Region-aware error triage for medical image segmentation with groupwise conformal risk control

Nishi Kanta Paul; Md Shihabul Islam Shovo; Hoileong Lee

cs.CVNov 6, 2026

Region-aware error triage for medical image segmentation with groupwise conformal risk control

Authors: Nishi Kanta Paul, Md Shihabul Islam Shovo, Hoileong Lee

Organizations: NIMISHES Lab, Dhaka, Bangladesh · Faculty of Electronic Engineering and Technology, Universiti Malaysia Perlis, 02600 Arau, Perlis, Malaysia

Abstract

Segmentation masks are increasingly reviewed rather than redrawn, which requires knowing where a mask is likely wrong. Learned error maps rank pixels by risk, but the number of pixels to warn is chosen ad hoc, and calibrating one warning threshold over all error pixels may leave errors far from the predicted contour under-flagged. We propose RACE-Triage, a post-hoc error triage framework that leaves the segmentation mask unchanged. A risk network is trained on out-of-fold predictions from backbone-independent inputs. The predicted contour partitions each image into a boundary band and a non-boundary region, and conformal risk control is applied separately in each region to bound the expected fraction of missed errors among cases with errors there. On ISIC 2018 dermoscopy and BUSI breast ultrasound (where patient identifiers are unavailable, so independence across images is assumed), pooled control met its overall target but missed on average 31.8% and 36.4% of non-boundary errors. Regional control reduced these fractions to 19.3% and 20.9% and raised the recall of size-filtered false-negative components from 0.722 to 0.818 and from 0.611 to 0.763. The mean warned area grew by 1.6 and 4.9 percentage points, and its 95th percentile reached 23.5% and 38.5%. The effect held across three seeds and a second backbone, and area-matched heuristics did not reproduce it. On the external PH $^2$ set, frozen thresholds exceeded both targets, whereas recalibration with 100 local images brought the mean over repeated splits below them. Regional calibration improves non-boundary error coverage at a measured increase in warned area.

Figures & tables

Method family	Produces or controls	RACE-Triage instead
Uncertainty estimation [ 10 , 13 , 14 ]	Pixel-wise uncertainty scores	Calibrates warning regions for a given score
Error and quality prediction [ 17 , 18 , 20 ]	Error maps or case-level quality scores	Converts an error map into warnings with a stated risk
Conformal segmentation [ 23 , 24 , 25 ]	Prediction sets or quality ranges for the segmentation	Leaves the mask unchanged
Conformal false-negative or false-discovery control [ 47 , 48 , 49 ]	Missed or spurious parts of the predicted object	Controls the share of all segmentation errors left unwarned
Group-wise risk control [ 26 , 29 ]	Risk per predefined group, e.g., per organ	Spatial groups derived from the predicted contour

Table 1: Positioning of RACE-Triage relative to related method families.

Figure 1: Overview of RACE-Triage. Top: out-of-fold error learning with backbone-independent inputs. Middle: deployment, in which the predicted contour splits each image into $G_{B}$ and $G_{N}$ and each group is warned by its own calibrated rule. Bottom: per-group conformal risk control on error-positive calibration cases. Map panels are schematic.

Dataset	Modality	Images	Partition sizes	Test Dice	Error pixels (%)	Cal. $n_{B}/n_{N}$	Test $n_{B}/n_{N}$	$G_{N}$ share (%)
ISIC 2018	Dermoscopy	2594	1686 / 259 / 324 / 325	0.883	3.2	322 / 128	323 / 131	21.6
BUSI	Breast ultrasound	780	507 / 78 / 97 / 98	0.751	5.6	82 / 49	83 / 52	37.9
PH 2	Dermoscopy (external)	200	– / – / – / 200	0.862	5.4	–	200 / 118	33.4

Table 2: Datasets, partitions and error audit of the frozen segmenters. Partitions are given as

\mathcal{D}_{\mathrm{fit}}/\mathcal{D}_{\mathrm{select}}/\mathcal{D}_{\mathrm{cal}}/\mathcal{D}_{\mathrm{test}}

n_{B}/n_{N}

counts error-positive cases in the boundary and non-boundary groups. Dice is the mean over test images with a non-empty reference mask; the last column gives the share of test error pixels that lie in

G_{N}

			ISIC 2018					BUSI
Error score	Params (M)	ms	AUPRC $\uparrow$	$q\uparrow$	Brier $\downarrow$	CR@10 $\uparrow$	$N$ @10 $\uparrow$	AUPRC $\uparrow$	$q\uparrow$	Brier $\downarrow$	CR@10 $\uparrow$	$N$ @10 $\uparrow$
Predictive entropy	0	0.1	0.318	0.192	–	0.512	0.214	0.276	0.163	–	0.384	0.152
Test-time augmentation	0	31.4	0.336	0.205	–	0.538	0.236	0.291	0.172	–	0.405	0.169
Fold-ensemble disagreement	38.8	19.7	0.352	0.221	–	0.561	0.268	0.309	0.188	–	0.431	0.204
Local head (P1)	0.01	0.4	0.394	0.251	2.48	0.603	0.331	0.341	0.207	4.71	0.466	0.248
P1 + distance $d$ (P2)	0.01	1.1	0.405	0.262	2.44	0.618	0.352	0.350	0.214	4.66	0.479	0.262

Table 3: Error-score quality on the held-out test sets with identical frozen segmentations. The first block contains uncertainty baselines that leave the segmenter unchanged: predictive entropy, test-time augmentation (entropy of the mean over 8 augmented views) and the disagreement of the five fold backbones

f_{k}

(parameters of the additional segmenters). AUPRC is computed for raw errors and, in the columns headed

q

, for

q

-distant errors (

q=2

px); Brier scores are

\times 10^{2}

and are not defined for unnormalized uncertainty scores. CR@10 is the size-filtered component recall and

N

@10 the recall of non-boundary error pixels when exactly 10% of each image is warned. Latency includes geometry computation (ms per

256\times 256

image).

Figure 2: Size-filtered component recall versus warned area for top-ranked warnings of each error score. Markers show the calibrated pooled and regional policies on the default risk map.

Figure 3: ISIC development pilot over 50 pseudo-calibration/pseudo-test splits. (a) Mean missed-error fraction per group. (b) Mean warned area per split; lines connect the two policies on the same split.

ISIC 2018 (in-domain, $n=325$ )
Policy	$\bar{M}$	$\bar{M}_{B}$	$\bar{M}_{N}$	Area (%)	Corr. (%)	Reg.	FP rec.	FN rec.
Pooled CRC, fixed	0.181	0.137	0.318	9.8	7.3	3.4	0.802	0.722
Regional CRC, fixed (ours)	0.190	0.188	0.193	11.4	8.9	4.7	0.829	0.818
Pooled CRC, adaptive	0.183	0.142	0.301	9.3	6.7	3.1	0.795	0.731
Regional CRC, adaptive	0.192	0.190	0.197	10.6	8.1	4.2	0.824	0.812
Top-ranked, 10% area	0.174	0.121	0.352	10.0	7.5	3.6	0.789	0.716

Table 4: Warning policies on one frozen risk map (

\delta_{B}=\delta_{N}=\delta=0.20

\bar{M}

\bar{M}_{B}

and

\bar{M}_{N}

are mean missed-error fractions over error-positive test cases (pooled, boundary and non-boundary). Area is the warned share of the image and Corr. the share of the image that is correctly segmented but warned; Reg. is the number of connected warning regions per image; FP/FN are size-filtered component recalls (

\rho=0.10

) at each policy’s own warned area, unlike CR@10 in Table 3 , which fixes the area at 10%. Top-ranked rows use exact per-image budgets and carry no calibration guarantee.

Figure 4: Held-out comparison on one frozen risk map. (a,b) Mean missed-error fraction per group with 95% bootstrap intervals; the dashed line is $\delta=0.20$ . (c) Distribution of per-image warned area for fixed-threshold pooled and regional CRC; long bars mark the means, white dots the medians and short bars the 95th percentiles.

		Primary split						Re-partitions
Policy	Guarantee	$\bar{M}$	$\bar{M}_{B}$	$\bar{M}_{N}$	Area (%)	Reg.	FN rec.	Mean $\bar{M}_{N}$	$>\delta$ (%)
ISIC 2018
Regional CRC, fixed (ours)	group	0.190	0.188	0.193	11.4	4.7	0.818	0.194	41
Regional plug-in (no $1/(n_{g}{+}1)$ term)	–	0.194	0.191	0.199	11.2	4.6	0.812	0.201	52
Fixed group quotas	–	0.219	0.214	0.236	11.4	4.1	0.792	0.231	89
Distance-boosted ranking + pooled CRC	pooled	0.186	0.171	0.238	11.0	4.3	0.781	0.234	91

Table 5: Regional CRC compared with spatial allocation rules on the same risk map (

\delta=0.20

). Guarantee indicates which expected missed-error fraction is controlled. For fixed group quotas, the ratio of boundary to non-boundary warning shares was tuned on

\mathcal{D}_{\mathrm{tune}}

, while the overall quota level was retrospectively matched to the regional CRC policy’s mean test warned area; as evaluated here, this baseline is diagnostic rather than a prospectively specified deployment policy. The distance-boosted ranking adds a constant to the risk scores in

G_{N}

before pooled CRC. The per-image matched diagnostic baseline warns, in each test image, exactly as many top-ranked pixels as the regional policy. The last two columns summarize 500 random re-partitions of

\mathcal{D}_{\mathrm{cal}}\cup\mathcal{D}_{\mathrm{test}}

into halves of the original sizes: the mean of

\bar{M}_{N}

and the percentage of re-partitions in which

\bar{M}_{N}

exceeded

\delta

ISIC 2018
				Pooled	Regional		Area (%)
Setting	Test Dice	$\hat{t}_{B}$	$\hat{t}_{N}$	$\bar{M}_{N}$	$\bar{M}_{B}$	$\bar{M}_{N}$	Pooled	Regional
U-Net, seed 1 (main results)	0.883	0.260	0.120	0.318	0.188	0.193	9.8	11.4
U-Net, 3 seeds (mean $\pm$ SD)	0.881 $\pm$ 0.003	0.258 $\pm$ 0.008	0.122 $\pm$ 0.003	0.309 $\pm$ 0.012	0.189 $\pm$ 0.003	0.197 $\pm$ 0.007	9.9 $\pm$ 0.3	11.6 $\pm$ 0.4
DeepLabV3+, own $h$	0.897	0.270	0.110	0.296	0.186	0.191	8.9	10.3
DeepLabV3+, U-Net $h$ , recalibrated	0.897	0.245	0.095	0.327	0.190	0.198	10.6	12.6

Table 6: Robustness of the pooled-versus-regional comparison to retraining and to the segmentation backbone (

\delta=0.20

, fixed thresholds). Each seed retrains the five fold backbones, the final segmenter and

h

, followed by a new calibration on the same

\mathcal{D}_{\mathrm{cal}}

. DeepLabV3+ is paired either with its own risk network, trained on its own out-of-fold predictions, or with the U-Net risk network, in which case only the thresholds are recalibrated.

\hat{t}_{g}=1-\hat{\lambda}_{g}

are the calibrated regional thresholds.

ISIC 2018 (US-Loss segmenter, test Dice 0.886)
		Pooled CRC		Regional CRC
Risk score	AUPRC	$\bar{M}_{N}$	Area	$\bar{M}_{N}$	Area
US-Loss entropy	0.361	0.402	10.6	0.198	16.9
Multiscale $h$ (refitted)	0.452	0.309	9.6	0.194	11.2
BUSI (US-Loss segmenter, test Dice 0.758)
US-Loss entropy	0.318	0.441	16.3	0.211	27.8

Table 7: Boundary-Guided US-Loss [ 14 ] as an alternative segmenter and risk score. Both scores are computed on the US-Loss segmentation, so the policies differ only in the ranking they calibrate (

\delta=0.20

, fixed thresholds). AUPRC refers to raw errors.

Figure 5: Warned area of regional CRC with fixed and load-adaptive families by tertile of predicted error load (tertile boundaries fixed on development data). Numbers above the bars give $\bar{M}_{N}$ .

Policy	$\bar{M}_{B}$	$\bar{M}_{N}$	Area	FN rec.	$>\delta$ (%)
Pooled, frozen ISIC	0.162	0.372	12.4	0.649	100
IQR	0.146–0.179	0.341–0.401	11.3–13.5
Regional, frozen ISIC	0.215	0.263	13.9	0.751	94
IQR	0.196–0.233	0.241–0.287	12.6–15.1
Pooled, recalibrated	0.144	0.349	13.4	0.668	100
IQR	0.129–0.160	0.318–0.379	12.0–14.9

Table 8: Target-site recalibration on PH 2 . Values are means over 100 stratified random splits into 100 calibration and 100 test images; frozen rows apply the ISIC thresholds to the same test halves. IQR rows give the interquartile range over splits. The last column is the percentage of splits in which

\bar{M}_{N}

exceeded

\delta=0.20

; the guarantee concerns the mean over splits, not each split.

Variant	$\bar{M}_{B}$	$\bar{M}_{N}$	Area	Reg.	FN rec.
Default ( $b=5$ px)	0.188	0.193	11.4	4.7	0.818
Narrower band ( $b=3$ px)	0.191	0.196	11.9	5.2	0.821
No distance channel in $h$	0.187	0.195	12.1	4.9	0.806
Local head P1 as $h$	0.190	0.198	17.8	7.6	0.734
Overlay dilated by 2 px	0.151	0.172	15.3	2.9	0.861
Union with component CRC	0.176	0.181	12.9	5.0	0.897

Table 9: Ablation of the regional fixed-threshold policy on ISIC 2018 test data. Each variant is recalibrated on

\mathcal{D}_{\mathrm{cal}}

except the two superset constructions, which enlarge the calibrated overlay.

Figure 6: Non-boundary missed fraction versus warned area for $\delta\in\{0.10,0.20,0.30\}$ and $b\in\{3,5,10\}$ pixels. Colours encode $\delta$ (dashed lines) and marker shapes the band width; arrows connect pooled CRC to regional CRC with $b=5$ at the same $\delta$ . Every setting is recalibrated on $\mathcal{D}_{\mathrm{cal}}$ with the same risk network.

Figure 7: Test cases from ISIC 2018 and BUSI. (a,b) Cases with the median paired reduction in $M_{N}$ ; (c,d) cases with the largest regional $M_{N}$ (failure cases). From left: image with reference (solid) and predicted (dashed) contours; errors over the groups $G_{B}$ and $G_{N}$ ; risk map; warnings of pooled and regional CRC, with unwarned errors in red. In (d) the prediction is empty.

Dataset	$k_{B}/k_{N}$	Viol. $B$	Viol. $N$	Either	Area (%)
ISIC 2018	16 / 6	0.046	0.053	0.058	23.7
		(0.026–0.075)	(0.022–0.107)	(0.036–0.090)
BUSI	4 / 2	0.060	0.077	0.082	38.6
		(0.020–0.135)	(0.021–0.185)	(0.036–0.155)

Table 10: Secondary tail control (

\alpha_{B}=\alpha_{N}=0.05

, fixed thresholds). Violation rates are the shares of error-positive test cases with

M_{g}>\delta

, with exact 95% intervals below; “Either” is the per-case union rate, bounded by 0.10.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

		ISIC 2018			BUSI
$\rho$	$s$	Pooled	Regional	$\Delta$	Pooled	Regional	$\Delta$
0.10	25	0.722	0.818	+0.096	0.611	0.763	+0.152
0.25	25	0.583	0.709	+0.126	0.472	0.642	+0.170
0.50	25	0.412	0.537	+0.125	0.318	0.469	+0.151
0.10	100	0.701	0.815	+0.114	0.598	0.771	+0.173
0.50	100	0.368	0.511	+0.143	0.291	0.447	+0.156

Appendix

Table 11: Size-filtered false-negative component recall of pooled and regional CRC under stricter detection criteria (fixed thresholds,

\delta=0.20

\rho

is the required overlap

|C\cap W|/|C|

and

s

the minimum component size in pixels; the first row is the default criterion of Table 4 .

Error type in $G_{N}$
Subset	$n$	Pooled $\bar{M}_{N}$	Regional $\bar{M}_{N}$	Area (%)
ISIC 2018, FP pixels	74	0.274	0.171	–
ISIC 2018, FN pixels	96	0.341	0.204	–
BUSI, FP pixels	38	0.318	0.188	–
BUSI, FN pixels	33	0.402	0.226	–
BUSI diagnostic category

Appendix

Table 12: Non-boundary missed-error fraction by error type and by BUSI category (fixed thresholds,

\delta=0.20

n

is the number of test cases with errors of the given type, or of any type, in

G_{N}

; areas are averaged over all test images of a category. These are empirical subgroup summaries without a separate guarantee.

Medical Image Segmentation

Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.

See all 822 papers

Bound-Aware Per-Organ Recall Risk Control for Multi-Organ CT Segmentation under Clinical Domain Shift
Aug 18, 2026Souraj Adhikary, Negar Chabi, Andre MastmeyerMulti-Organ Segmentation Medical Image Segmentation
Department of Engineering Sciences, Jade Hochschule Wilhelmshaven, Germany
A Decision-Aware Evaluation Protocol for Uncertainty in Medical Image Segmentation
Oct 15, 2026Saket Maganti, Shreyank N GowdaUncertainty Quantification Segmentation Uncertainty Estimation
Independent Researcher · School of Computer Science, University of Nottingham, Nottingham, UK
Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation
Jul 23, 2026Jingguo Qu, Xinyang Han, Xiang Wang +8Ultrasound Image Segmentation Implicit Neural Representations
Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China · Centre for Smart Health and School of Nursing, The Hong Kong Polytechnic University, Hong Kong, China · Department of Imaging and Interventional Radiology, The Chinese University of Hong Kong, Hong Kong, China

Aug 18, 2026J/K move · Enter open · S save

Bound-Aware Per-Organ Recall Risk Control for Multi-Organ CT Segmentation under Clinical Domain Shift

Souraj Adhikary, Negar Chabi, Andre Mastmeyer

Department of Engineering Sciences, Jade Hochschule Wilhelmshaven, Germany

Multi-Organ Segmentation Medical Image Segmentation Medical Image Analysis Conformal Risk Control Computed Tomography

Open paper AskPDF arXiv

Distribution-free risk control adds organ-specific recall guarantees to frozen segmentation. We calibrate per-organ thresholds for an AMOS-trained nnU-Net, audit transfer to RAOS, and estimate local re-certification cost using case-level voxel false-negative rate (FNR). The AMOS control passes, but

7/12

organs exceed

α{=}0.10

after transfer; smaller calibration sets can mask exceedances with conservative or vacuous thresholds. Risk-Controlling Prediction Sets (RCPS) give high-probability control of population-mean risk, whereas Conformal Risk Control (CRC) gives weaker expectation control. Both require exchangeability; fixed and global thresholds give no per-organ guarantee. The Waudby--Smith--Ramdas (WSR) betting bound re-certifies six Tier-1 organs with 25 local cases, versus 30--40 for Hoeffding--Bentkus (HB). CRC needs 10--15 but has a heavier individual-case tail. No Tier-2 organ meets our illustrative precision criterion with 25 cases.

Abstract

Medical Image Segmentation

Explore similar work

Bound-Aware Per-Organ Recall Risk Control for Multi-Organ CT Segmentation under Clinical Domain Shift

A Decision-Aware Evaluation Protocol for Uncertainty in Medical Image Segmentation

Risk-Routed Implicit Boundary Refinement for Robust Ultrasound Image Segmentation