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

Figure 1: Conceptual synthesis of native FLAIR-space brain masking, T 1 -to-FLAIR mask transfer, and multidimensional quality assessment. The schematic summarizes the clinical relevance of native FLAIR-space brain masking, major classes of brain-masking approaches, potential failure modes of T 1 -to-FLAIR mask transfer, and design principles for improving rigid-registration robustness. The A 3 PreCore strategy is illustrated as a full-T 1 -guided, optimization-based rigid registration approach in which a subject-specific six-degree-of-freedom transformation is estimated between the full bias-corrected T 1 image and native FLAIR using normalized mutual information, with translational and rotational parameters optimized jointly before transfer of the established T 1 -space anatomical mask. The schematic further illustrates the proposed extended quantitative quality-control framework (E-QC), which retains method-specific mask descriptors and pairwise geometric disagreement as complementary dimensions rather than combining them into a single aggregate score. Global, FLAIR-support, boundary-sensitive, inferior-morphology, and geometric descriptors are used to characterize mask behaviour and participant-specific inter-method disagreement. The final panel summarizes the principal literature gaps and the evaluation strategy used in the present study, including comparison with fsl_anat , registration-independent comparison with BET, preservation analysis in successful registrations, gross-failure characterization, E-QC analysis, and blinded neuroscientist ranking.

Parameter	T 1 -weighted	FLAIR
Scanner	GE SIGNA Pioneer	GE SIGNA Pioneer
Field strength	3 T	3 T
Receiver coil	Head 32	Head 32
Sequence	3D sagittal BRAVO	3D sagittal T 2 CUBE FLAIR
Acquisition type	3D	3D
Orientation	Sagittal	Sagittal

Table 1: MRI acquisition and reconstruction parameters for the T 1 -weighted and FLAIR sequences.

Figure 2: Native-FLAIR brain-mask transfer and comparison workflows. BET generated a brain mask directly from the native FLAIR image,

F

. The fsl_anat and A 3 PreCore workflows propagated the same T 1 -space anatomical mask,

M_{T}^{\mathrm{anat}}

, but differed in transformation estimation. The fsl_anat registered the skull-stripped T 1 image,

T_{B}

, to

F

using FLIRT with mutualinfo , whereas A 3 PreCore registered the full bias-corrected T 1 image,

T

, to

F

using normmi . Both workflows estimated subject-specific six-degree-of-freedom rigid transformations comprising rotational and translational parameters and applied the resulting transformation to the same anatomical source mask using nearest-neighbour resampling. All candidate masks were evaluated on the native FLAIR grid.

Figure 3: Overview of the extended quantitative quality-control (E-QC) framework for native-FLAIR brain-mask assessment. E-QC receives the native FLAIR image,

F

, a candidate mask,

M_{a}

, and a prespecified comparator,

M_{b}

, after common input validation and native-space preprocessing. The framework separates into two complementary analytical branches. Method-specific mask characterization independently retains physical volume, field-of-view-relative position, FLAIR intensity support, boundary-gradient support, peripheral foreground occupation, and inferior-tail morphology. Pairwise geometric characterization quantifies signed relative volume difference, physical centroid displacement, principal-axis orientation difference, and the corresponding PCA-orientation-induced boundary displacement, with PCA-based descriptors reported only when the predefined eigengap stability criterion is satisfied. The resulting method-specific and pairwise profiles are retained as separate multidimensional descriptors and are not reduced to a composite quality score. The fsl_anat serves as the primary methodological comparator and BET as the registration-independent native-FLAIR comparator; neither is treated as anatomical ground truth.

Figure 4: Study flow and primary paired registration outcome. a , Study cohort, processing completeness, and construction of the analysis populations. Of the 80 baseline CQDM participants considered, two were excluded from the comparative analysis because the required FLAIR input was unavailable, resulting in a final common analysis cohort of 78 participants. BET, the fsl_anat , and A 3 PreCore were evaluated in this common cohort, and all three workflows produced readable, native-FLAIR-grid-compatible outputs in all 78 participants. b , Paired gross rigid-body misregistration outcome for the primary comparison between the fsl_anat and A 3 PreCore. Gross misregistration was identified in 21/78 fsl_anat outputs (26.9%) and in none of the corresponding A 3 PreCore outputs (0/78, 0.0%). The 21 paired discordances all consisted of fsl_anat gross misregistration with no corresponding gross misregistration in A 3 PreCore. fsl_anat failures comprised 14 translation-dominant, one rotation-dominant, and six combined translational and rotational cases. The proportion of gross-misregistration-free outputs increased from 57/78 (73.1%) to 78/78 (100.0%), corresponding to an absolute paired increase of 26.9 percentage points (exact two-sided McNemar test,

p=9.54\times 10^{-7}

).

Configuration	Moving T 1	Cost	Search	Gross failure	Rescued a	New failures b	McNemar $p$
fsl_anat	Skull-stripped	MI	Default	21/78 (26.9%)	–	–	–
No full T 1	Skull-stripped	NMI	$\pm 90^{\circ}$	4/78 (5.1%)	17/21	0/57	$1.53\times 10^{-5}$
No NMI	Full	MI	$\pm 90^{\circ}$	3/78 (3.8%)	18/21	0/57	$7.63\times 10^{-6}$
No extended search	Full	NMI	Default	2/78 (2.6%)	19/21	0/57	$3.81\times 10^{-6}$
A 3 PreCore	Full	NMI	$\pm 90^{\circ}$	0/78 (0.0%)	21/21	0/57	$\mathbf{9.54\times 10^{-7}}$

Table 2: Leave-one-component-out ablation of the A 3 PreCore registration strategy, using gross rigid-body misregistration as the primary endpoint. Each ablated configuration retained two of the three registration components, whereas the complete A 3 PreCore configuration combined full-T 1 registration, NMI, and extended angular search.

Figure 5: Gross rigid-body misregistration across A 3 PreCore ablation configurations. Rates were 26.9% for fsl_anat , 5.1% for No full T 1 , 3.8% for No NMI , 2.6% for No extended search , and 0.0% for the complete A 3 PreCore configuration, all evaluated in the common 78-participant cohort.

Figure 6: Controlled NMI characterization of fsl_anat and A 3 PreCore registration transforms. a , Participant-level NMI after the fsl_anat and A 3 PreCore transforms were reapplied to the same full bias-corrected T 1 image and evaluated against the same native FLAIR reference. The identity line denotes equal NMI between the two transforms. b , Participant-level paired NMI difference,

\Delta\operatorname{NMI}=\operatorname{NMI}_{\mathrm{A3P}}-\operatorname{NMI}_{\mathrm{fsl\_anat}}

, stratified by fsl_anat gross-misregistration status. Positive differences were concentrated in participants with gross fsl_anat misregistration, whereas paired differences remained centred near zero among participants without gross fsl_anat misregistration. NMI is interpreted as a registration-similarity descriptor rather than as independent anatomical validation.

Figure 7: Pairwise Dice agreement according to comparator and fsl_anat registration status. Participant-level Dice coefficients are shown for A 3 PreCore–BET across the complete cohort and for A 3 PreCore– fsl_anat across the complete cohort and separately among participants with and without gross fsl_anat misregistration. Agreement remained high when the fsl_anat showed no gross misregistration, whereas substantially lower overlap was concentrated in the fsl_anat gross-misregistration subgroup.

	A 3 PreCore–BET	A 3 PreCore– fsl_anat
Metric	Overall	Overall	Gross misregistration	No gross misregistration
	( $n=78$ )	( $n=78$ )	( $n=21$ )	( $n=57$ )
Dice	0.955 [0.95–0.96]	0.990 [0.93–0.99]	0.889 [0.86–0.92]	0.993 [0.98–0.99]
Jaccard	0.914 [0.90–0.91]	0.981 [0.87–0.99]	0.800 [0.76–0.85]	0.985 [0.98–0.99]
Comparator coverage	0.948 [0.94–0.96]	0.990 [0.93–0.99]	0.884 [0.86–0.92]	0.992 [0.98–0.99]
A 3 PreCore-contained fraction	0.965 [0.95–0.97]	0.990 [0.93–0.99]	0.894 [0.87–0.92]	0.993 [0.98–0.99]

Table 4: Pairwise spatial agreement of A 3 PreCore with BET and the fsl_anat .

Metric	A 3 PreCore– fsl_anat	A 3 PreCore–BET
Normalized symmetric difference	0.019 (0.013–0.128)	0.086 (0.081–0.091)
HD95 (mm)	0.777 (0.693–5.423)	4.005 (3.638–4.513)
ASSD (mm)	0.274 (0.186–2.272)	1.477 (1.384–1.583)
Surface Dice, 1 mm	0.996 (0.318–0.999)	0.476 (0.455–0.500)
Surface Dice, 2 mm	1.000 (0.537–1.000)	0.732 (0.707–0.760)
Surface Dice, 3 mm	1.000 (0.729–1.000)	0.891 (0.862–0.907)

Table 5: Boundary and surface disagreement between A 3 PreCore and the fsl_anat and BET comparators.

Figure 8: Boundary and surface agreement of A 3 PreCore with the fsl_anat and BET comparators. a , HD95. b , Average symmetric surface distance (ASSD). c , Surface Dice at a 1-mm tolerance. d , Normalized symmetric difference. These measures characterize pairwise inter-method agreement and do not represent error against an independently delineated anatomical ground-truth reference.

Metric	BET ( $n=78$ )	fsl_anat ( $n=78$ )	A 3 PreCore ( $n=78$ )
$V$ (mL)	1503.7 (1391.3–1593.3)	1465.8 (1372.9–1569.2)	1465.5 (1372.8–1566.8)
$C$	0.07972 (0.07215–0.08637)	0.08076 (0.07386–0.08897)	0.08075 (0.07289–0.08745)
$R_{I}$	5.2696 (4.7234–5.8034)	5.2794 (4.7560–5.8867)	5.4330 (4.8589–5.9479)
$B_{G}$	2.2592 (1.9709–2.5025)	2.4010 (2.2319–3.0117)	2.7034 (2.3276–3.0739)
$P_{\mathrm{peripheral}}$	0.00705 (0.00525–0.00875)	0.00622 (0.00472–0.00800)	0.00618 (0.00459–0.00796)
$P_{\mathrm{tail}}$	0.02449 (0.02107–0.02945)	0.02602 (0.02114–0.03201)	0.02720 (0.02277–0.03260)

Table 6: Method-specific E-QC descriptor distributions for BET, fsl_anat , and A 3 PreCore in the common 78-participant cohort.

Comparator	Descriptor	A 3 PreCore median	Comparator median	Paired difference	Bootstrap 95% CI	$r_{\mathrm{rb}}$	$p$
fsl_anat	$V$ (mL)	1465.502	1465.765	$-0.370$	$-0.475$ to $-0.080$	$-0.505$	$<0.001$
	$C$	0.080750	0.080761	$-0.000326$	$-0.000640$ to $0.000064$	$-0.248$	0.057
	$R_{I}$	5.432968	5.279435	$+0.001397$	$-0.000601$ to $0.006260$	0.420	0.0013
	$B_{G}$	2.703394	2.401043	$+0.025187$	$0.004760$ to $0.058166$	0.552	$<0.001$
	$P_{\mathrm{peripheral}}$	0.006176	0.006222	$-5.83\times 10^{-6}$	$-1.59\times 10^{-5}$ to $7.31\times 10^{-6}$	$-0.055$	0.674
	$P_{\mathrm{tail}}$	0.027196	0.026017	$+0.000071$	$-0.000127$ to $0.000337$	0.175	0.179

Table 7: Participant-level paired comparisons of six method-specific E-QC descriptors for A 3 PreCore relative to fsl_anat and BET (

n=78

).

Figure 9: Participant-level multidimensional E-QC profiles. Rows represent the 78 participants and columns represent the six method-specific descriptors (

V

,

C

,

R_{I}

,

B_{G}

,

P_{\mathrm{peripheral}}

, and

P_{\mathrm{tail}}

) for BET, fsl_anat , and A 3 PreCore. Values were robustly standardized within each descriptor for visualization only; all statistical analyses used the original descriptor values.

Figure 10: Paired E-QC descriptor effect profile. Points represent paired rank-biserial effect sizes for A 3 PreCore relative to fsl_anat and BET across the six method-specific E-QC descriptors. Positive and negative values indicate the numerical direction of the paired difference and do not denote anatomical superiority.

Comparison	Centroid displacement	PCA orientation	PCA-orientation-induced displacement	Orientation $>10^{\circ}$
	(mm)	( ∘ )	(mm)	$n$
A 3 PreCore– fsl_anat	0.386 (0.247–2.578)	0.539 (0.383–1.414)	0.551 (0.383–1.412)	13/78
A 3 PreCore–BET	1.170 (0.698–1.586)	3.442 (2.273–4.810)	3.496 (2.292–4.951)	2/78

Table 8: Pairwise geometric characterization of A 3 PreCore relative to fsl_anat and BET in the common 78-participant cohort.

Figure 11: Pairwise geometric characterization of inter-method mask disagreement. a , Participant-level physical centroid displacement. b , PCA-derived principal-axis orientation difference. c , PCA-orientation-induced boundary displacement. d , Relationship between centroid displacement and orientation difference, illustrating that pronounced orientation disagreement may occur despite only modest centroid displacement. Magenta denotes A 3 PreCore– fsl_anat comparisons and green denotes A 3 PreCore–BET comparisons. Distributions are shown for the common 78-participant cohort. These descriptors quantify inter-method geometric disagreement and are not interpreted as anatomical error relative to an independent ground-truth reference.

Method	Rank 1, $n$ (%)	Rank 2, $n$ (%)	Rank 3, $n$ (%)
A 3 PreCore	35 (87.5%)	3 (7.5%)	2 (5.0%)
BET	3 (7.5%)	11 (27.5%)	26 (65.0%)
fsl_anat	2 (5.0%)	26 (65.0%)	12 (30.0%)

Table 9: Method-blinded expert ranking of brain-mask quality in the random 40-participant evaluation subset.

Figure 12: Representative native-FLAIR comparisons between A 3 PreCore and fsl_anat . Sagittal, coronal, and axial views illustrate three distinct patterns of inter-method disagreement. The top row (CQDM02_BL) demonstrates pronounced orientation disagreement despite limited centroid displacement. The middle row (CQDM66_BL) shows severe gross geometric disagreement involving substantial positional and orientational differences. The bottom row (CQDM70_BL) illustrates near-complete overlap with minimal residual boundary differences. Green indicates A 3 PreCore-exclusive regions, red indicates fsl_anat -exclusive regions, and light blue indicates shared mask extent. Quantitative measurements are provided beneath each representative case. The difference maps characterize inter-method disagreement rather than anatomical ground-truth error.

Figure 13: Representative native-FLAIR comparisons between A 3 PreCore and BET. Sagittal, coronal, and axial views demonstrate distinct directional patterns of inter-method disagreement. The top row (CQDM95_BL) shows extensive overlap with nearly balanced method-exclusive volumes. The middle row (CQDM25_BL) illustrates greater A 3 PreCore-exclusive extent, whereas the bottom row (CQDM78_BL) demonstrates substantially greater BET-exclusive extent despite a comparable Dice coefficient. Green indicates A 3 PreCore-exclusive regions, red indicates BET-exclusive regions, and light blue indicates shared mask extent. Quantitative measurements are reported beneath each case. The overlays describe relative mask extent and spatial disagreement without establishing anatomical superiority of either method.

Evaluation support	Bins	Mean $\Delta$ NMI	Median $\Delta$ NMI (IQR)	A 3 PreCore/ fsl_anat higher, $n$
Common FLAIR foreground	32	0.021211	0.000144 ( $-0.000105$ –0.043475)	46/32
Common FLAIR foreground	64	0.015939	0.000063 ( $-0.000171$ –0.032549)	44/34
Common FLAIR foreground	128	0.012836	0.000047 ( $-0.000127$ –0.025678)	46/32
Common full field of view	32	0.015441	0.000098 ( $-0.000121$ –0.028743)	47/31
Common full field of view	64	0.012883	0.000128 ( $-0.000075$ –0.023143)	52/26
Common full field of view	128	0.011072	0.000147 ( $-0.000010$ –0.019317)	56/22

Table S1 : Sensitivity of the controlled NMI comparison to evaluation support and histogram resolution. Positive

\Delta\mathrm{NMI}

values indicate higher NMI for A 3 PreCore than for fsl_anat . The configured main analysis is shown in bold.

Figure S1 : Sensitivity of the paired controlled-NMI comparison to voxel support and histogram discretization. Median participant-level differences,

\Delta\mathrm{NMI}=\mathrm{NMI}_{\mathrm{A3P}}-\mathrm{NMI}_{\mathrm{fsl\_anat}}

, are shown for common FLAIR foreground, common non-zero support, common full-field-of-view support, and BET-mask support using 32, 64, and 128 histogram bins. Positive values indicate higher NMI for A 3 PreCore. BET-mask support was evaluated as an exploratory support-restriction analysis.

Metric	A 3 PreCore– fsl_anat	A 3 PreCore–BET
Inferior-third Dice	0.961 (0.838–0.980)	0.870 (0.800–0.898)
Middle-third Dice	0.991 (0.939–0.994)	0.962 (0.957–0.965)
Superior-third Dice	0.989 (0.935–0.993)	0.942 (0.931–0.947)
Mean slice Dice	0.975 (0.869–0.983)	0.883 (0.872–0.897)
Median slice Dice	0.990 (0.940–0.994)	0.956 (0.952–0.959)
10th-percentile slice Dice	0.962 (0.795–0.977)	0.776 (0.704–0.828)

Table S2 : Regional and slice-wise spatial agreement between A 3 PreCore and the two comparators.

Metric	A 3 PreCore– fsl_anat	A 3 PreCore–BET
Signed volume difference (mL)	$-0.370$ ( $-0.647$ –0.136)	$-24.013$ ( $-50.048$ –4.521)
Absolute volume difference (mL)	0.464 (0.217–0.839)	30.277 (17.690–56.709)
Absolute volume difference (%)	0.030 (0.015–0.061)	2.136 (1.114–3.910)
A 3 PreCore volume smaller/larger, $n$	53/25	56/22
Normalized symmetric difference $>0.10$ , $n$	21	9
HD95 $>5$ mm, $n$	21	12

Table S3 : Whole-mask volume disagreement and frequencies of marked spatial disagreement between A 3 PreCore and the two comparators.

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

Abstract

Figures & tables

Medical Image Quality Assessment

Cycle Inverse-Consistent TransMorph: A Balanced Deep Learning Framework for Brain MRI Registration

Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

CoFlu-Brain: An fMRI Connectivity Foundation Model with Cofluctuation-Based Pretraining

Abstract

Medical Image Quality Assessment

Explore similar work

Cycle Inverse-Consistent TransMorph: A Balanced Deep Learning Framework for Brain MRI Registration

Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

CoFlu-Brain: An fMRI Connectivity Foundation Model with Cofluctuation-Based Pretraining