algorithmic, so readers need some familiarity with registration and segmentation metrics.

Scope: A retrospective 149-patient benchmark that reframes MRI–MRA co-registration for trigeminal neuralgia as a local, vessel-aware evaluation problem rather than a whole-brain alignment task.

What This Paper Is About

In trigeminal neuralgia, surgeons need to see where a blood vessel touches the trigeminal nerve, which requires combining structural MRI (which shows the nerve) with TOF-MRA (which shows vessels). The clinically relevant region is tiny — a small volume around the nerve — and existing whole-brain registration benchmarks do not measure whether fused images are actually usable at that scale. This paper builds a clinician-annotated benchmark to determine how registration quality should be measured when the goal is local neurovascular visualization, and shows that common metrics can rank methods in misleading ways.

Key Contributions

1. Problem reformulation: Reframes TN MRI–MRA fusion as an ROI-centered, vessel-aware, contrast-sensitive, FOV-aware registration-evaluation problem, distinct from generic multimodal brain registration.

2. Clinical benchmark: Assembles 149 TN patients (298 bilateral trigeminal ROIs) with clinician-annotated per-voxel nerve and vessel labels drawn on MRI alone, independent of MRA and registration outputs.

3. Multi-axis evaluation: Evaluates six registration pipelines (ANTs Affine, ANTs SyN, ConvexAdam, FireANTs, EasyReg, SynthMorph) using complementary image-based metrics, segmentation-derived vessel localization metrics, predicted-volume analysis, and stratified contrast/FOV comparisons.

4. Confound diagnosis: Quantifies how partial vessel annotations and prediction extent can systematically bias one-sided distance metrics, and demonstrates via a blinded reader study that favorable local metrics can coexist with globally unusable registrations.

Main Findings

• Metrics disagree with each other. Vessel AUC (vessel-vs-background intensity separability) ranked EasyReg (0.578) and SynthMorph (0.581) highest, while ROI NMI (local multimodal intensity correspondence) ranked FireANTs highest (1.072). Directional vessel distances also ranked methods in opposite directions: SynthMorph and FireANTs looked best under annotation-to-prediction distance but worst under prediction-to-annotation distance.

• Predicted vessel volume confounds one-sided distances. Annotation-to-prediction distance correlated strongly with predicted vessel volume (Spearman ρ = −0.53), meaning methods producing larger vessel masks are rewarded regardless of geometric accuracy. Prediction-to-annotation distance showed essentially no such coupling (ρ = +0.08). Prediction volumes varied by more than an order of magnitude across methods (median 1148 to 2876 voxels).

• Local contrast determines when Vessel AUC is informative. 63% of ROIs fell in a Low-contrast tier where AUC collapsed toward chance for all methods. Clear method separation appeared only in Mid and High tiers, where ANTs SyN reached ~0.76 median AUC.

• Visibility does not guarantee localization. In the Mid tier, ANTs SyN achieved median Vessel AUC of 0.600, yet 55.4% of ROIs missed the annotated vessel under a 2 mm criterion. Even among ROIs with AUC > 0.60, 48.5% missed. Vessel AUC and annotation-to-prediction distance were not significantly associated within this tier (ρ = −0.12, p = 0.35).

• Deformable refinement gives only a small, FOV-dependent gain. In Good-FOV cases, SyN modestly reduced annotation-to-prediction distance versus affine (median Δ = −0.18 mm, p < 0.001), but this benefit disappeared in Bad-FOV cases (Δ = −0.07 mm, p = 0.32).

• Reader study suggests venous biology underlies some Low-contrast ROIs. Of 100 reviewer-assessed ROIs, 66 were clinically evaluable; among those, 23 (34.8%) had some venous component, consistent with TOF-MRA's preferential arterial depiction rather than simple acquisition failure.

• Global failure can hide behind good local numbers. SynthMorph produced globally shifted or failed registrations in 71 of 100 reviewed ROIs, yet sometimes yielded low local distances because predicted vessels happened to fall near the annotation within the small ROI. ANTs SyN showed the most balanced overall profile, though at ~20 minutes per pair.

Methodology in Plain English

The researchers collected preoperative MRI and MRA scans from 149 patients who later underwent microvascular decompression surgery. Neurosurgeons manually outlined the trigeminal nerve and the clinically relevant vessel segment in small 48×48×48 voxel boxes centered on the nerve, working only from the structural MRI so the annotations would not be biased by the registration results.

They then ran six different registration algorithms, each warping the MRA into MRI space, and cropped the same trigeminal ROI from every warped output. Because there is no ground-truth deformation to compare against, they measured quality indirectly: how well MRA intensity separates annotated vessel voxels from background (Vessel AUC), how well MRI and MRA intensities correspond locally (ROI NMI), and how close a standard vessel-segmentation model's predictions land to the clinician annotations.

The key analytical move was stratification. They grouped ROIs by local vessel-to-background contrast and by how well the MRI and MRA fields of view overlapped, then checked whether metric behavior changed across these groups. They also tracked predicted vessel volume as a potential confounder. Finally, two attending neurosurgeons reviewed 100 stratified ROIs blinded to method identity, rating clinical evaluability and vessel type.

Why This Matters

Impact on research: The paper argues that TN-specific registration cannot be validated with a single whole-brain similarity score. It gives the field a concrete set of evaluation principles — report vessel AUC, ROI NMI, and symmetric vessel distance together with predicted volume, stratify by contrast and FOV, and interpret local metrics alongside global plausibility — and releases a public benchmark and code.

Real-world applications:

• Preoperative planning for microvascular decompression, where accurate nerve-vessel visualization could inform surgical targeting.
• Quality control for clinical image-fusion pipelines, flagging cases where a locally plausible overlay actually sits on a globally distorted registration.
• Design of TN-specific imaging protocols, since FOV mismatch between MRI and MRA demonstrably limits deformable alignment.
• Morphologic NVC assessment — contact location, severity, and arterial versus venous type — which has been linked to postoperative pain outcomes.

Industry relevance: The findings apply directly to vendors of neuroimaging software and surgical navigation platforms that offer multimodal fusion. The demonstration that a widely used learning-based registration method (SynthMorph) can fail globally while scoring well locally is a caution for anyone shipping automated registration without global sanity checks, and the FOV finding favors explicit affine initialization stages over end-to-end learned pipelines in clinical data with partial overlap.

Future Directions

• Decouple registration from segmentation. Because vessel segmentation runs after registration, the distance metrics reflect the combined pipeline. A "segment-once-then-warp" design — extracting vessels from the original MRA and warping them with each registration field — would isolate registration geometry from segmentation-model sensitivity.

• Validate the venous-offender hypothesis. The reader study is hypothesis-generating. Confirming that low TOF-MRA contrast reflects venous compression would require complementary venous imaging or intraoperative ground truth.

• Expand beyond a single institution. The cohort is predominantly Siemens 3T from one center; multi-vendor and multi-protocol validation is needed to establish whether the contrast and FOV thresholds generalize.

• Define a clinical usability endpoint. The blinded reader study shows that local metrics and clinical evaluability diverge. A scoring system that combines whole-brain plausibility with ROI-level vessel accuracy could provide a single actionable quality gate.

Target Audience

Medical image analysis researchers working on multimodal registration, particularly those evaluating methods on focal clinical targets rather than whole-brain alignment. Also relevant to neuroradiologists and neurosurgeons involved in trigeminal neuralgia workup, clinical imaging scientists designing preprocessing pipelines for surgical planning, and machine-learning engineers deploying registration or vessel-segmentation models where evaluation metrics must reflect downstream clinical usefulness.

Authors’ abstract

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.