MIRTO: a registration-gated, multiverse-tested evaluation protocol for unsupervised anomaly segmentation in brain MRI

Negin Kafee Hernashki; Soumick Chatterjee

cs.CVOct 1, 2026

MIRTO: a registration-gated, multiverse-tested evaluation protocol for unsupervised anomaly segmentation in brain MRI

Authors: Negin Kafee Hernashki, Soumick Chatterjee

Organizations: Università degli Studi di Milano, Milan, Italy · Politecnico di Milano, Milan, Italy · Human Technopole, Milan, Italy · Otto von Guericke University Magdeburg, Germany

Abstract

Unsupervised anomaly detection (UAD) methods for brain MRI are ranked by a single score, yet that score rests on choices that are rarely reported: how each anomaly map is aligned with the reference, how and on which data the threshold is set, and which false-positive budget, metric, aggregation and lesion definition are used. We present MIRTO, an evaluation protocol that makes these choices explicit and measures their effect. It gates the geometry of every comparison with a registration check and label-free diagnostics of known power, sets thresholds on validation data alone and reports the false-positive volume actually realised on test, repeats each comparison over 15,552 defensible evaluation pipelines, and attaches paired subject-bootstrap intervals with multiplicity control. Applied to four UAD methods trained on the same healthy data and tested on 312 BraTS 2020 subjects, MIRTO showed that an axis-order mismatch between stored maps and the reference lowered a diffusion model's voxel AUROC from 0.873 to 0.583 whilst barely moving its slice-level AUROC. Within each metric, the method explained at least 0.95 of the variance in voxel AUROC and AUPRC and 0.77 in Dice, but only 0.14 in lesion sensitivity, where the lesion definition and hit criterion dominated. A Dice advantage that was significant at validation thresholds vanished at equal realised false-positive burden, and an exact identity attributes it to threshold transfer. A training-free change to REFLECT's latent aggregation raised Dice at equal burden by 0.052. Nine hypotheses were tested against explicit criteria; because the same cohort served to develop the protocol, all inference is exploratory.

Figures & tables

Figure 1: The MIRTO protocol. Frozen anomaly maps from every method enter four components. R : one canonical reference and a repaired, gated transform per method. T : thresholds chosen on validation data alone, with the realised (not the nominal) false-positive burden, the threshold-transfer gap and a conformal per-scan guarantee. Exact identities accompany the metrics and explain how they relate. M : a multiverse of defensible evaluation pipelines. I : paired subject-bootstrap intervals with Holm adjustment inside declared families. Dashed boxes give, for each component, a failure it caught in this study (numbers from Section 3 ).

Figure 2: Registration gating (schematic). (a) Each stored map is mapped to the canonical grid by a per-method transform

T_{m}

(Eq. 7 ); voxels outside the field of view

F

score zero. (b) The registration gate compares the applied transform with the other candidate transforms on the overlap between the method’s own reference, mapped, and the canonical reference inside

F

(Eq. 8 ). (c) Label-free diagnostics: an in-plane error pushes per-subject voxel AUROC below

0.5

(anti-location rate) and scrambles its ranks (rank stability). Bars are illustrative, not data.

Figure 3: Operating-point geometry (schematic). (a) Each method traces mean Dice against mean realised burden on test,

\Psi_{m}

. The validation threshold lands at a realised burden

\hat{\beta}_{m}

that need not equal the nominal

b

; the transfer gap

\bar{g}_{m}

is the rise of

\Psi_{m}

between

\hat{\beta}_{m}

and

b

(Identity 1 ). (b) A post-processed map at threshold

\tau

realises a smaller burden than plain thresholding at the same

\tau

; its gain over plain thresholding at equal burden splits exactly into an effect at fixed

\tau

and a relaxation of the plain threshold from

\tau_{m}

\tau

(Identity 5 ). (c) For the loss

\mathbf{1}[\mathrm{FP}>b]

, CRC selects an order statistic of the per-subject critical thresholds, which yields the known Beta law of Lemma 1 .

Axis	Levels
reference	canonical $C$ ; method’s own reference, mapped; round trip via method grid
support	full brain; field of view $\cap$ brain
voxel volume	true 1 mm 3 ; declared legacy value
threshold	validation (Eq. 10 ) ; CRC $\alpha=0.10$ ; test-matched ; fixed $0.5$ ; per-subject optimum
burden $b$	5, 10, 20, 40 cm 3
post-processing	plain thresholding; hysteresis; hysteresis with pruning

Table 1: Axes of the evaluation multiverse. Levels in bold enter the defensible subset; all levels of the remaining axes do.

Name	Source	Training contrasts
IXI-HH	IXI, Hammersmith (3 T)	T2
IXI-Guys	IXI, Guy’s (1.5 T)	T2
IXI-HH+Guys	IXI, both sites above	T2
MOOD+IXI-HH	MOOD and IXI Hammersmith	T1 and T2
MOOD+IXI	MOOD and all of IXI	T1 and T2

Table 2: Healthy training sets. All models were tested on BraTS 2020 T2. IXI-HH+Guys is used for every method in the main text.

Figure 4: Registration. (a) Per-subject overlap of each method’s own reference, mapped to the canonical grid, with the canonical reference inside the field of view; the dashed line is the gate’s floor (0.85). (b) Change in Dice, voxel AUROC and slice-level AUROC under injected corruptions of two verified pipelines: in-plane errors devastate the voxel metrics and leave slice-level AUROC unchanged or improved.

	Method share	Axis share †
voxel AUROC	0.954 [0.933, 0.966]	0.046 [0.033, 0.067]
voxel AUPRC	0.960 [0.941, 0.972]	0.040 [0.028, 0.059]
slice AUROC, marginal	0.974 [0.964, 0.980]	0.025 [0.019, 0.036]
Dice	0.770 [0.738, 0.787]	0.207 [0.190, 0.234]
slice AROC, adjusted	0.718 [0.627, 0.783]	0.267 [0.207, 0.355]
lesion sensitivity	0.139 [0.127, 0.152]	0.657 [0.648, 0.666]

Table 3: Sum-of-squares shares across defensible universes within each metric and, last, pooled over metrics, where the metric main effect mostly reflects the scales of the metrics. Within-metric shares test the exploratory hypothesis and carry no decision criterion. † Axis main effects plus method

\times

axis interactions.

Figure 5: Multiverse. Specification curves, one panel per metric; each column is a defensible universe ( $2{,}592$ per metric), ordered by REFLECT’s score. Dice and lesion sensitivity reorder the three residual models from universe to universe; voxel AUROC does not.

Figure 6: What decides a ranking, metric by metric. (a) Share of the variance across defensible universes explained by the method and by the evaluation axes (main effects plus interactions with method), within each metric and pooled over metrics; the hatched bar is the main effect of the metric axis, which reflects scale. (b) Within-metric flip shares of the three residual-model pairs with 95% bootstrap intervals;

\phi=0.5

means that the pair is ordered each way equally often. Read (b) together with the size of each difference (Section 3.2 ).

Method	$b$	$\hat{\tau}_{b}$	Dice mean	median	pooled	FP mean	median	p95	max	$>b$	gap $\bar{g}(b)$
REFLECT	5	0.930	0.241 [0.223, 0.260]	0.235	0.311	3.3 [1.5, 6.0]	0.4	6.0	293	6.7%	$+0.073$ [0.070, 0.077]
	10	0.870	0.382 [0.359, 0.405]	0.409	0.458	7.2 [4.4, 11.2]	2.0	17.1	431	10.3%	$+0.051$ [0.049, 0.053]
	20	0.785	0.497 [0.473, 0.520]	0.549	0.568	16.3 [12.2, 22.0]	8.6	37.2	604	15.7%	$+0.020$ [0.019, 0.022]
	40	0.685	0.554 [0.531, 0.577]	0.604	0.615	35.1 [29.4, 42.6]	24.9	76.6	801	20.2%	$+0.003$ [0.003, 0.004]
UCCD	5	0.500	0.441 [0.426, 0.455]	0.440	0.448	5.3 [4.0, 7.0]	1.0	21.2	166	24.4%	$-0.008$ [ $-0.008$ , $-0.007$ ]
	10	0.445	0.497 [0.482, 0.511]	0.506	0.504	9.6 [7.6, 12.1]	2.8	34.2	226	28.8%	$+0.003$ [0.002, 0.003]

Table 4: Operating points at the validation threshold (plain thresholding). Dice: subject mean with interval, median and pooled value. Realised FP burden on test (cm 3 ): mean with interval, median, 95th percentile, maximum and share of subjects above

b

. Last column: threshold-transfer gap

\bar{g}(b)

(Eq. 12 ).

Figure 7: Threshold transfer. (a) Mean realised burden on test against the nominal budget (left) and share of subjects above the budget (right) at the validation threshold. (b) Threshold-transfer gap

\bar{g}(b)

with 95% intervals; the dashed line is the decision criterion of H6.

Figure 8: Paired Dice contrasts between residual models at the validation threshold (black) and at equal realised burden (orange), by budget and aggregation. Filled markers: Holm-adjusted

p<0.05

in the declared 24-contrast family. The UCCD

-

cDDPM advantage at validation thresholds disappears at equal burden; REFLECT

-

UCCD changes sign with budget and aggregation.

		mean exceedance		share $\leq\alpha$		97.5th pct.
$n$	law	observed	exact	observed	exact	observed	exact
24	Beta(2, 23)	0.078–0.082	0.080	69.5–71.2%	70.8%	0.205–0.222	0.211
50	Beta(5, 46)	0.097–0.099	0.098	55.6–59.3%	56.9%	0.191–0.206	0.192
100	Beta(10, 91)	0.096–0.100	0.099	53.7–57.7%	54.9%	0.170–0.179	0.164

Table 5: CRC (

\alpha=0.10

b=20\,\mathrm{cm}^{3}

) over 1,000 random calibration splits of the test cohort: observed range over the four methods against the exact law

\mathrm{Beta}(n+1-k,k)

of Lemma 1 .

	voxel AUROC			voxel AUPRC
	mean	med.	pool	mean	med.	pool
REFLECT	0.932 [0.925, 0.939]	0.951	0.945	0.611 [0.582, 0.639]	0.689	0.642
UCCD	0.865 [0.857, 0.872]	0.880	0.862	0.608 [0.589, 0.626]	0.655	0.580
cDDPM	0.873 [0.865, 0.881]	0.889	0.867	0.630 [0.611, 0.649]	0.686	0.593
AnomalyDINO	0.832 [0.822, 0.841]	0.848	0.826	0.311 [0.290, 0.331]	0.303	0.286

Table 6: Threshold-free metrics per method: subject mean [95% CI], subject median and pooled value.

	total	fixed- $\tau$ effect	relaxation	sens. global $\to$ pp
REFLECT	$+0.034$	$-0.013$ [ $-0.022$ , $-0.005$ ]	$+0.046$ [0.042, 0.050]	0.608 $\to$ 0.549
UCCD	$+0.085$	$+0.006$ [ $-0.004$ , 0.016]	$+0.078$ [0.073, 0.083]	0.674 $\to$ 0.629
cDDPM	$+0.116$	$-0.002$ [ $-0.012$ , 0.008]	$+0.117$ [0.110, 0.125]	0.665 $\to$ 0.619
AnomalyDINO	$-0.001$	$-0.044$ [ $-0.055$ , $-0.035$ ]	$+0.044$ [0.036, 0.051]	0.210 $\to$ 0.121

Table 7: Exact decomposition (Eq. 22 ) of the gain from hysteresis with pruning at equal realised burden,

b=20\,\mathrm{cm}^{3}

, and above-floor lesion sensitivity with plain thresholding and with post-processing.

Figure 9: Post-processing at equal burden. (a) Gain of hysteresis with pruning over plain thresholding, split into the fixed-threshold effect and the threshold relaxation (Identity 5 ). (b) Above-floor lesion sensitivity at

20\,\mathrm{cm}^{3}

falls with post-processing for every method.

Scorer	marg.	strat.	m. 0.2	m. 0.05	reg. 5	obs. $-$ borrowed
REFLECT	0.887	0.754	0.734	0.726	0.600	$+0.229$ [0.201, 0.257]
UCCD	0.907	0.759	0.769	0.743	0.649	$+0.266$ [0.244, 0.288]
cDDPM	0.905	0.762	0.768	0.744	0.646	$+0.271$ [0.249, 0.293]
AnomalyDINO	0.598	0.619	0.620	0.624	0.504	$+0.124$ [0.100, 0.148]
brain area	0.884	0.691	0.668	0.560	0.626	–
mid-brain closeness	0.857	0.576	0.550	0.486	0.530	–

Table 8: Slice-level AUROC: marginal, covariate-adjusted (stratified; matched with two calipers; spline regression, 5 df), and observed minus borrowed-map AROC.

Figure 10: Slice-level AUROC under six estimators. Covariate-only controls (crosses) should sit at 0.5 if an estimator removes anatomy; none does.

Figure 11: Detection at

20\,\mathrm{cm}^{3}

, validation threshold. (a) Error modes of the 472 above-floor lesions. (b) Detected (coloured) and missed (grey) lesions by standardised log volume and T2 contrast: volume separates them almost completely.

Figure 12: Baselines minus models for voxel AUROC, voxel AUPRC and Dice (

20\,\mathrm{cm}^{3}

, validation threshold) under three aggregations, with Holm-adjusted families. FLAIR beats the residual models only in subject-averaged threshold-free metrics.

Panel of Figure 13 — Figure 13: Training data. (a) Dice, voxel AUPRC and lesion sensitivity in the reference universe when all four methods share one training set (Table 2 ), and in the test-selected set, in which each method used the training set with its best test Dice. (b) Training set chosen on the 24 validation subjects by six selection rules; each cell gives the choice, the share of $10^{4}$ bootstrap resamples that reproduce it (also shown by the shading) and the test Dice lost relative to the best training set.

Figure 14: Latent aggregation for REFLECT. Dice of each score rule minus the published map at

20\,\mathrm{cm}^{3}

, matched to the published map’s realised burden over the cohort (black) or per subject (blue), for REFLECT trained on two different training sets.

Hypothesis	Decision criterion	Verdict; estimate
H1 ranking instability	a pair flips in $\geq 20\%$ of defensible universes (lower bound)	supported; $\phi_{\max}=0.472$ [0.414, 0.499]
H2 evaluation dominance	evaluation choices explain more variance than method, pooled over metrics	supported, driven by the metric axis (scale); 0.614 vs 0.162
H3 intensity baseline	T2 within 0.05 Dice of $\geq 1$ model at equal FP	not supported; closest $+0.098$ [0.076, 0.119]
H4 contrast dependence	detection rises with T2 contrast and oedema fraction in every residual model	partly testable; not supported
H5 pruning trade-off	pruning helps detection at 5, hurts at 40 cm 3	not supported (point estimates)
H6 test-tuning optimism	test-matched $\tau$ overstates Dice by $\geq 0.02$ at $20\,\mathrm{cm}^{3}$ (lower bound)	not supported; cDDPM $+0.0234$ [0.0199, 0.0269], REFLECT $+0.0203$ [0.0191, 0.0216]

Table 9: Hypotheses H1 to H9, their decision criteria and verdicts (all methods trained on IXI-HH+Guys).

Appendix figures & tables54 assets

Supplementary material from the paper’s appendix.

Appendix

Split	Canonical array (x $\times$ y $\times$ z voxels, 1 mm)	Whole-tumour volume, median [IQR] (cm 3 )	Range (cm 3 )	Distinct reference hashes
test (312)	240 $\times$ 240 $\times$ [98, 149], median 139	91.7 [53.2, 147.8]	7.3 to 361.8	312
validation (24)		89.6 [43.2, 137.8]	12.5 to 180.7	24

Appendix

Table 10: Cohort and canonical arrays.

Legacy bundle (method	grid)	Median ratio legacy/canonical voxel count	Range	$\approx$ 1.0004 (same 240-px grid)
AnomalyDINO\|fourway	1.0004	0.974 to 1.063	True	False
AnomalyDINO\|pairA	1.0004	0.974 to 1.063	True	False
REFLECT\|fourway	1.0188	0.974 to 1.146	False	False
REFLECT\|pairA	1.0188	0.974 to 1.146	False	False
UCCD\|fourway	1.0514	0.992 to 1.167	False	False
UCCD\|pairB	1.1380	1.122 to 1.196	False	True

Appendix

Table 11: Voxel-volume check from voxel counts: (tumour volume in the method’s files / declared voxel volume) / canonical voxel count.

Method (test-selected pairB grid)	Voxel volume used	Declared vox (mm 3 )	Effective budget for ’nominal 20 cm 3 ’	$\tau$	Dice	Sensitivity (all comps.)	Mean FP (cm 3 )
UCCD	as declared	1.166	17.2	0.600	0.360	0.049	15.3
UCCD	under B3 vox=1.0	1.166	20.0	0.575	0.382	0.051	18.4
cDDPM	as declared	0.604	33.1	0.600	0.029	0.003	30.5
cDDPM	under B3 vox=1.0	0.604	20.0	0.700	0.018	0.001	17.9

Appendix

Table 12: Illustration with the uncorrected cDDPM mapping: the declared voxel volume changes the physical budget behind a nominal

20\,\mathrm{cm}^{3}

Metric	REFLECT	UCCD	cDDPM	AnomalyDINO
Dice	0.515 [0.428, 0.584]; 35.2%	0.543 [0.472, 0.577]; 38.9%	0.496 [0.420, 0.568]; 25.9%	0.097 [0.030, 0.157]; 0%
voxel AUROC	0.945 [0.932, 0.951]; 100%	0.869 [0.865, 0.880]; 0%	0.877 [0.871, 0.889]; 0%	0.832 [0.826, 0.848]; 0%
voxel AUPRC	0.642 [0.611, 0.689]; 50.0%	0.613 [0.590, 0.655]; 0%	0.636 [0.604, 0.686]; 50.0%	0.303 [0.286, 0.311]; 0%
lesion sens.	0.341 [0.064, 0.699]; 35.3%	0.343 [0.077, 0.817]; 53.5%	0.333 [0.060, 0.793]; 11.3%	0.011 [0.000, 0.134]; 0%
slice AUROC marg.	0.886 [0.878, 0.921]; 0%	0.909 [0.891, 0.940]; 100%	0.905 [0.890, 0.938]; 0%	0.598 [0.587, 0.625]; 0%
slice AROC adj.	0.787 [0.754, 0.843]; 33.3%	0.800 [0.759, 0.837]; 0%	0.808 [0.762, 0.842]; 66.7%	0.650 [0.619, 0.661]; 0%

Appendix

Table 13: Each method’s score across the 2,592 defensible universes of a metric: median [IQR] and share of universes in which it ranks first (descriptive; the spread is over evaluation choices, not sampling).

cDDPM (IXI-HH+Guys) quantity	Statistic	Repaired $-$ pre-repair [95 % CI]
voxel AUROC	mean	+0.290 [0.277, 0.302]
voxel AUROC	median	+0.293 [0.277, 0.306]
voxel AUPRC	mean	+0.535 [0.517, 0.553]
voxel AUPRC	median	+0.577 [0.549, 0.599]
slice-max AUROC	mean	+0.112 [0.099, 0.125]
slice-max AUROC	median	+0.083 [0.072, 0.095]

Appendix

Table 14: cDDPM with the incorrect and the correct mapping (same maps; correct minus incorrect, 312 subjects). Dice rows compare each pipeline at its own validation threshold.

Method	gate pass (test + validation)	Dice in FOV, median [IQR]	min	centroid error median (mm)	max (mm)	FOV tumour fraction, median	subjects with FOV fraction $<$ 0.95	pass at overlap $\geq$ 0.95	gate failures (clause)
REFLECT	336/336	0.966 [0.957, 0.972]	0.890	0.58	1.02	1.000	0	275	none
UCCD	335/336	0.947 [0.935, 0.957]	0.845	0.85	1.07	1.000	45	93	BraTS20_Training_0044 (iii)
cDDPM	335/336	0.947 [0.935, 0.957]	0.845	0.85	1.07	1.000	45	93	BraTS20_Training_0044 (iii)
AnomalyDINO	336/336	0.966 [0.957, 0.972]	0.901	0.59	1.01	1.000	0	273	none

Appendix

Table 15: Registration gate per method on 312 test and 24 validation subjects. “Dice in FOV” is the overlap of the method’s own reference, mapped to the canonical grid, with the canonical reference inside the field of view; the column for the fixed 0.95 overlap rule counts the subjects that would pass it.

Injected corruption	Pipeline	n	$\Delta$ Dice (20 cm 3 )	$\Delta$ voxel AUROC	$\Delta$ voxel AUPRC	$\Delta$ slice-max AUROC	$\Delta$ lesion sensitivity	rank stability	anti-location rate
LR flip	UCCD	312	$-$ 0.502	$-$ 0.260	$-$ 0.510	+0.000	$-$ 0.797	0.253	0.122
LR flip	AnomalyDINO	312	$-$ 0.145	$-$ 0.307	$-$ 0.231	+0.021	$-$ 0.252	0.297	0.449
AP flip	UCCD	312	$-$ 0.341	$-$ 0.198	$-$ 0.386	+0.012	$-$ 0.467	0.140	0.170
AP flip	AnomalyDINO	312	$-$ 0.086	$-$ 0.147	$-$ 0.118	+0.181	$-$ 0.158	0.360	0.151
transpose $xy$	UCCD	312	$-$ 0.400	$-$ 0.207	$-$ 0.430	+0.005	$-$ 0.560	0.073	0.215
transpose $xy$	AnomalyDINO	312	$-$ 0.106	$-$ 0.184	$-$ 0.163	+0.099	$-$ 0.192	0.243	0.208

Appendix

Table 16: Bias of the headline metrics (corrupted minus unaltered, cohort mean) and the rank-stability (Spearman) and anti-location diagnostics under every injected corruption. The Jaccard bound was never violated.

Injected corruption	Pipeline	rank-stability power, n = 60	anti-location power, n = 60	smallest n with rank-stability power $\geq$ 0.95	smallest n with anti-location power $\geq$ 0.95
LR flip	UCCD	1.000	1.000	30	30
LR flip	AnomalyDINO	1.000	1.000	30	30
AP flip	UCCD	1.000	1.000	30	30
AP flip	AnomalyDINO	1.000	1.000	30	30
transpose $xy$	UCCD	1.000	1.000	30	30
transpose $xy$	AnomalyDINO	1.000	1.000	30	30

Appendix

Table 17: Power of the rank-stability and anti-location diagnostics at a false-alarm rate fixed at 0.05 by construction; cohort sizes 30, 60 and 312.

Metric	Variant	b	Aggregation	Contrast	n = 312	n = 311 (without BraTS20_Training_0044)
Dice	interpolated	40	pooled	UCCD $-$ cDDPM	$-$ 0.0087 [ $-$ 0.0174, $-$ 0.0001]	$-$ 0.0081 [ $-$ 0.0167, 0.0008]
Dice	testmatched	40	median	REFLECT $-$ UCCD	+0.0184 [0.0003, 0.0466]	+0.0184 [ $-$ 0.0000, 0.0460]
Dice	validation	5	pooled	T2 $-$ AnomalyDINO	$-$ 0.0202 [ $-$ 0.0400, 0.0001]	$-$ 0.0202 [ $-$ 0.0400, $-$ 0.0002]
Dice	validation	20	pooled	REFLECT $-$ UCCD	+0.0202 [0.0013, 0.0386]	+0.0190 [ $-$ 0.0001, 0.0370]
Dice	validation	40	median	REFLECT $-$ UCCD	+0.0242 [0.0006, 0.0480]	+0.0220 [ $-$ 0.0001, 0.0480]

Appendix

Table 18: Contrasts whose interval-excludes-zero status changes when subject BraTS20_Training_0044 is removed (

n=311

); 5 of 588 contrasts, largest absolute change of any point estimate 0.0078.

Method / contrast	Quantity	Full 312	FOV-complete subset	Interval-excludes-0 status changed
cDDPM	FP (cm 3 ) / correct ( $\tau$ 0.5) $-$ incorrect ( $\tau$ 0.875), mean	+0.195 [ $-$ 1.141, 1.613]	$-$ 0.010 [ $-$ 1.453, 1.565]	False
cDDPM	FP (cm 3 ) / correct ( $\tau$ 0.375) $-$ incorrect ( $\tau$ 0.7), mean	+0.385 [ $-$ 3.418, 4.281]	$-$ 0.124 [ $-$ 4.048, 4.090]	False
pairwise	REFLECT - UCCD, auprc_vox, mean	+0.003 [ $-$ 0.015, 0.022]	$-$ 0.009 [ $-$ 0.029, 0.011]	False
pairwise	REFLECT - cDDPM, auprc_vox, median	+0.006 [ $-$ 0.019, 0.032]	$-$ 0.007 [ $-$ 0.023, 0.022]	False

Appendix

Table 19: Estimates whose sign differs between the full cohort and the FOV-complete subset (270 subjects); no interval-excludes-zero status changes.

Method	Grid	dz* (selection 20)	dz* (replication 20)	$\Delta$ voxel AUROC at dz* vs 0 (292 subjects)	AUROC at dz = 0	Offset real by declared rule
AnomalyDINO	native	3	3	+0.0082 [0.0063, 0.0101]	0.826	True
REFLECT	native	1	1	$-$ 0.0005 [ $-$ 0.0009, $-$ 0.0001]	0.933	False
UCCD	native	0	0	+0.0000 [0.0000, 0.0000]	0.896	False
cDDPM	native	2	2	$-$ 0.0005 [ $-$ 0.0011, 0.0002]	0.899	False
AnomalyDINO	canonical	3	3	+0.0053 [0.0034, 0.0071]	0.833	True
REFLECT	canonical	1	1	$-$ 0.0004 [ $-$ 0.0008, 0.0000]	0.933	False

Appendix

Table 20: Through-plane offset check: offset

dz^{\ast}

chosen on 20 subjects, replicated on the next 20, and its effect on voxel AUROC in the remaining 292. “Native” refers to the method’s own grid.

Method grid	Subjects	6-conn. components, canonical $\to$ round trip	Above-floor lesions	Tumour volume ratio (round trip / canonical)	Round trip equals method’s own reference
AnomalyDINO	312	5,473 $\to$ 5,050	472 $\to$ 474	0.9999	100.0 %
REFLECT	312	5,473 $\to$ 5,065	472 $\to$ 472	0.9998	100.0 %
UCCD	312	5,473 $\to$ 2,355	472 $\to$ 543	0.9823	100.0 %
cDDPM	312	5,473 $\to$ 2,355	472 $\to$ 543	0.9823	100.0 %

Appendix

Table 21: Round trip of the canonical reference through each method’s native grid (nearest neighbour). The UCCD/cDDPM grid fragments the reference and loses 1.8% of tumour volume, which is why no MIRTO metric uses a method’s own reference.

Method	brain voxels tied (median share)	outside FOV (median share)	mean $\Delta$ AUROC	max $\|\Delta$ AUROC $\|$	max $\|\Delta$ AUPRC $\|$
REFLECT	0.038	0.001	-0.0000	0.0000	0.0000
UCCD	0.649	0.091	-0.0020	0.0464	0.0020
cDDPM	0.256	0.091	-0.0016	0.0391	0.0020
AnomalyDINO	0.165	0.000	-0.0000	0.0076	0.0037

Appendix

Table 22: Tie audit over 312 subjects per method: ordinal minus mid-rank voxel AUROC (

\Delta

AUROC) and ordinal minus tie-collapsed AUPRC.

Method	unshifted	$-2$	$+2$	$-5$	$+5$	$-10$	$+10$	perm. within	perm. across
REFLECT	0.497	0.493	0.494	0.472	0.476	0.406	0.413	0.183	0.066
UCCD	0.539	0.530	0.537	0.497	0.514	0.418	0.442	0.192	0.068
cDDPM	0.519	0.511	0.517	0.481	0.495	0.404	0.425	0.183	0.063
AnomalyDINO	0.166	0.166	0.167	0.161	0.164	0.141	0.149	0.069	0.026

Appendix

Table 23: Spatial randomisation controls: subject-mean Dice at

20\,\mathrm{cm}^{3}

, validation threshold, for the unshifted map, through-plane shifts of the map by the given number of slices, and slice permutation within or map exchange across subjects.

Method	quantity	F1 export grid	F2 reference source	F3 support	F4 threshold
REFLECT	Dice	+0.0008 [0.0007, 0.0010]	+0.0002 [0.0000, 0.0003]	+0.0002 [0.0001, 0.0002]	$-$ 0.0605 [ $-$ 0.0662, $-$ 0.0547]
REFLECT	voxel AUROC	+0.0002 [0.0001, 0.0004]	+0.0003 [0.0002, 0.0004]	$-$ 0.0000 [ $-$ 0.0001, 0.0000]	+0.0000 [0.0000, 0.0000]
UCCD	Dice	$-$ 0.0017 [ $-$ 0.0023, $-$ 0.0012]	$-$ 0.0016 [ $-$ 0.0022, $-$ 0.0011]	+0.0001 [0.0001, 0.0001]	+0.2358 [0.2187, 0.2527]
UCCD	voxel AUROC	+0.0029 [0.0016, 0.0040]	$-$ 0.0038 [ $-$ 0.0050, $-$ 0.0027]	$-$ 0.0262 [ $-$ 0.0311, $-$ 0.0216]	+0.0000 [0.0000, 0.0000]
cDDPM	Dice	$-$ 0.0017 [ $-$ 0.0023, $-$ 0.0011]	$-$ 0.0017 [ $-$ 0.0023, $-$ 0.0012]	+0.0000 [0.0000, 0.0000]	+0.1926 [0.1731, 0.2123]
cDDPM	voxel AUROC	+0.0037 [0.0023, 0.0049]	$-$ 0.0041 [ $-$ 0.0054, $-$ 0.0029]	$-$ 0.0210 [ $-$ 0.0251, $-$ 0.0171]	+0.0000 [0.0000, 0.0000]

Appendix

Table 24: Geometry factorial: main effects with 95% intervals on subject-mean Dice and voxel AUROC (second level minus first).

Quantity	Universe set	Flip share [95 % CI]
$\phi_{\max}$	defensible	0.472 [0.414, 0.499]
REFLECT vs UCCD	defensible	0.472 [0.389, 0.499]
REFLECT vs cDDPM	defensible	0.472 [0.386, 0.498]
REFLECT vs AnomalyDINO	defensible	0.010 [0.009, 0.014]
UCCD vs cDDPM	defensible	0.393 [0.250, 0.493]
UCCD vs AnomalyDINO	defensible	0.009 [0.008, 0.011]

Appendix

Table 25: Pairwise flip shares (share of universes in which a pair takes its minority order) with 1,000-replicate bootstrap intervals; defensible set (15,552 universes) and all universes (155,520).

Sum-of-squares share	Defensible universes	All universes
method share	0.162 [0.147, 0.177]	0.169 [0.153, 0.184]
axis share	0.614 [0.595, 0.634]	0.595 [0.575, 0.614]
remainder	0.223 [0.212, 0.235]	0.237 [0.226, 0.249]
main effect: metric	0.539 [0.516, 0.561]	0.519 [0.497, 0.541]
method $\times$ metric	0.033 [0.030, 0.037]	0.030 [0.028, 0.034]
main effect: lesion definition	0.014 [0.012, 0.015]	0.014 [0.013, 0.016]

Appendix

Table 26: Sum-of-squares shares, defensible and all universes.

Metric	REFLECT vs UCCD	REFLECT vs cDDPM	UCCD vs cDDPM
voxel AUROC	0.000 [0.000, 0.000]	0.000 [0.000, 0.000]	0.000 [0.000, 0.000]
voxel AUPRC	0.167 [0.000, 0.500]	0.500 [0.167, 0.500]	0.000 [0.000, 0.000]
Dice	0.403 [0.264, 0.495]	0.491 [0.366, 0.500]	0.356 [0.125, 0.412]
lesion sensitivity	0.261 [0.219, 0.302]	0.340 [0.282, 0.392]	0.286 [0.238, 0.398]
slice AUROC, marginal	0.000 [0.000, 0.333]	0.167 [0.000, 0.333]	0.000 [0.000, 0.500]
slice AROC, adjusted	0.333 [0.333, 0.500]	0.333 [0.333, 0.500]	0.000 [0.000, 0.500]

Appendix

Table 27: Within-metric flip shares for the three residual-model pairs (defensible universes; post hoc).

Metric	Distinct four-method orders	Three most frequent orders (share of universes)
Dice	6	REFLECT $>$ UCCD $>$ cDDPM $>$ AnomalyDINO (25.5 %); UCCD $>$ cDDPM $>$ REFLECT $>$ AnomalyDINO (25.0 %); cDDPM $>$ UCCD $>$ REFLECT $>$ AnomalyDINO (20.8 %)
voxel AUROC	1	REFLECT $>$ cDDPM $>$ UCCD $>$ AnomalyDINO (100.0 %)
voxel AUPRC	3	REFLECT $>$ cDDPM $>$ UCCD $>$ AnomalyDINO (50.0 %); cDDPM $>$ REFLECT $>$ UCCD $>$ AnomalyDINO (33.3 %); cDDPM $>$ UCCD $>$ REFLECT $>$ AnomalyDINO (16.7 %)
lesion sensitivity	8	UCCD $>$ cDDPM $>$ REFLECT $>$ AnomalyDINO (43.7 %); REFLECT $>$ UCCD $>$ cDDPM $>$ AnomalyDINO (31.2 %); UCCD $>$ REFLECT $>$ cDDPM $>$ AnomalyDINO (9.7 %)
slice AUROC, marginal	2	UCCD $>$ cDDPM $>$ REFLECT $>$ AnomalyDINO (83.3 %); UCCD $>$ REFLECT $>$ cDDPM $>$ AnomalyDINO (16.7 %)
slice AROC, adjusted	2	cDDPM $>$ UCCD $>$ REFLECT $>$ AnomalyDINO (66.7 %); REFLECT $>$ cDDPM $>$ UCCD $>$ AnomalyDINO (33.3 %)

Appendix

Table 28: Distinct four-method orders per metric across the 2,592 defensible universes of each metric, with the three most frequent.

Method	b (cm 3 )	validation $\tau$	test-matched $\tau$	Realised FP at validation $\tau$ (cm 3 )	Dice at validation $\tau$	Dice at test-matched $\tau$	Threshold-transfer gap, mean [95 % CI]	Gap, median of per-subject differences [95 % CI]
REFLECT	5	0.930	0.902	3.3	0.241	0.314	+0.073 [0.070, 0.077]	+0.082 [0.079, 0.086]
REFLECT	10	0.870	0.839	7.2	0.382	0.433	+0.051 [0.049, 0.053]	+0.054 [0.052, 0.055]
REFLECT	20	0.785	0.761	16.3	0.497	0.517	+0.020 [0.019, 0.022]	+0.021 [0.019, 0.022]
REFLECT	40	0.685	0.667	35.1	0.554	0.558	+0.003 [0.003, 0.004]	+0.003 [0.002, 0.004]
UCCD	5	0.500	0.507	5.3	0.441	0.433	$-$ 0.008 [ $-$ 0.008, $-$ 0.007]	$-$ 0.008 [ $-$ 0.008, $-$ 0.008]
UCCD	10	0.445	0.442	9.6	0.497	0.499	+0.003 [0.002, 0.003]	+0.003 [0.003, 0.003]

Appendix

Table 29: Threshold-transfer gap: Dice at the test-matched minus Dice at the validation threshold, mean and median of per-subject differences, with the thresholds and the realised burden at the validation threshold.

Method	Threshold rule (20 cm 3 )	$\tau$	Test exceedance P(FP_i $>$ 20 cm 3 ) [95 % CI]	Mean FP (cm 3 )	Dice, subject mean
REFLECT	test-matched	0.761	0.228 [0.179, 0.276]	19.9	0.517
REFLECT	validation threshold	0.785	0.157 [0.115, 0.199]	16.3	0.497
REFLECT	CRC validation $\alpha$ = 0.05	0.989	0.013 [0.003, 0.026]	0.6	0.040
REFLECT	CRC validation $\alpha$ = 0.1	0.831	0.074 [0.045, 0.103]	10.8	0.444
REFLECT	CRC validation $\alpha$ = 0.2	0.790	0.154 [0.115, 0.196]	15.7	0.492
REFLECT	RCPS validation $\alpha$ = 0.1	0.989	0.013 [0.003, 0.026]	0.6	0.040

Appendix

Table 30: Every threshold rule at

20\,\mathrm{cm}^{3}

: test exceedance

\mathbb{P}[\mathrm{FP}_{i}>20\,\mathrm{cm}^{3}]

, mean realised burden and subject-mean Dice.

Method	b	test-matched	validation	CRC $\alpha$ 0.05	CRC $\alpha$ 0.10	CRC $\alpha$ 0.20	RCPS $\alpha$ 0.10	label-free CRC $\alpha$ 0.10
REFLECT	5	0.135 / 0.314	0.067 / 0.241	0.013 / 0.010	0.051 / 0.227	0.183 / 0.352	0.013 / 0.010	0.013 / 0.014
REFLECT	10	0.186 / 0.433	0.103 / 0.382	0.013 / 0.021	0.067 / 0.328	0.157 / 0.418	0.013 / 0.021	0.013 / 0.029
REFLECT	20	0.228 / 0.517	0.157 / 0.497	0.013 / 0.040	0.074 / 0.444	0.154 / 0.492	0.013 / 0.040	0.013 / 0.066
REFLECT	40	0.256 / 0.558	0.202 / 0.554	0.010 / 0.085	0.096 / 0.535	0.109 / 0.540	0.010 / 0.085	0.013 / 0.157
UCCD	5	0.231 / 0.433	0.244 / 0.441	0.010 / 0.113	0.074 / 0.310	0.138 / 0.384	0.010 / 0.113	0.000 / 0.027
UCCD	10	0.288 / 0.499	0.288 / 0.497	0.010 / 0.188	0.093 / 0.394	0.144 / 0.446	0.010 / 0.188	0.000 / 0.051

Appendix

Table 31: Every threshold rule at every budget; each cell gives test exceedance / subject-mean Dice. RCPS at

\alpha=0.05

is infeasible with 24 calibration subjects and is omitted.

Method	Calibration n	Mean exceedance	2.5th to 97.5th percentile over splits	Splits with exceedance $\leq$ $\alpha$
REFLECT	24	0.078	[0.007, 0.212]	69.5 %
REFLECT	50	0.097	[0.031, 0.206]	59.3 %
REFLECT	100	0.099	[0.042, 0.175]	54.9 %
UCCD	24	0.079	[0.007, 0.215]	70.1 %
UCCD	50	0.099	[0.031, 0.206]	55.6 %
UCCD	100	0.098	[0.038, 0.175]	57.7 %

Appendix

Table 32: CRC (

\alpha=0.10

20\,\mathrm{cm}^{3}

) over 1,000 random calibration/evaluation splits of the 312 test subjects, per method and calibration size.

Contrast	Operating point	Budget (cm 3 )	Subject mean	Subject median	Pooled
REFLECT $-$ UCCD	validation $\tau$	5	$-$ 0.200 [ $-$ 0.219, $-$ 0.180], p_H $<$ 0.005	$-$ 0.205 [ $-$ 0.235, $-$ 0.180], p_H $<$ 0.005	$-$ 0.137 [ $-$ 0.161, $-$ 0.113], p_H $<$ 0.005
REFLECT $-$ UCCD	validation $\tau$	10	$-$ 0.115 [ $-$ 0.135, $-$ 0.095], p_H $<$ 0.005	$-$ 0.097 [ $-$ 0.136, $-$ 0.057], p_H $<$ 0.005	$-$ 0.046 [ $-$ 0.069, $-$ 0.025], p_H $<$ 0.005
REFLECT $-$ UCCD	validation $\tau$	20	$-$ 0.042 [ $-$ 0.060, $-$ 0.025], p_H $<$ 0.005	$-$ 0.008 [ $-$ 0.040, 0.015], p_H 0.74	+0.020 [0.001, 0.039], p_H 0.070
REFLECT $-$ UCCD	validation $\tau$	40	+0.003 [ $-$ 0.012, 0.018], p_H 0.72	+0.024 [0.001, 0.048], p_H 0.14	+0.056 [0.038, 0.073], p_H $<$ 0.005
REFLECT $-$ UCCD	matched realised FP	5	$-$ 0.118 [ $-$ 0.209, $-$ 0.048], p_H $<$ 0.005	$-$ 0.100 [ $-$ 0.225, $-$ 0.014], p_H 0.23	$-$ 0.049 [ $-$ 0.147, 0.023], p_H 0.79
REFLECT $-$ UCCD	matched realised FP	10	$-$ 0.067 [ $-$ 0.129, $-$ 0.026], p_H $<$ 0.005	$-$ 0.046 [ $-$ 0.121, 0.017], p_H 1.00	+0.001 [ $-$ 0.061, 0.043], p_H 1.00

Appendix

Table 33: Paired Dice contrasts between residual models at every budget, Holm-adjusted within each 24-contrast family. Val: validation threshold; matched: equal realised burden (interpolated).

Contrast	Metric	Subject mean	Subject median	Pooled
REFLECT $-$ UCCD	voxel AUROC	+0.067 [0.060, 0.074], p_H $<$ 0.005	+0.072 [0.064, 0.080], p_H $<$ 0.005	+0.083 [0.076, 0.088], p_H $<$ 0.005
REFLECT $-$ cDDPM	voxel AUROC	+0.059 [0.051, 0.066], p_H $<$ 0.005	+0.062 [0.053, 0.073], p_H $<$ 0.005	+0.078 [0.071, 0.084], p_H $<$ 0.005
UCCD $-$ cDDPM	voxel AUROC	$-$ 0.008 [ $-$ 0.011, $-$ 0.006], p_H $<$ 0.005	$-$ 0.010 [ $-$ 0.015, $-$ 0.003], p_H 0.018	$-$ 0.005 [ $-$ 0.008, $-$ 0.002], p_H 0.016
REFLECT $-$ AnomalyDINO	voxel AUROC	+0.100 [0.093, 0.108], p_H $<$ 0.005	+0.104 [0.092, 0.112], p_H $<$ 0.005	+0.119 [0.108, 0.130], p_H $<$ 0.005
UCCD $-$ AnomalyDINO	voxel AUROC	+0.033 [0.023, 0.043], p_H $<$ 0.005	+0.032 [0.018, 0.042], p_H $<$ 0.005	+0.036 [0.024, 0.049], p_H $<$ 0.005
cDDPM $-$ AnomalyDINO	voxel AUROC	+0.041 [0.031, 0.052], p_H $<$ 0.005	+0.041 [0.026, 0.052], p_H $<$ 0.005	+0.041 [0.029, 0.053], p_H $<$ 0.005

Appendix

Table 34: Paired voxel AUROC and AUPRC contrasts, Holm-adjusted within the voxel-pair families (mean and median, 24 contrasts; pooled, 12).

Family	Contrasts	raw p $<$ 0.05	Holm p $<$ 0.05
Dice, model vs baseline, equal FP, mean	32	31	31
Dice, model vs baseline, equal FP, median	32	32	32
Dice, model vs baseline, equal FP, pooled	32	30	29
Dice, model vs baseline, test-matched, mean	32	31	31
Dice, model vs baseline, test-matched, median	32	32	32
Dice, model vs baseline, test-matched, pooled	32	30	29

Appendix

Table 35: Holm family summary: number of contrasts per declared family with raw and Holm-adjusted

p<0.05

Method	slice max	slice 99th pct.	slice mean
REFLECT	0.887 [0.873, 0.899]	0.874 [0.860, 0.888]	0.646 [0.622, 0.671]
UCCD	0.907 [0.896, 0.918]	0.900 [0.887, 0.912]	0.869 [0.852, 0.885]
cDDPM	0.905 [0.894, 0.916]	0.893 [0.880, 0.905]	0.843 [0.825, 0.861]
AnomalyDINO	0.598 [0.579, 0.618]	0.580 [0.559, 0.600]	0.425 [0.403, 0.448]

Appendix

Table 36: Marginal slice-level AUROC (subject mean, 95% interval) under three slice summaries of the normalised map.

Method	b (cm 3 )	validation $\tau$ (hysteresis + pruning)	thresholding $\tau$ at matched FP	Matched realised FP (cm 3 )	Total: hysteresis + pruning $-$ thresholding	Post-processing at fixed $\tau$	Threshold relaxation
REFLECT	5	0.930	0.941	2.77	+0.0219 [+0.0177, +0.0259]	$-$ 0.0008 [ $-$ 0.0024, +0.0000]	+0.0227 [+0.0188, +0.0265]
REFLECT	10	0.870	0.901	5.04	+0.0497 [+0.0419, +0.0569]	$-$ 0.0012 [ $-$ 0.0036, +0.0000]	+0.0510 [+0.0435, +0.0578]
REFLECT	20	0.785	0.825	11.45	+0.0335 [+0.0232, +0.0428]	$-$ 0.0128 [ $-$ 0.0216, $-$ 0.0050]	+0.0463 [+0.0422, +0.0502]
REFLECT	40	0.685	0.741	23.36	+0.0294 [+0.0163, +0.0414]	+0.0013 [ $-$ 0.0099, +0.0115]	+0.0281 [+0.0245, +0.0317]
UCCD	5	0.500	0.593	1.98	+0.0954 [+0.0846, +0.1055]	$-$ 0.0157 [ $-$ 0.0250, $-$ 0.0072]	+0.1111 [+0.1063, +0.1156]
UCCD	10	0.445	0.535	3.67	+0.0939 [+0.0825, +0.1047]	$-$ 0.0079 [ $-$ 0.0177, +0.0010]	+0.1018 [+0.0970, +0.1065]

Appendix

Table 37: Exact decomposition of the gain from hysteresis with pruning at equal realised burden, at every budget (Identity 5 ).

Method	Post-processing	validation $\tau$	Above-floor sens., IoU $\geq$ 0.10 (n = 472)	IoU $\geq$ 0.25	IoU $\geq$ 0.50	Large (n = 307)	Small (n = 165)	Hits / all 5,473 components (not a sensitivity)
REFLECT	thresholding	0.785	0.608	0.479	0.222	0.915	0.036	0.052
REFLECT	hysteresis	0.785	0.593	0.479	0.222	0.906	0.012	0.051
REFLECT	hysteresis + pruning	0.785	0.549	0.460	0.220	0.844	0.000	0.047
UCCD	thresholding	0.385	0.674	0.614	0.186	0.997	0.073	0.058
UCCD	hysteresis	0.385	0.674	0.614	0.186	0.997	0.073	0.058
UCCD	hysteresis + pruning	0.385	0.629	0.591	0.174	0.954	0.024	0.054

Appendix

Table 38: Lesion sensitivity by post-processing at

20\,\mathrm{cm}^{3}

, validation threshold: above-floor sensitivity at three IoU criteria, by lesion size, and hits over all 5,473 components (not a sensitivity).

Scorer	Marginal	Stratified (quintiles)	Matched, 0.2 SD	Matched, 0.05 SD	Regression, df 5	Regression, df 8	Observed $-$ borrowed-map AROC (matched 0.2 SD)
REFLECT	0.887 [0.873, 0.899]	0.754 [0.728, 0.780]	0.734 [0.704, 0.765]	0.726 [0.689, 0.762]	0.600 [0.563, 0.636]	0.551 [0.510, 0.593]	+0.229 [0.201, 0.257]
UCCD	0.907 [0.896, 0.918]	0.759 [0.737, 0.780]	0.769 [0.745, 0.792]	0.743 [0.713, 0.773]	0.649 [0.616, 0.682]	0.505 [0.466, 0.545]	+0.266 [0.244, 0.288]
cDDPM	0.905 [0.894, 0.916]	0.762 [0.740, 0.784]	0.768 [0.744, 0.793]	0.744 [0.714, 0.774]	0.646 [0.612, 0.680]	0.497 [0.457, 0.536]	+0.271 [0.249, 0.293]
AnomalyDINO	0.598 [0.579, 0.618]	0.619 [0.590, 0.649]	0.620 [0.589, 0.652]	0.624 [0.588, 0.662]	0.504 [0.466, 0.542]	0.449 [0.409, 0.489]	+0.124 [0.100, 0.148]
control: brain area	0.884 [0.869, 0.897]	0.691 [0.676, 0.706]	0.668 [0.654, 0.682]	0.560 [0.545, 0.576]	0.626 [0.592, 0.659]	0.719 [0.691, 0.746]	n/a
control: mid-brain closeness	0.857 [0.839, 0.874]	0.576 [0.535, 0.619]	0.550 [0.505, 0.596]	0.486 [0.437, 0.539]	0.530 [0.489, 0.569]	0.545 [0.508, 0.582]	n/a

Appendix

Table 39: Slice-level AUROC: marginal, covariate-adjusted and against the borrowed-map null, with 95% intervals.

Scorer	Matched on area, 0.2 SD	0.1 SD	0.05 SD	Matched on area + position, 0.2 SD
control: brain area	0.668 [0.654, 0.682]	0.610 [0.596, 0.624]	0.560 [0.545, 0.576]	0.929 [0.906, 0.950]
control: mid-brain closeness	0.550 [0.505, 0.596]	0.518 [0.471, 0.568]	0.486 [0.437, 0.539]	0.915 [0.889, 0.939]
REFLECT	0.734 [0.704, 0.765]	0.723 [0.690, 0.756]	0.726 [0.689, 0.762]	0.577 [0.544, 0.611]
UCCD	0.769 [0.745, 0.792]	0.748 [0.721, 0.776]	0.743 [0.713, 0.773]	0.631 [0.598, 0.664]
cDDPM	0.768 [0.744, 0.793]	0.750 [0.722, 0.777]	0.744 [0.714, 0.774]	0.639 [0.606, 0.671]
AnomalyDINO	0.620 [0.589, 0.652]	0.617 [0.584, 0.651]	0.624 [0.588, 0.662]	0.501 [0.471, 0.532]

Appendix

Table 40: Matched-pair AROC by caliper, and matched on brain area and relative position jointly; last row: matched fraction of positive slices.

Method	Connectivity	Floor (mm 3 )	Reference components	Hits	Pooled sensitivity	Subject-mean sensitivity
AnomalyDINO	6	0	5473	99	0.018	0.052
AnomalyDINO	6	50	472	99	0.210	0.264
AnomalyDINO	26	0	1685	100	0.059	0.127
AnomalyDINO	26	50	440	100	0.227	0.275
REFLECT	6	0	5473	287	0.052	0.173
REFLECT	6	50	472	287	0.608	0.765

Appendix

Table 41: Lesion definition and lesion sensitivity at

20\,\mathrm{cm}^{3}

, validation threshold, IoU

\geq 0.10

Method	5 cm 3	10 cm 3	20 cm 3	40 cm 3
REFLECT	0.394 (large 0.603, small 0.006)	0.523 (large 0.801, small 0.006)	0.608 (large 0.915, small 0.036)	0.644 (large 0.967, small 0.042)
UCCD	0.653 (large 0.977, small 0.048)	0.663 (large 0.987, small 0.061)	0.674 (large 0.997, small 0.073)	0.669 (large 0.997, small 0.061)
cDDPM	0.614 (large 0.909, small 0.067)	0.648 (large 0.964, small 0.061)	0.665 (large 0.984, small 0.073)	0.672 (large 0.993, small 0.073)
AnomalyDINO	0.040 (large 0.062, small 0.000)	0.106 (large 0.163, small 0.000)	0.210 (large 0.316, small 0.012)	0.339 (large 0.518, small 0.006)

Appendix

Table 42: Above-floor lesion sensitivity (IoU

\geq 0.10

, plain thresholding, validation threshold) by budget, with large- and small-lesion sensitivity.

Missed-lesion Jaccard	REFLECT	UCCD	cDDPM	AnomalyDINO
REFLECT	1.000	0.813	0.805	0.476
UCCD	0.813	1.000	0.938	0.402
cDDPM	0.805	0.938	1.000	0.416
AnomalyDINO	0.476	0.402	0.416	1.000

Appendix

Table 43: Jaccard index between the sets of above-floor lesions missed at

20\,\mathrm{cm}^{3}

, validation threshold.

Method	Detected (IoU $\geq$ 0.10)	Missed: evidence not selected	Missed: no evidence	Missed: selected, poorly delineated	Total
REFLECT	287	135	15	35	472
UCCD	318	123	17	14	472
cDDPM	314	137	9	12	472
AnomalyDINO	99	170	14	189	472

Appendix

Table 44: Error modes of the 472 above-floor lesions at

20\,\mathrm{cm}^{3}

, validation threshold.

Method	b (cm 3 )	Dice at validation $\tau$	Gain: per-subject $\tau$ upper bound	Gain: per-lesion $\tau$ upper bound	Gain: component selection upper bound	Gain: boundary replacement	Within-lesion coverage at boundary rung
REFLECT	5	0.241	+0.198 [0.182, 0.214]	+0.353 [0.331, 0.373]	$-$ 0.008 [ $-$ 0.016, $-$ 0.001]	+0.349 [0.309, 0.389]	0.231
REFLECT	20	0.497	+0.042 [0.034, 0.050]	+0.153 [0.138, 0.166]	$-$ 0.002 [ $-$ 0.017, 0.013]	+0.400 [0.375, 0.425]	0.439
REFLECT	40	0.554	+0.011 [0.006, 0.015]	+0.108 [0.096, 0.119]	+0.023 [0.002, 0.043]	+0.395 [0.372, 0.417]	0.560
UCCD	5	0.441	+0.081 [0.068, 0.094]	+0.203 [0.191, 0.215]	+0.001 [ $-$ 0.009, 0.011]	+0.527 [0.509, 0.544]	0.314
UCCD	20	0.539	+0.029 [0.022, 0.036]	+0.122 [0.113, 0.132]	+0.015 [ $-$ 0.003, 0.031]	+0.452 [0.436, 0.467]	0.454
UCCD	40	0.551	+0.028 [0.023, 0.033]	+0.113 [0.102, 0.125]	+0.037 [0.019, 0.055]	+0.433 [0.416, 0.450]	0.516

Appendix

Table 45: Label-informed upper bounds at 5, 20 and

40\,\mathrm{cm}^{3}

: Dice gain over the validation threshold. Every bound keeps

\mathrm{FP}_{i}\leq b

; they use the reference labels and are ceilings, not methods.

Method	best attainable Dice	mean FP (cm 3 )	gain over validation $\tau$ at 5	at 20	at 40
REFLECT	0.580	51.6	+0.338 [0.326, 0.351]	+0.083 [0.075, 0.091]	+0.025 [0.021, 0.030]
UCCD	0.585	31.0	+0.145 [0.134, 0.156]	+0.046 [0.041, 0.052]	+0.034 [0.030, 0.039]
cDDPM	0.605	26.3	+0.209 [0.195, 0.223]	+0.086 [0.077, 0.095]	+0.076 [0.068, 0.084]
AnomalyDINO	0.373	155.8	+0.309 [0.293, 0.326]	+0.207 [0.191, 0.223]	+0.143 [0.129, 0.158]

Appendix

Table 46: Per-subject optimal threshold without a burden cap: the best Dice attainable by any single threshold per subject, its mean false-positive volume, and its gain over the validation threshold.

Method	Model	log volume (per SD)	T2 contrast (per SD)	relative z position (per SD)
REFLECT	Firth, profile likelihood	+4.84 [3.73, 6.24]	+1.98 [1.22, 2.82]	$-$ 0.05 [ $-$ 0.43, 0.34]
REFLECT	GEE exchangeable, robust SE	+5.09 [3.65, 6.52]	+2.02 [1.33, 2.72]	$-$ 0.04 [ $-$ 0.44, 0.36]
REFLECT	GLMM, variational Bayes	+5.04 [4.28, 5.79]	+1.94 [1.34, 2.54]	$-$ 0.04 [ $-$ 0.43, 0.36]
REFLECT	GLMM, lme4 glmer (Laplace)	+24.52 [6.70, 42.35]	+6.53 [ $-$ 2.49, 15.55]	+0.72 [ $-$ 3.81, 5.25]
UCCD	Firth, profile likelihood	+6.03 [4.53, 8.31]	+0.76 [ $-$ 0.65, 2.08]	$-$ 0.33 [ $-$ 1.01, 0.28]
UCCD	GEE exchangeable, robust SE	+6.82 [4.30, 9.33]	+0.74 [ $-$ 0.13, 1.62]	$-$ 0.35 [ $-$ 0.84, 0.13]

Appendix

Table 47: Determinants of detection: log-odds of a hit per standard deviation of each predictor under four estimators (472 lesions in 312 subjects). The Laplace fit did not converge for REFLECT; its estimates are not interpretable.

Model $-$ T2 intensity, Dice	5 cm 3	10 cm 3	20 cm 3	40 cm 3
REFLECT	+0.283 [0.263, 0.302]	+0.383 [0.361, 0.405]	+0.437 [0.415, 0.460]	+0.431 [0.408, 0.453]
UCCD	+0.401 [0.386, 0.417]	+0.449 [0.431, 0.467]	+0.461 [0.441, 0.481]	+0.423 [0.401, 0.446]
cDDPM	+0.412 [0.395, 0.429]	+0.452 [0.433, 0.470]	+0.463 [0.441, 0.484]	+0.430 [0.406, 0.454]
AnomalyDINO	+0.044 [0.031, 0.058]	+0.070 [0.053, 0.088]	+0.098 [0.076, 0.119]	+0.113 [0.088, 0.138]

Appendix

Table 48: Model minus z-scored T2 intensity, Dice at matched realised burden, subject mean (unadjusted).

Contrast	Voxel AUROC, subject mean	Voxel AUROC, pooled	Voxel AUPRC, subject mean	Voxel AUPRC, pooled	Dice 20 cm 3 validation $\tau$ , subject mean
FLAIR $-$ REFLECT	+0.018 [0.009, 0.027], p_H $<$ 0.005	$-$ 0.071 [ $-$ 0.085, $-$ 0.057], p_H $<$ 0.005	+0.105 [0.085, 0.126], p_H $<$ 0.005	$-$ 0.127 [ $-$ 0.190, $-$ 0.059], p_H $<$ 0.005	$-$ 0.173 [ $-$ 0.205, $-$ 0.142], p_H 0.006
FLAIR $-$ UCCD	+0.085 [0.076, 0.094], p_H $<$ 0.005	+0.012 [ $-$ 0.005, 0.027], p_H 0.43	+0.108 [0.087, 0.129], p_H $<$ 0.005	$-$ 0.065 [ $-$ 0.110, $-$ 0.019], p_H 0.026	$-$ 0.216 [ $-$ 0.247, $-$ 0.184], p_H 0.006
FLAIR $-$ cDDPM	+0.077 [0.067, 0.086], p_H $<$ 0.005	+0.007 [ $-$ 0.009, 0.022], p_H 0.43	+0.086 [0.065, 0.106], p_H $<$ 0.005	$-$ 0.078 [ $-$ 0.123, $-$ 0.033], p_H 0.007	$-$ 0.195 [ $-$ 0.228, $-$ 0.163], p_H 0.006
FLAIR $-$ AnomalyDINO	+0.118 [0.106, 0.130], p_H $<$ 0.005	+0.048 [0.029, 0.066], p_H $<$ 0.005	+0.405 [0.381, 0.430], p_H $<$ 0.005	+0.229 [0.184, 0.274], p_H $<$ 0.005	+0.158 [0.123, 0.192], p_H 0.006
T2 $-$ REFLECT	$-$ 0.096 [ $-$ 0.103, $-$ 0.088], p_H $<$ 0.005	$-$ 0.130 [ $-$ 0.143, $-$ 0.119], p_H $<$ 0.005	$-$ 0.353 [ $-$ 0.373, $-$ 0.333], p_H $<$ 0.005	$-$ 0.418 [ $-$ 0.464, $-$ 0.367], p_H $<$ 0.005	$-$ 0.411 [ $-$ 0.434, $-$ 0.389], p_H 0.006
T2 $-$ UCCD	$-$ 0.029 [ $-$ 0.039, $-$ 0.018], p_H $<$ 0.005	$-$ 0.048 [ $-$ 0.062, $-$ 0.034], p_H $<$ 0.005	$-$ 0.350 [ $-$ 0.371, $-$ 0.328], p_H $<$ 0.005	$-$ 0.356 [ $-$ 0.382, $-$ 0.328], p_H $<$ 0.005	$-$ 0.454 [ $-$ 0.474, $-$ 0.433], p_H 0.006

Appendix

Table 49: FLAIR and T2 intensity minus each model, Holm-adjusted within the baseline families.

Universe set	Baseline	Model	Share of universes baseline $>$ model	Tied
defensible	T2 intensity	REFLECT	0.1 %	1.7 %
defensible	T2 intensity	UCCD	0.0 %	1.7 %
defensible	T2 intensity	cDDPM	0.2 %	1.8 %
defensible	T2 intensity	AnomalyDINO	27.0 %	7.6 %
defensible	FLAIR intensity	REFLECT	25.1 %	2.2 %
defensible	FLAIR intensity	UCCD	32.5 %	2.2 %

Appendix

Table 50: Share of universes in which a model-free baseline outranks a model.

Score rule (REFLECT)	IXI-HH+Guys: cohort-matched burden	IXI-HH+Guys: subject-matched burden	MOOD+IXI: cohort-matched	MOOD+IXI: subject-matched
MAD (L2)	+0.052 [0.045, 0.058]	+0.053 [0.046, 0.060]	+0.038 [0.033, 0.043]	+0.034 [0.028, 0.040]
robust max (L4)	+0.030 [0.024, 0.035]	+0.022 [0.016, 0.027]	+0.031 [0.027, 0.036]	+0.022 [0.019, 0.026]
Euclidean norm (L3)	+0.029 [0.023, 0.034]	+0.023 [0.018, 0.028]	+0.029 [0.024, 0.033]	+0.022 [0.018, 0.025]
whitened mean (L5)	+0.030 [0.024, 0.035]	+0.021 [0.015, 0.026]	+0.027 [0.022, 0.031]	+0.015 [0.012, 0.019]
mean (L1)	+0.027 [0.022, 0.033]	+0.018 [0.013, 0.023]	+0.028 [0.024, 0.032]	+0.018 [0.015, 0.022]
image only	+0.016 [0.013, 0.020]	+0.014 [0.010, 0.018]	+0.014 [0.011, 0.016]	+0.010 [0.006, 0.014]

Appendix

Table 51: Score rule minus the published REFLECT map, Dice at

20\,\mathrm{cm}^{3}

, matched to the published map’s realised burden over the cohort or per subject; REFLECT trained on IXI-HH+Guys and on MOOD+IXI.

Score rule	b (cm 3 )	own validation $\tau$	Test Dice, mean [95 % CI]	Test Dice, median [95 % CI]	Test FP mean (cm 3 )	Above-floor sensitivity
published map	5	0.93	0.241 [0.223, 0.260]	0.235 [0.207, 0.256]	3.3 [1.5, 6.0]	0.394
published map	20	0.79	0.492 [0.468, 0.516]	0.542 [0.510, 0.569]	15.7 [11.6, 21.2]	0.606
published map	40	0.69	0.553 [0.530, 0.576]	0.604 [0.577, 0.628]	33.9 [28.3, 41.3]	0.644
MAD (L2)	5	0.70	0.364 [0.344, 0.384]	0.385 [0.353, 0.411]	4.5 [3.0, 6.6]	0.521
MAD (L2)	20	0.57	0.544 [0.521, 0.566]	0.597 [0.570, 0.624]	15.3 [11.5, 20.2]	0.631
MAD (L2)	40	0.47	0.599 [0.577, 0.621]	0.653 [0.626, 0.680]	33.3 [27.8, 40.3]	0.655

Appendix

Table 52: Score rules at their own validation thresholds (0.01 grid), REFLECT trained on IXI-HH+Guys.

Matching	Growth	b (cm 3 )	MAD $-$ published	Best other family member	Best other $-$ published	Verdict
cohort	0	5	+0.068 [0.056, 0.079]	whitened mean (L5)	+0.073 [0.062, 0.085]	MAD demoted
cohort	0	10	+0.048 [0.039, 0.058]	whitened mean (L5)	+0.040 [0.031, 0.049]	MAD demoted
cohort	0	20	+0.052 [0.045, 0.058]	robust max (L4)	+0.030 [0.024, 0.035]	MAD retained
cohort	0	40	+0.046 [0.042, 0.051]	robust max (L4)	+0.027 [0.023, 0.030]	MAD retained
cohort	1	5	+0.057 [0.043, 0.070]	Euclidean norm (L3)	+0.059 [0.045, 0.072]	MAD demoted
cohort	1	10	+0.034 [0.023, 0.046]	whitened mean (L5)	+0.010 [ $-$ 0.000, 0.021]	MAD retained

Appendix

Table 53: Retention criterion for MAD (REFLECT, IXI-HH+Guys): MAD is retained only if its interval does not overlap that of the best other aggregation. Growth: 1 with hysteresis, 0 without.

Model (training set)	Small-lesion definition	Lesions	Subjects	Local AUPRC [95 % CI] (chance 0.5)	Local AUROC [95 % CI]
REFLECT (MOOD+IXI)	pixel	122	71	0.531 [0.511, 0.554]	0.537 [0.516, 0.558]
REFLECT (MOOD+IXI)	physical	133	77	0.557 [0.533, 0.584]	0.561 [0.538, 0.587]
cDDPM (IXI-HH+Guys)	pixel	100	62	0.535 [0.504, 0.567]	0.484 [0.432, 0.536]
cDDPM (IXI-HH+Guys)	physical	127	81	0.584 [0.552, 0.617]	0.538 [0.491, 0.584]
UCCD (IXI-HH)	pixel	120	74	0.509 [0.484, 0.536]	0.495 [0.462, 0.528]
UCCD (IXI-HH)	physical	136	85	0.544 [0.514, 0.575]	0.527 [0.493, 0.562]

Appendix

Table 54: Local contrast of small lesions against an equal-area surrounding region (maps produced with the uncorrected mapping; chance level 0.5). Small lesions defined in pixels or in physical volume.

Method	Phase	Wall clock	GPU-hours	s per volume
REFLECT	train (stage 1)	6d 17:25:53	484.29
REFLECT	train (stage 2)	4d 07:31:24	207.05
REFLECT	inference	03:07:18	3.12	32.5
AnomalyDINO	memory-native build	00:01:07	0.02
AnomalyDINO	inference	00:15:23	0.26	2.7
IterMask3D	train	09:15:03	18.50

Appendix

Table 55: Computational cost (Tesla V100-SXM2-32GB). Dinomaly2 times are approximate and were not taken from job accounting.

Training set	Sources	REFLECT	UCCD	cDDPM	AnomalyDINO	REFLECT $-$ UCCD
IXI-HH	IXI Hammersmith, T2	0.425	0.500	0.467	0.148
IXI-Guys	IXI Guy’s, T2	0.407	0.576	0.507	0.158
IXI-HH+Guys	IXI, both sites, T2	0.497	0.539	0.519	0.166	$-0.042$ [ $-0.060$ , $-0.025$ ]
MOOD+IXI-HH	MOOD, IXI Hammersmith, T1 + T2	0.521	0.496	0.417	0.152
MOOD+IXI	MOOD, all IXI, T1 + T2	0.552	0.452	0.216	0.161	$+0.100$ [0.079, 0.121]

Appendix

Table 56: Dice in the reference universe (

20\,\mathrm{cm}^{3}

, validation threshold, subject mean) when all four methods share one training set, and the REFLECT

-

UCCD contrast. T2 and FLAIR do not depend on training data (0.085 and 0.324).

Comparison set vs IXI-HH+Guys	Configurations	All metrics	Dice	Voxel AUROC	Voxel AUPRC	Lesion sens.	Slice AUROC marg.	Slice AUROC adj.
test-selected set	REFLECT (MOOD+IXI), UCCD (IXI-Guys), cDDPM (IXI-HH+Guys), AnomalyDINO (MOOD+IXI-HH)	69.3 %	70.8 %	100.0 %	100.0 %	28.3 %	16.7 %	100.0 %
provisional set	REFLECT (MOOD+IXI), UCCD (IXI-HH), cDDPM (IXI-HH+Guys), AnomalyDINO (MOOD+IXI)	50.4 %	81.5 %	0.0 %	50.0 %	54.0 %	83.3 %	33.3 %
validation-selected set	REFLECT (IXI-HH+Guys), UCCD (IXI-Guys), cDDPM (IXI-HH+Guys), AnomalyDINO (MOOD+IXI-HH)	62.0 %	50.5 %	100.0 %	100.0 %	21.5 %	0.0 %	100.0 %
validation-tuned set	REFLECT (MOOD+IXI), UCCD (IXI-Guys), cDDPM (IXI-Guys), AnomalyDINO (IXI-HH+Guys)	81.4 %	71.8 %	100.0 %	100.0 %	33.2 %	83.3 %	100.0 %
all IXI-HH	all IXI-HH	79.0 %	71.8 %	100.0 %	100.0 %	52.3 %	100.0 %	50.0 %
all IXI-Guys	all IXI-Guys	71.9 %	75.0 %	100.0 %	100.0 %	39.7 %	16.7 %	100.0 %

Appendix

Table 57: Share of defensible universes in which the four-method order differs from the order with IXI-HH+Guys, per comparison set and metric.

Metric subset	Method share	Configuration main effect	Method $\times$ configuration
all metrics	0.129 [0.116, 0.141]	0.005 [0.004, 0.006]	0.018 [0.016, 0.020]
Dice	0.617 [0.588, 0.637]	0.028 [0.022, 0.033]	0.115 [0.104, 0.127]
voxel AUROC	0.849 [0.825, 0.868]	0.043 [0.033, 0.053]	0.068 [0.057, 0.080]
voxel AUPRC	0.764 [0.734, 0.786]	0.056 [0.045, 0.067]	0.142 [0.125, 0.161]
lesion sensitivity	0.118 [0.108, 0.128]	0.004 [0.003, 0.005]	0.020 [0.018, 0.022]
slice AUROC, marginal	0.919 [0.901, 0.931]	0.006 [0.004, 0.008]	0.038 [0.031, 0.048]

Appendix

Table 58: Variance shares with the training set as an eleventh axis (five training sets; defensible universes, 1,000 resamples).

Common training set	Metric (subject mean)	Contrast	Estimate [95 % CI]	p_Holm
all IXI-HH+Guys	Dice	REFLECT $-$ UCCD	$-$ 0.042 [ $-$ 0.060, $-$ 0.025]	$<$ 0.005
all IXI-HH+Guys	Dice	REFLECT $-$ cDDPM	$-$ 0.022 [ $-$ 0.042, $-$ 0.002]	0.062
all IXI-HH+Guys	Dice	UCCD $-$ cDDPM	+0.020 [0.011, 0.029]	$<$ 0.005
all IXI-HH+Guys	voxel AUROC	REFLECT $-$ UCCD	+0.067 [0.060, 0.074]	$<$ 0.005
all IXI-HH+Guys	voxel AUROC	REFLECT $-$ cDDPM	+0.059 [0.051, 0.066]	$<$ 0.005
all IXI-HH+Guys	voxel AUROC	UCCD $-$ cDDPM	$-$ 0.008 [ $-$ 0.011, $-$ 0.006]	$<$ 0.005

Appendix

Table 59: Paired contrasts with a common training set (all methods on IXI-HH+Guys, all on MOOD+IXI), subject mean at

20\,\mathrm{cm}^{3}

, validation threshold, Holm-adjusted. On MOOD+IXI cDDPM was trained on 72% T1 volumes.

Method	Config	validation median slice Dice [95 % CI]	validation pooled Dice	validation median volume Dice	validation canonical Dice 20 cm 3	validation canonical AUPRC	test canonical Dice 20 cm 3 (validation $\tau$ )	test voxel AUPRC	test regret vs best config	chosen by	primary
AnomalyDINO	IXI-HH	0.198 [0.159, 0.232]	0.319	0.309	0.176	0.293	0.148	0.312	0.018	AUPRC
AnomalyDINO	IXI-Guys	0.190 [0.155, 0.228]	0.317	0.305	0.175	0.277	0.158	0.310	0.007	none
AnomalyDINO	IXI-HH+Guys	0.227 [0.195, 0.251]	0.320	0.297	0.189	0.286	0.166	0.311	0.000	pooled Dice, Dice at 20 cm 3	yes
AnomalyDINO	MOOD+IXI-HH	0.247 [0.218, 0.273]	0.309	0.291	0.164	0.261	0.152	0.297	0.013	slice Dice, test
AnomalyDINO	MOOD+IXI	0.169 [0.131, 0.208]	0.309	0.316	0.180	0.273	0.161	0.289	0.005	volume Dice
REFLECT	IXI-HH	0.538 [0.516, 0.564]	0.524	0.479	0.404	0.509	0.425	0.569	0.127	none

Appendix

Table 60: All twenty method–training-set combinations: validation criteria of the six selection rules, test Dice and AUPRC, and test regret relative to the method’s best training set (Dice at

20\,\mathrm{cm}^{3}

, validation threshold, subject mean). “test” denotes the median-slice-Dice rule applied to the test cohort.

Selection rule	Method	Point choice	Bootstrap frequency of point choice	Test regret of point choice	Mean test regret over resamples	2.5th to 97.5th pct	Label
validation median slice Dice	REFLECT	IXI-HH+Guys	36.8 %	0.055	0.031	[0.000, 0.055]	unstable
validation median slice Dice	UCCD	IXI-Guys	97.8 %	0.000	0.001	[0.000, 0.000]	stable
validation median slice Dice	cDDPM	IXI-HH+Guys	57.1 %	0.000	0.010	[0.000, 0.052]	unstable
validation median slice Dice	AnomalyDINO	MOOD+IXI-HH	27.3 %	0.013	0.008	[0.000, 0.018]	unstable
validation pooled Dice	REFLECT	MOOD+IXI	86.4 %	0.000	0.006	[0.000, 0.055]	stable
validation pooled Dice	UCCD	IXI-Guys	99.3 %	0.000	0.000	[0.000, 0.000]	stable

Appendix

Table 61: Training-set selection on the 24 validation subjects: point choice, its frequency over

10^{4}

bootstrap resamples and its test regret.

Hypothesis	provisional set	test-selected set	validation-selected set	primary
ranking instability	supported (max. flip share 0.302 [0.278, 0.374])	supported (0.459 [0.357, 0.497])	supported (0.472 [0.393, 0.498])	supported (0.472 [0.414, 0.499])
evaluation dominance	supported (0.607 vs 0.173)	supported (0.598 vs 0.184)	supported (0.613 vs 0.169)	supported (0.614 vs 0.162)
intensity baseline	not	not	not	not
contrast dependence	not	not	not	partly tested; not
pruning trade-off	not	not	not	not (point estimates)
test-tuning optimism	REFLECT only	REFLECT (MOOD+IXI) only (lower bound 0.022)	none (lower bound 0.019)	not (lower bound 0.0191 REFLECT, 0.0199 cDDPM)

Appendix

Table 62: Verdicts of H1 to H9 in every set of training data.

Secondary hypothesis	Estimate	Verdict
UCCD, $20\,\mathrm{cm}^{3}$ : the threshold-relaxation term is positive, its interval excludes 0, and it exceeds the fixed-threshold term	$+0.078$ [0.073, 0.083] vs $+0.006$	held
AnomalyDINO, $20\,\mathrm{cm}^{3}$ : the total gain from hysteresis with pruning is $\leq 0$	$-0.001$ [ $-0.011$ , 0.010]	held
Above-floor sensitivity falls with hysteresis and pruning (UCCD, AnomalyDINO)	$-0.045$ ; $-0.089$	held
UCCD small-lesion sensitivity $<0.10$	0.073	held
Jaccard index of the missed-lesion sets of UCCD with REFLECT and with cDDPM $\geq 0.70$	0.813; 0.938	held
UCCD transfer gap at $5\,\mathrm{cm}^{3}$ $>0$ , interval excludes 0	$-0.008$ [ $-0.008$ , $-0.007$ ]	failed

Appendix

Table 63: Secondary hypotheses and their outcome (all methods trained on IXI-HH+Guys).

Explore similar work

Sep 21, 2026cs.AI

Unsupervised Brain Anomaly Detection as a Bayesian Inverse Problem with Diffusion Prior

Unsupervised anomaly detection (UAD) aims to localize abnormal regions in medical scans without pixel-level annotations. A typical strategy seeks to reconstruct a pseudo-healthy image that preserves subject-specific anatomy. Recently, diffusion models have been proposed to perform UAD. However, these methods rely on heuristic noise schedules or synthetic corruptions to balance subject-specificity and anomaly removal. In this work, we propose an alternative formulation of UAD as a Bayesian inverse problem under a diffusion prior. First, we introduce a latent spatial anomaly mask that models pixel-wise consistency between a test image and its latent corresponding pseudo-healthy image. Then, we propose an approximation of the unknown generation process that links healthy anatomy, anomalies, and the observed image, enabling a well-defined likelihood within the Bayesian framework. Building on recent advances in diffusion-based inverse problem methods, we jointly infer the pseudo-healthy image and the anomaly mask via annealed posterior sampling. We evaluate our approach on FDG PET (ADNI) and FLAIR MRI (BraTS 2021), demonstrating improved anomaly localization performance compared to other diffusion-based approaches and validating the contribution of our introduced model. Our code is available at https://github.com/HuguesRoy/UAD_DAPS.

Hugues Roy, Reuben Dorent, Ninon Burgos

Sorbonne Université, Institut du Cerveau - Paris Brain Institute - ICM, CNRS, Inria, Inserm, AP-HP, Hôpital de la Pitié Salpêtrière, F-75013, Paris, France

Graph Anomaly Detection Diffusion Priors Anatomical Priors Bayesian Inverse Problems

Ask PDF

Aug 17, 2026cs.CV

Unsupervised Anomaly Detection for Image Dataset Quality Assurance in Multi-Center Breast MRI

Corrupted, inconsistent, or anomalous data silently threatens the safety and reliability of medical AI. Despite growing regulatory recognition of dataset quality assurance (QA) for high-risk medical AI, scalable automated detection remains underdeveloped. We employ unsupervised anomaly detection (AD) and out-of-distribution (OOD) detection as an automated dataset QA mechanism for multi-center dynamic contrast-enhanced breast MRI. We build a controlled AD benchmark of 17 realistic QA-relevant anomaly types from six public datasets (protocol violations, processing errors, incorrect anatomical regions) and propose a taxonomy of radiological image anomalies based on human visual perception, enabling fine-grained analysis of AD failure modes. The benchmark includes near-, medium-far-, far-OOD samples, as well as in-distribution and external normal data. Four methods are evaluated: a projection-based method extended with a domain-specific feature extractor and a novel positional encoding, a reconstruction-based approach extended to full 3D volumes with an augmented training objective, and two unmodified hybrid OOD detection methods. Medium-far- and far-OOD samples are detected reliably, whereas near-OOD samples and external normal data from unseen institutions expose method-specific differences. The 3D reconstruction-based approach best balances detection performance (AUROC: 0.936) and generalization to unseen institutions. The projection-based method with positional encoding achieves the highest overall detection performance (AUROC: 0.954). Both hybrid methods exhibit critical failure modes, confirming that methods validated for one modality or anatomy may not generalize without domain-specific adaptation. Implants and mastectomies remain an open challenge for all methods. Our results establish a foundation and practical guidance on scalable unsupervised QA in medical AI pipelines.

Chiara Tappermann, Steffen Renisch, Lars Ole Schwen +3

Fraunhofer Institute for Digital Medicine MEVIS, Max-Von-Laue-Straße 2, Bremen, 28359, Bremen, Germany · University of Bremen, Bibliothekstraße 1, Bremen, 28359, Bremen, Germany · Institute of Diagnostic and Interventional Radiology, Hannover Medical School, Carl-Neuberg-Str. 1, Hannover, 28359, Lower Saxony, Germany

Graph Anomaly Detection Detection Performance Image Quality Assessment Data Quality

Ask PDF

May 23, 2026physics.med-ph

Catching magnetic resonance imaging outliers in artificial intelligence-supported radiotherapy workflows: unsupervised detection and localization of image anomalies using deep learning

Artificial intelligence is increasingly integrated into radiotherapy workflows, yet such pipelines remain vulnerable to out-of-distribution image data that may introduce unexpected behavior in clinical tasks. Deep learning-based anomaly detection for pelvic magnetic resonance imaging (MRI) remains largely unexplored, and transparent evaluation of its feasibility for full automation is limited. We developed and evaluated a fully automated, unsupervised anomaly-detection framework for pelvic and brain MRI. A two-stage framework was trained on reference images from public datasets: LUND-PROBE for pelvic MRI, and IXI, fastMRI, and fastMRI+ for brain MRI. In the first stage, MRI slices were compressed into discrete tokens; in the second, the distribution of normal tokens was modeled. Anomaly evidence was estimated by combining perceptual image differences with token-surprisal scores based on negative log-likelihood. Automated detection was evaluated on pelvic MRI with synthetic global and real clinical anomalies, and on brain MRI with clinically annotated fastMRI+ abnormalities. Sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and false-positive behavior in held-out normal cases were assessed. The framework achieved robust detection across hidden evaluation cohorts, with AUCs of 0.97 (95% CI, 0.95-0.98) and 0.81 (95% CI, 0.74-0.87) for pelvic and brain MRI, respectively. Heatmap analysis showed strong spatial agreement between detected anomalies and ground-truth locations, supporting localization accuracy and interpretability. These results support the potential of unsupervised anomaly detection as an automated MRI quality-control layer for radiotherapy workflows, with transparent visualization of image regions likely to compromise downstream AI-based tasks.

Mustafa Kadhim, Viktor Rogowski, Emilia Persson +7

Medical Radiation Physics, Lund University, Lund, Sweden. · Radiation Physics, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden. · Medical Radiation Physics, Department of Translational Medicine, Lund University, Malmö, Sweden +4

Magnetic Resonance Imaging Graph Anomaly Detection Radiotherapy Unsupervised Detection Deep Learning

Ask PDF