PaperlayerView exact revision on arXiv

Source-linked AI summary

A$^3$PreCore: Robust Rigid Registration for Native FLAIR-Space Brain-Mask Transfer

Elmira Mirzabeigi, Zahra Khalilzad, Taylor Snowden-Richardson, Jesus Eduardo Lugo Arce, Jodie R. Gawryluk, Brian R. Christie, Jocelyn Faubert

arXiv:2610.20561v1cs.CVeess.IV

TL;DR

Unstable T1-to-FLAIR registration can leave a plausible anatomical mask badly placed in native FLAIR, while global descriptors may miss localized or orientation-related disagreement. A3PreCore changes transformation estimation while retaining the T1-space mask, and E-QC characterizes masks across complementary dimensions. In this cohort, A3PreCore avoided the gross placement failures observed with fsl_anat and remained concordant with successful registrations and direct-FLAIR masks.

  • Problem

    A plausible T1-space source mask can still be substantially displaced in native FLAIR, and localized mask defects may evade global descriptors.

    Source: §1.2 · p. 2, §1.3 · p. 3

  • Method

    A3PreCore applies a subject-specific six-degree-of-freedom rigid transformation estimated between full bias-corrected T1 and native FLAIR to the existing T1-space mask.

    Source: §7.2 · p. 44, §8 · p. 48

  • Results

    Gross rigid-body placement failures observed with the implemented fsl_anat were not reproduced by A3PreCore, which remained closely concordant with successful registrations and showed high overall agreement with BET.

    Source: §8 · pp. 48–49

  • Takeaways & Limitations

    E-QC's complementary geometric measures revealed disagreement patterns not apparent from global mask volume or overlap alone.

    Source: §8 · p. 49

    Without independently delineated native-FLAIR reference masks, the study could not establish absolute voxel-level anatomical accuracy.

    Source: §7.5 · p. 47

Abstract

from arXiv · show

Reliable brain-mask transfer into native fluid-attenuated inversion recovery (FLAIR) space can fail when T1-to-FLAIR rigid registration is unstable. We developed A$^3$PreCore, a six-degree-of-freedom rigid-registration strategy that aligns the full bias-corrected T1-weighted image to native FLAIR using normalized mutual information while preserving the established T1-space anatomical mask. We also introduced extended quantitative quality control (E-QC), which characterizes mask extent, field-of-view-relative position, FLAIR intensity support, boundary support, inferior morphology, and pairwise geometric disagreement. In 78 participants with paired T1-weighted and FLAIR imaging, A$^3$PreCore was compared with fsl_anat and the Brain Extraction Tool (BET). Gross rigid-body misregistration occurred in 21/78 fsl_anat outputs (26.9%) and 0/78 A$^3$PreCore outputs (exact two-sided McNemar p=9.54 x 10^-7). Among 57 participants without gross fsl_anat misregistration, A$^3$PreCore preserved close agreement (median Dice=0.993; median paired Delta NMI=-0.00007, p=0.429). A$^3$PreCore remained broadly concordant with BET (median Dice=0.955), with less inferior and extracranial extension in representative comparisons. Method-blinded neuroscientists ranked A$^3$PreCore first in 35/40 participants (87.5%). E-QC additionally exposed severe orientation-related disagreement not evident from global descriptors or centroid displacement alone. These findings support A$^3$PreCore as a robust native-FLAIR brain-mask transfer strategy and E-QC as a complementary framework for multidimensional mask characterization and participant-specific geometric quality control.

1 Introduction

Reliable native-FLAIR brain masks matter for lesion-focused analysis, but direct extraction and T1-to-FLAIR transfer can both produce errors that global measures may miss. The paper proposes full-T1-guided rigid registration and multidimensional E-QC to assess mask placement and disagreement.

1.1 Clinical and Methodological Importance of Native FLAIR-Space Brain Masking

  • Errors in native-FLAIR masks can affect lesion volume, regional distribution, boundary measurements, and downstream analyses, making reliable masking important for lesion-focused workflows.This matters near cortical, juxtacortical, periventricular, cerebellar, and inferior boundaries.Source: §1.1 · p. 2

1.2 Challenges of Brain-Mask Generation in Native FLAIR Space

  • Heterogeneous contrast, coverage, artifacts, and pathology make direct FLAIR extraction challenging, while localized boundary defects may escape global volume or centroid checks.Potential defects include inferior extracranial extension and exclusion of cortical, temporal, cerebellar, or inferior brain regions.Source: §1.2 · p. 2

1.3 Limitations of T1-to-FLAIR Brain-Mask Transfer

  • T1-to-FLAIR transfer separates source-mask anatomy from geometric placement, so a plausible T1 mask can still be unusable if transformation estimation fails.Evaluation should test both reduction of severe placement failures and preservation of already successful registrations.Source: §1.3 · pp. 2–3

1.4 Rationale for Robust T1-to-FLAIR Rigid Registration

  • A3PreCore uses the full bias-corrected T1 image, normalized mutual information, and joint six-degree-of-freedom rigid optimization to place the established T1-space mask in native FLAIR.The study also compares it with direct-FLAIR BET, a registration-independent comparator rather than anatomical ground truth.Source: §1.4 · p. 3

1.5 Need for Multidimensional Quantitative Quality Control

  • E-QC combines method-specific descriptors of extent, position, FLAIR and boundary support, and inferior morphology with pairwise geometric measures, without collapsing them into one score.This complements overlap measures, which do not specify the location or physical distance of disagreement and cannot establish anatomical accuracy without a manual reference.Source: §1.5 · p. 4

1.6 Study Contributions

  • The study evaluates selective registration robustness and preservation, compares A3PreCore with BET, and combines E-QC with agreement measures, visual review, and blinded neuroscientist ranking.The A3PreCore strategy preserves the established T1-space anatomical source mask while changing transformation estimation.Source: §1.6 · pp. 5–6, Figure 1 · p. 5

2 Materials and Data

This cross-sectional study used baseline MRI data from a community-dwelling older-adult cohort. Its analyses required paired T1-weighted and FLAIR images, evaluated masks in native FLAIR space, and lacked an independently delineated native-FLAIR reference.

2.1 Study Cohort and Ethics

  • The baseline cohort comprised 80 participants, with a mean age of 65.6 ± 7.8 years; 78 had paired T1-weighted and FLAIR images for registration analyses.The study retained cases with subsequently observed registration failures and used a single-site acquisition setting.Source: §2.1 · p. 6

2.2 MRI Acquisition and Image Inputs

  • The analysis used T1-weighted BRAVO and three-dimensional sagittal T2-weighted CUBE FLAIR acquisitions on a GE SIGNA Pioneer 3 T system.The T1 and FLAIR series contained 380 and 252 reconstructed sagittal images, respectively.Source: §2.2 · pp. 6–7

2.3 Analysis Population and Mask Data

  • Native FLAIR defined the target grid; BET generated a direct-FLAIR comparator, while fsl_anat and A3PreCore transferred the same T1-derived mask using different registration inputs.No independently delineated native-FLAIR masks were available, so pairwise comparisons assessed inter-method agreement rather than anatomical accuracy.Source: §2.2 · p. 7, §2.3 · pp. 7–8

3 A3PreCore: Full-T1-Guided Registration for Native-FLAIR Brain-Mask Transfer

A3PreCore changes how the T1-to-FLAIR transformation is estimated while retaining the same T1-derived anatomical mask. The workflow uses full-T1 rigid registration, then applies the transformation to the mask and evaluates outputs on the native FLAIR grid.

3.1 Methodological Rationale

  • A3PreCore treats anatomical mask definition and geometric placement as distinct operations, addressing failures in the latter without redefining the T1-space mask.The full T1 image retains head contour and extracranial structures represented in raw FLAIR, which may provide broader spatial context for registration.Source: §3.1 · p. 8

3.2 fsl_anat T1-to-FLAIR Transfer Comparator

  • The fsl_anat comparator registers skull-stripped T1 to raw FLAIR, then transfers the same anatomical source mask used by A3PreCore.This controls source-mask anatomy; the differing anatomical support between the moving and fixed images may reduce spatial constraints in challenging scans.Source: §3.2 · pp. 8–9

3.3 A3PreCore Full-T1 Registration Strategy

  • A3PreCore registers the full bias-corrected T1 image to raw FLAIR with FLIRT's normmi cost function and jointly optimizes six rigid-body parameters.The configured initial angular search spans −90° to +90° on each axis; these are search bounds, not fixed rotations.Source: §3.3 · p. 9

3.4 Rigid Registration Model and Multimodal Similarity Objectives

  • The workflows use six-degree-of-freedom rigid transformations, excluding scaling, shearing, and nonlinear deformation; A3PreCore uses normalized mutual information for multimodal alignment.The conceptual objectives describe registration criteria, and NMI analyses are not independent evidence of anatomical accuracy because NMI is part of A3PreCore's optimization objective.Source: §3.4 · pp. 10–11

3.5 Anatomical Mask Propagation and Native-FLAIR Outputs

  • Each estimated transformation is applied to the same binary anatomical source mask, which is resampled with nearest-neighbour interpolation on the native FLAIR grid.The configured dilation step was applied to the registration-derived masks used in the primary analyses.Source: §3.5 · p. 11

3.6 Methodological Contrast and Technical Output Validation

  • The implemented workflows share the reference, source mask, rigid model, resampling, dilation, and output geometry, but differ in moving-image representation, similarity objective, and rotational search.The comparison therefore evaluates the implemented strategies as a whole, not the moving-image representation alone; adverse spatial outputs were retained after technical validation.Source: §3.6 · pp. 11–12

4 E-QC: Multidimensional Quality Control for Native-FLAIR Brain Masks

Global mask descriptors can miss participant-specific displacement, orientation, and localized boundary abnormalities. E-QC addresses this with separate mask-characterization and pairwise-geometric profiles, retaining their dimensions rather than combining them into one score.

4.1 Rationale and Design Principles

  • Global volume, field-of-view position, and signal characteristics can appear plausible despite displacement, orientation differences, or localized boundary abnormalities.Direct FLAIR extraction may also have localized defects despite plausible global volume and centroid.Source: §4.1 · p. 12
  • Separate mask characterization from pairwise geometric characterization so individual mask properties and disagreement between masks remain distinct.Neither descriptor level is interpreted as anatomical ground truth.Source: §4.1 · pp. 12–13
  • E-QC retains complementary descriptor dimensions without scalarizing them, preserving discordant mask behaviours that aggregation could obscure.The framework’s two analytical branches are summarized in Figure 3.Source: §4.1 · p. 13, Figure 3 · p. 14

4.2 Common Image Domain and Input Validation

  • E-QC inputs undergo common validation and native-space preprocessing before the framework characterizes candidate masks and their prespecified comparator.The figure describes the native FLAIR image, candidate mask, and comparator as inputs.Source: §4.2 · p. 13, Figure 3 · p. 14

4.3 FLAIR Intensity and Spatial Preprocessing

E-QC robustly normalizes positive finite FLAIR intensities for descriptors that use intensity or gradient information, and derives inferior–superior orientation from the native FLAIR affine.

4.3 FLAIR Intensity and Spatial Preprocessing

  • Positive finite FLAIR intensities are robustly normalized using their 1st and 99th percentiles when the 99th percentile exceeds the 1st.A degenerate distribution invalidates the affected descriptor rather than receiving an artificial score.Source: §4.3 · p. 14
  • When intensity normalization is degenerate, only the affected descriptor is treated as invalid.No artificial score is assigned.Source: §4.3 · p. 14
  • The native FLAIR affine determines the physical superior direction and thus the inferior–superior coordinate used for morphology.This uses affine-encoded anatomical orientation rather than raw voxel-index ordering alone.Source: §4.3 · pp. 14–15

4.4 Method-Specific Mask Characterization

E-QC characterizes each candidate mask through complementary descriptors of extent, position, image support, and morphology. These measures are interpreted separately and descriptively, not as a single anatomical-accuracy score.

4.4 Method-Specific Mask Characterization

  • The six method-specific descriptors are retained separately because they differ in units, interpretation, and sensitivity to acquisition geometry.They cover volume, FOV-relative position, intensity support, boundary-gradient support, peripheral occupation, and inferior-tail morphology.Source: §4.4 · p. 15

4.4.1 Physical Mask Volume

  • Mask volume measures spatial extent but does not locate disagreement, so it is interpreted descriptively rather than as evidence of anatomical superiority.Volume is reported in millilitres.Source: §4.4.1 · p. 15

4.4.2 Geometric Centroid and FOV-Relative Position

  • The dimensionless FOV-relative centroid offset describes position relative to acquisition extent, not anatomical centring.Head position, cropping, and inferior coverage can alter it.Source: §4.4.2 · p. 15

4.4.3 Mask-to-Complement FLAIR Intensity Ratio

  • The mask-to-complement squared-intensity ratio is method-specific, not an invariant tissue-contrast or direct anatomical-inclusion measure.It is undefined if either sampling region is empty or the denominator is zero.Source: §4.4.3 · p. 16

4.4.4 Spacing-Aware Boundary-Gradient Support

  • Boundary-gradient support measures local FLAIR-gradient support at the mask boundary relative to the whole-image gradient distribution, not anatomical boundary accuracy.Strong gradients may also occur at skull, scalp, sinus, or neck interfaces.Source: §4.4.4 · p. 16

4.4.5 FLAIR-Derived Foreground Support

  • The FLAIR-derived foreground support is built independently of the candidate mask and is an image-derived reference, not an anatomical segmentation.Its peripheral band uses three voxel-grid erosion iterations; its thickness is not a fixed physical distance.Source: §4.4.5 · pp. 16–17

4.4.6 Peripheral Foreground Occupation

  • Peripheral occupation is the fraction of a candidate mask in the outer band of the FLAIR-derived foreground, but excludes voxels entirely outside that support.It therefore does not measure all extracranial inclusion or every possible overextension.Source: §4.4.6 · p. 17

4.4.7 Inferior-Tail Fraction

  • The inferior-tail fraction targets a narrow, low-cross-sectional-area extension from the mask’s inferior end, but broad extensions may change it little.Valid brainstem or cerebellar anatomy can also contribute, so it is a morphological descriptor rather than a contamination label.Source: §4.4.7 · p. 18

4.5 Pairwise Geometric Characterization

Pairwise E-QC complements method-specific mask descriptors with measures of relative volume, centroid position, and principal-axis disagreement. These geometric measures characterize mask differences but do not directly estimate registration parameters or anatomical superiority.

4.5 Pairwise Geometric Characterization

  • Pairwise E-QC retains complementary measures of global extent, position, and principal-axis orientation for predefined mask pairs.The descriptors characterize how candidate masks differ from one another.Source: §4.5 · p. 18

4.5.1 Signed Relative Volume Difference

  • The signed relative volume difference changes sign when method order is reversed, while its magnitude is symmetric and related to volumetric similarity.Its sign indicates which candidate has the larger physical volume, not anatomical superiority.Source: §4.5.1 · p. 18

4.5.2 Physical Centroid Displacement

  • Centroid displacement reports the Euclidean distance between candidate-mask centroids in millimetres, with voxel-coordinate displacement also retained.It is not a direct estimate of registration translation because shape differences can also shift centroids.Source: §4.5.2 · p. 18

4.5.3 Principal-Axis Frames and Stability

  • Principal-axis frames use ordered covariance eigenvectors, but orientation descriptors are treated as interpretable only when eigengaps meet a stability criterion.If either mask fails the criterion, PCA orientation and displacement values are recorded as unavailable; the threshold is not an anatomical or diagnostic cutoff.Source: §4.5.3 · p. 19

4.5.4 Principal-Axis Orientation Difference

  • The sign-resolved PCA orientation difference compares dominant geometric axes, not the actual rotational parameters of T1-to-FLAIR registration.Localized inclusion, asymmetric extraction, inferior extension, or dilation can alter the PCA frame.Source: §4.5.4 · pp. 19–20

4.5.5 PCA-Orientation-Induced Boundary Displacement

  • PCA-orientation-induced boundary displacement expresses the same relative PCA-frame rotation as a physical boundary-scale displacement in millimetres.It is related to θPCA, not statistically independent, and is not equivalent to common surface-distance measures or actual registration displacement.Source: §4.5.5 · p. 20

5 Experimental Evaluation

The evaluation compares native-FLAIR masks and registration workflows using complementary characterization, agreement, and multimodal-similarity analyses. Because independently delineated native-FLAIR masks were unavailable, these measures describe mask behaviour and inter-method agreement rather than anatomical accuracy.

5.1 Evaluation Design and Prespecified Comparisons

  • Without independently delineated native-FLAIR masks, the evaluation treats mask descriptors and inter-method measures as characterizations of behaviour and agreement, not anatomical accuracy.Primary comparisons used a common cohort of 78 participants with paired T1-weighted and FLAIR imaging.Source: §5.1 · p. 21

5.2 Method-Specific E-QC and Visual Registration Assessment

  • Six method-specific E-QC descriptors characterize volume, FOV-relative centroid, FLAIR intensity and boundary support, and inferior morphology; visual review separately classified gross rigid-body displacement.Three neuroscientists independently identified the same 21 fsl_anat outputs as grossly misregistered, and QC descriptors were not treated as equivalent to that visual endpoint.Source: §5.2 · pp. 21–22

5.3 Pairwise Inter-Method Mask Agreement

  • Pairwise mask comparisons use neutral agreement measures, including overlap, directional exclusive extent, and volume differences relative to a prespecified comparator.Exclusive fractions describe unshared mask proportions, not false-positive or false-negative fractions; relative volume percentages are directional.Source: §5.3 · p. 22, §5.3.1 · p. 22, §5.3.2 · p. 23

5.3.3 Surface and Boundary Disagreement

  • Surface disagreement is measured on voxel-centre boundaries using physical voxel spacing, with 2-mm Surface Dice prespecified as the central tolerance.Surface agreement does not establish which boundary is anatomically correct; maximum Hausdorff distance is exploratory because isolated voxels can dominate it.Source: §5.3.3 · p. 23

5.3.4 Regional and Slice-Wise Disagreement

  • Regional and slice-wise analyses localize disagreement across inferior, middle, and superior FOV thirds, but cropping and neck coverage can affect the exploratory partition.Per-slice summaries include median and 10th-percentile Dice among slices with a non-empty mask union.Source: §5.3.4 · p. 24

5.4 Controlled Registration-Related Multimodal Similarity

  • Controlled registration-related NMI comparisons reapply both transformations to the same full bias-corrected T1 image on the native-FLAIR reference grid.Higher NMI is optimization-consistent multimodal-similarity evidence, not independent evidence of anatomical correctness; sensitivity analyses vary histogram bins and evaluation support.Source: §5.4 · pp. 24–25

5.5 Method-Blinded Expert Ranking of Brain-Mask Quality

  • Three method-blinded neuroscientists ranked A3PreCore, fsl_anat, and BET masks for overall quality and anatomical plausibility, with majority ordering used for analysis.The blinded assessment covered a random subset of 40 participants; the first two reviewers were primary, and the third adjudicated disagreements.Source: §5.5 · p. 25

5.6 Statistical Analysis

  • Paired participant-level contrasts used Wilcoxon or exact McNemar tests, while regional, slice-wise, maximum-Hausdorff, sensitivity, subgroup, and association analyses were exploratory.Pairwise mask-agreement measures were interpreted primarily as descriptive disagreement, and no scalar E-QC score was defined.Source: §5.6 · pp. 25–26

6 Results

In the shared 78-participant cohort, the results characterize native-FLAIR mask compatibility, registration outcomes, inter-method agreement, participant-specific geometric differences, and reviewer preferences across A3PreCore, fsl_anat, and BET. These comparisons describe relative mask quality and agreement, not accuracy against an independent anatomical reference.

6.1 Primary Outcome: Gross Rigid-Body Misregistration

  • 21/78 fsl_anat outputs (26.9%) showed gross misregistration, compared with 0/78 A3PreCore outputs; all 21 paired discordances favored A3PreCore (exact two-sided McNemar p=9.54 x 10^-7).The proportion of gross-misregistration-free outputs increased from 57/78 (73.1%) to 78/78 (100.0%), an absolute paired increase of 26.9 percentage points; this measures registration robustness, not absolute anatomical accuracy.Source: §6.1 · pp. 26–27, Figure 4 · p. 28

6.2 Ablation and Registration-Strategy Characterization

  • 26.9% gross failure with fsl_anat fell to 5.1% without full-T1 registration, 3.8% without NMI, and 2.6% without extended search, versus 0.0% for complete A3PreCore.The complete strategy combined full-T1 registration, NMI, and extended angular search; each two-component configuration rescued most baseline failures, but only the full combination eliminated gross misregistration across the cohort.Source: §6.2 · pp. 27–29, Figure 5 · p. 29
  • Across 78 participants, median NMI was 1.111 for A3PreCore and 1.102 for fsl_anat, with median paired difference 0.00006 (Wilcoxon p=0.0022; sign-test p=0.308).NMI was a controlled registration-similarity descriptor, not an independent measure of anatomical accuracy; the distributional shift did not indicate a consistent increase across participants.Source: §6.2 · pp. 29–30, Figure 6 · p. 32
  • Among 21 grossly misregistered fsl_anat cases, all paired NMI differences favored A3PreCore (median paired difference 0.04567; p<0.001), while the 57 non-failure cases had median paired difference −0.00007 (p=0.429).All 21 cases had higher NMI with A3PreCore; among non-failure cases, median Dice was 0.993, indicating close mask agreement. These comparisons characterize registration behavior and inter-method agreement, not anatomical accuracy.Source: §6.2 · p. 30, §6.3 · p. 32, Figure 6 · p. 32

6.3 Spatial Agreement and Boundary Concordance

  • Median A3PreCore–fsl_anat Dice was 0.993 in 57 non-failure cases versus 0.889 in 21 gross-misregistration cases, concentrating the greatest spatial disagreement in the latter subgroup.In the misregistration subgroup, median directional non-overlap volumes were 158.4 mL for A3PreCore-only voxels and 170.8 mL for fsl_anat-only voxels.Source: §6.3 · p. 32, Figure 7 · p. 33
  • A3PreCore–BET median Dice was 0.955; median HD95 and ASSD were 4.005 mm and 1.477 mm, compared with 0.777 mm and 0.274 mm for A3PreCore–fsl_anat.BET provided a registration-independent native-FLAIR mask comparison; these pairwise measures characterize inter-method agreement rather than anatomical accuracy.Source: §6.3 · pp. 33–34, Table 4 · p. 33, Table 5 · p. 34
  • 21/78 A3PreCore–fsl_anat comparisons had normalized symmetric difference above 0.10 despite none having absolute volume difference above 2%, showing that similar total volumes can mask spatial disagreement.The reported median absolute volume difference was 0.464 mL for A3PreCore–fsl_anat; geometric and spatial measures were needed to characterize discrepancies beyond global volume.Source: §6.3 · pp. 34–35

6.4 Method-Specific Extended Quantitative Quality-Control Characterization

  • Scalar QC scores did not fully characterize participant-specific geometric disagreement, which E-QC assessed alongside mask volume, FOV-relative centroid, FLAIR intensity, boundary support, peripheral occupation, and inferior-tail fraction.The method-specific E-QC comparison evaluated six descriptors independently and used no composite score.Source: §6.3 · p. 36, §6.4 · p. 36
  • E-QC differences were descriptor-specific: versus fsl_anat, A3PreCore shifted RI by +0.001397 and BG by +0.025187, while C, Pperipheral, and Ptail showed no clear paired shift.Versus BET, median differences were −24.013 mL for V, +0.001028 for C, +0.132618 for RI, +0.442448 for BG, −0.000638 for Pperipheral, and +0.002093 for Ptail; these descriptors characterize masks independently and do not directly measure anatomical accuracy.Source: §6.4 · pp. 36–37, Table 7 · p. 37

6.5 Geometric Characterization of Inter-Method Mask Disagreement

  • A3PreCore–fsl_anat medians were 0.386 mm centroid, 0.539° orientation, and 0.551 mm boundary displacement; A3PreCore–BET medians were 1.170 mm, 3.442°, and 3.496 mm, respectively.Both comparisons had right-skewed distributions and participant-level extremes; orientation disagreement could be pronounced despite modest centroid displacement, so no single geometric measure captured the full disagreement pattern.Source: §6.5 · pp. 37–40, Table 8 · p. 40

6.6 Blinded Expert Ranking of Brain-Mask Quality

  • A3PreCore ranked first in 35/40 participants (87.5%), versus BET in 3/40 (7.5%) and fsl_anat in 2/40 (5.0%), in blinded review.Reviewers commonly noted the perceived precision and spatial plausibility of A3PreCore boundaries; without an independent native-FLAIR mask reference, these rankings represent relative expert judgments rather than segmentation accuracy.Source: §6.6 · p. 40, Table 9 · p. 40

6.7 Representative Inter-Method Qualitative Comparisons

  • CQDM25_BL and CQDM78_BL had nearly identical Dice scores (0.9400 and 0.9403) but opposite directional disagreement: A3PreCore-exclusive volume dominated in the former, whereas BET-exclusive volume dominated in the latter.Separately, CQDM02_BL had only 1.63 mm centroid displacement despite an 81.28° orientation difference (Dice 0.8877).Source: §6.7 · p. 42

7 Discussion

A3PreCore had fewer gross placement failures than the implemented fsl_anat workflow while remaining concordant in successful registrations. The discussion presents E-QC as a multidimensional aid, but notes that the study cannot establish absolute anatomical accuracy or generalizability.

7.1 Principal Findings

  • 21/78 fsl_anat outputs had gross rigid-body misregistration, compared with 0/78 A3PreCore outputs; the authors caution that this is not an estimate of absolute anatomical accuracy.In this cohort, A3PreCore remained broadly concordant with direct BET masks and successful fsl_anat transfers.Source: §7.1 · pp. 43–44

7.2 A3PreCore as a Full-T1-Guided Geometric-Transfer Strategy

  • A3PreCore optimizes six-degree-of-freedom placement of an existing T1-space mask; it does not redefine the source mask or correct its systematic boundary errors.The two registration branches used the same T1-derived mask, separating placement differences from source-mask definition.Source: §7.2 · p. 44

7.3 Multidimensional E-QC and Interpretation Without Manual Ground Truth

  • E-QC descriptors and pairwise coordinates reveal complementary mask differences that global volume, centroid position, or overlap alone can miss.PCA-derived orientation difference can identify orientation disagreement without a comparably large centroid shift; orientation-induced boundary displacement expresses it in physical-distance units.Source: §7.3 · p. 45

7.4 Relationship to Existing Methods and Pipeline Implications

  • High global agreement can coexist with local boundary differences, especially in inferior regions, and agreement between automatic methods does not establish which boundary is anatomically preferable.BET-exclusive extent was concentrated near inferior boundaries in representative comparisons; no manual native-FLAIR reference was available.Source: §7.3 · pp. 45–46, §7.4 · p. 46
  • A3PreCore complements T1-processing pipelines, while direct FLAIR extraction remains useful without a suitable T1 image; validation should assess completion, placement, and local boundaries separately.The study does not establish superiority over modern learning-based systems or comprehensive direct-FLAIR approaches; more flexible registration also remains to be assessed.Source: §7.4 · pp. 46–47

7.5 Limitations and Future Directions

  • Absolute accuracy and cross-dataset generalizability remain unestablished: there were no independent native-FLAIR reference masks, the cohort was single-site, and E-QC thresholds were not validated.A3PreCore is also rigid-only, and local boundary comparisons relied on measures and visual review rather than manual voxel-level ground truth.Source: §7.5 · pp. 47–48

7.2 A3PreCore as a Full-T1-Guided Geometric-Transfer Strategy

  • The comparison varied image representation, similarity criterion, and angular search together, so the failure reduction cannot be attributed to any one component of A3PreCore.The controlled NMI analysis supported the interpretation but was not independent anatomical validation.Source: §7.2 · pp. 44–45, §7.5 · p. 47

7.5 Limitations and Future Directions

  • The study evaluated mask placement, not lesion-segmentation performance; downstream studies should test lesion and regional outcomes and validate across populations, scanners, and protocols.The authors propose a modular pipeline combining T1 guidance, direct FLAIR information, failure detection, and lesion segmentation, with processing stages kept auditable.Source: §7.5 · p. 48

8 Conclusion

A3PreCore transfers an established T1-derived mask to native FLAIR using six-degree-of-freedom rigid optimization. In this cohort it avoided the gross placement failures seen with the implemented fsl_anat while remaining concordant with successful transfers and BET masks.

8 Conclusion

  • A3PreCore estimates a subject-specific six-degree-of-freedom rigid transform between full bias-corrected T1 and native FLAIR, then applies it to the existing T1-space mask.It is a geometric transfer strategy, not a replacement for anatomical brain extraction.Source: §8 · pp. 48–49
  • Gross rigid-body placement failures observed with the implemented fsl_anat were not reproduced by A3PreCore, which remained concordant with successful fsl_anat registrations and showed high overall agreement with BET masks.The conclusion emphasizes that workflow completion and plausible global mask characteristics alone do not ensure reliable placement.Source: §8 · pp. 48–49
  • E-QC descriptors and pairwise geometric measures revealed disagreements missed by global volume or overlap, but independent native-FLAIR references and multi-site validation are needed to establish accuracy and generalizability.Centroid displacement, orientation difference, and orientation-induced boundary displacement may help prioritize outputs for tri-planar review.Source: §8 · p. 49

Data and Code Availability

Individual-level MRI data are not publicly available under the study's privacy and governance constraints, while the source code and analysis workflow materials are publicly available.

Data and Code Availability

  • The individual-level MRI data analyzed in the study are not publicly available.The passage cites human-participant imaging, ethics approval, informed consent, and privacy restrictions.Source: Data and Code Availability · p. 49
  • Requests for access to de-identified data may be considered through the study investigators, subject to institutional data-governance requirements.Source: Data and Code Availability · p. 49
  • The A3PreCore source code, analysis scripts, and workflow documentation are publicly available, with a permanent release identifier planned upon publication.Source: Data and Code Availability · p. 49

CRediT authorship contribution statement

The contribution statement assigns conceptualization, analysis, methodology, validation, writing, and other project roles across the named authors.

CRediT authorship contribution statement

S1 Supplementary Results

Supplementary analyses show that controlled-NMI comparisons depend more on evaluation support than histogram resolution, regional agreement varies by anatomy, and whole-mask volume can conceal spatial disagreement.

S1.1 Sensitivity Analysis of Controlled NMI

  • Under common FLAIR-foreground support with 64 bins, median ΔNMI was 0.000063, with higher NMI for A3PreCore in 44/78 participants and fsl_anat in 34/78.Across common-support definitions, median differences remained slightly positive and close to zero.Source: S1.1 Sensitivity Analysis of… · p. 55, Table S1 · p. 56, Figure S1 · p. 60
  • Mean ΔNMI was 0.015939 while median ΔNMI was 0.000063; all 21 participants with gross fsl_anat misregistration had positive differences across common-support configurations.The mean–median separation was consistent with large positive differences concentrated in a subset, while medians among participants without gross misregistration stayed near zero.Source: S1.1 Sensitivity Analysis of… · p. 55
  • On BET-mask support, median ΔNMI was −0.001203, −0.000981, and −0.000827 at 32, 64, and 128 bins, respectively; fsl_anat had higher NMI in 50/78 participants.The authors retained this as an exploratory support restriction; mean paired differences remained positive.Source: S1.1 Sensitivity Analysis of… · p. 55, Figure S1 · p. 60
  • Across 78 leave-one-participant-out analyses, median ΔNMI ranged from 0.00005219 to 0.00007301, with maximum absolute deviation of 1.04 × 10^-5 from the full-sample estimate.This tested sensitivity to individual participants, not invariance to support, histogram resolution, preprocessing, or NMI implementation.Source: S1.1 Sensitivity Analysis of… · p. 55

S1.2 Regional and Slice-Wise Spatial Agreement

  • For A3PreCore–BET, median Dice was 0.870 in the inferior third, 0.962 in the middle third, and 0.942 in the superior third; A3PreCore–fsl_anat values were 0.961, 0.991, and 0.989.Both comparisons showed lower agreement in the inferior third; the regional difference was more pronounced for A3PreCore–BET.Source: S1.2 Regional and Slice-Wise Spatial… · p. 56, Table S2 · p. 56
  • Lower-tail 10th-percentile slice Dice was substantially lower than median slice Dice, particularly for A3PreCore–BET, consistent with disagreement concentrated near inferior mask boundaries.The 10th percentile was used because minimum Dice was frequently zero in small terminal slices and less stable as a cohort-level descriptor.Source: S1.2 Regional and Slice-Wise Spatial… · p. 56

S1.3 Whole-Mask Volume Disagreement

  • For A3PreCore–fsl_anat, median absolute volume difference was 0.464 mL (0.030%), yet 21/78 participants exceeded a normalized symmetric difference of 0.10 and 21/78 exceeded HD95 of 5 mm.No participant exceeded an absolute volume difference of 2%, despite substantial spatial discrepancies.Source: S1.3 Whole-Mask Volume Disagreement · p. 57, Table S3 · p. 57
  • For A3PreCore–BET, median absolute volume difference was 30.277 mL and 41/78 participants exceeded 2%; 12/78 exceeded HD95 of 5 mm and 2/78 exceeded 10 mm.Volume difference alone did not capture marked spatial disagreement.Source: S1.3 Whole-Mask Volume Disagreement · p. 57, Table S3 · p. 57
Loading 2610.20561v1…