SPARCL: Spatially Adaptive Regional Coverage for Active Learning in Chest X-ray Analysis

Juho Jung; Seongjae Kang; Mijoo Kim; Sungduk Cho; Yunkyoung Jun; Sunghoon Joo; Yoona Hwang

cs.CVOct 13, 2026

SPARCL: Spatially Adaptive Regional Coverage for Active Learning in Chest X-ray Analysis

Authors: Juho Jung, Seongjae Kang, Mijoo Kim, Sungduk Cho, Yunkyoung Jun, Sunghoon Joo, Yoona Hwang

Organizations: VUNO Inc., Seoul, South Korea

Abstract

Pixel-level annotation for chest X-ray (CXR) lesion segmentation is expensive, making active learning (AL) essential. However, existing AL strategies fail to account for the spatial distribution of selected samples, resulting in inconsistent performance across focal (e.g., aortic calcification) and diffuse (e.g., nodules) pathologies. In this paper, we propose SPARCL (Spatially Adaptive Regional Coverage for Active Learning) to dynamically balance predictive uncertainty with cumulative spatial coverage. By employing a concave coverage function, SPARCL imposes diminishing returns on previously annotated regions, naturally directing the annotation budget toward under-labeled areas. Furthermore, an entropy gate reverts the algorithm to standard uncertainty sampling when spatial variation is absent. This coverage objective is monotone submodular, guaranteeing near-optimal greedy batch selection. To address how aggregate metrics like Dice obscure regional failures, we further introduce two metrics---Spatial Inconsistency (SI) and Performance Coverage (PC)---to quantify anatomical fairness across AL rounds. Evaluated on an in-house calcification cohort and public CXR benchmarks, SPARCL reaches 95% of fully supervised performance using only 24% of the data, compared to 29% for the best baseline. Ultimately, SPARCL ensures robust spatial reliability, surfacing regional gains that aggregate metrics miss.

Figures & tables

Fig. 1: Overall framework of SPARCL . Step 1: Each unlabeled image $x\in\mathcal{U}_{t-1}$ yields an uncertainty score $U(x)$ and a spatial distribution $q(x)$ of its pixel entropy. Over the top- $rB$ candidates, the entropy $\mathcal{H}(\bar{q})$ of their pooled distribution sets the positional-diversity gate $\lambda^{(t)}$ , lowering the spatial weight when the candidates’ uncertainty concentrates in the same few bins. Step 2: Rather than starting from zero, the coverage state $H(\mathcal{S})$ is initialized to $H_{0}$ , the ground-truth lesion distribution $g(x)$ of the labeled set. Greedy marginal-gain maximization then adds one image at a time, updating $H(\mathcal{S})$ after each pick, so that it rewards uncertainty and bins underrepresented by both the labeled set and the current batch. Step 3: The selected samples are annotated, merged into $\mathcal{L}_{t}$ , and used to retrain the model.

Fig. 2: Active learning performance across annotation budgets (10–100%) . SPARCL reaches the highest Dice(

\uparrow

) and lowest SI(

\downarrow

) at every budget up to 50% and the highest PC(

\uparrow

) at

\leq 20\%

. PC is a per-round gain and decays for all methods as learning saturates.

	In-house Calcification			SIIM-ACR
Method	Dice@20% $\uparrow$	AULC $\uparrow$	Budget@95% $\downarrow$	Dice@20% $\uparrow$	AULC $\uparrow$	Budget@95% $\downarrow$
Random	37.3 $\pm$ 1.2	42.7 $\pm$ 1.3	29.7 $\pm$ 3.3%	52.9 $\pm$ 2.2	56.9 $\pm$ 1.5	25.5 $\pm$ 3.2%
Area-Random	38.3 $\pm$ 0.4	43.0 $\pm$ 1.2	29.9 $\pm$ 1.6%	54.1 $\pm$ 2.1	59.3 $\pm$ 2.3	24.7 $\pm$ 2.1%
Uncertainty	41.2 $\pm$ 1.1	43.7 $\pm$ 0.5	29.3 $\pm$ 0.9%	53.6 $\pm$ 1.8	58.9 $\pm$ 1.6	22.4 $\pm$ 1.1%
Area-Uncertainty	41.5 $\pm$ 0.2	43.4 $\pm$ 0.8	29.5 $\pm$ 0.7%	53.8 $\pm$ 0.5	58.1 $\pm$ 2.2	24.2 $\pm$ 1.6%
Coreset	40.9 $\pm$ 2.2	43.4 $\pm$ 0.9	29.9 $\pm$ 2.2%	52.7 $\pm$ 1.0	57.5 $\pm$ 0.7	31.3 $\pm$ 2.8%

TABLE I: Label efficiency on segmentation benchmarks, mean

\pm

std over five shared random seeds. Dice@20% is the per-image Dice (

\times

100) at a 20% annotation budget, Random held-out split; AULC is the normalized area under the Dice-vs-budget curve. Budget@95% is the minimum budget to reach 95% of full-data Dice (in-house 46.8

\pm

1.7, SIIM-ACR 61.8

\pm

0.4 at 100% budget). Bold = best in column, Underline = second best in column.

Method	Global Dice $\uparrow$	Mean Bin Dice $\uparrow$	Dead-Bin % $\downarrow$	Anat-Zone Worst $\uparrow$	SI $\downarrow$	PC $\uparrow$
Random	28.8 $\pm$ 3.1	22.0 $\pm$ 2.1	23.0 $\pm$ 3.3	11.3 $\pm$ 3.3	0.931 $\pm$ 0.181	0.138 $\pm$ 0.039
Area-Random	29.2 $\pm$ 2.8	27.3 $\pm$ 3.1	17.6 $\pm$ 2.0	10.6 $\pm$ 2.7	0.847 $\pm$ 0.109	0.137 $\pm$ 0.042
Uncertainty	39.8 $\pm$ 1.1	31.0 $\pm$ 2.0	16.0 $\pm$ 1.2	16.6 $\pm$ 2.9	0.750 $\pm$ 0.045	0.111 $\pm$ 0.035
Area-Uncertainty	38.1 $\pm$ 1.2	33.7 $\pm$ 2.4	12.0 $\pm$ 1.5	16.4 $\pm$ 3.3	0.879 $\pm$ 0.119	0.137 $\pm$ 0.022
Coreset	36.9 $\pm$ 1.3	29.4 $\pm$ 2.3	16.1 $\pm$ 2.5	14.5 $\pm$ 2.9	0.751 $\pm$ 0.023	0.098 $\pm$ 0.009
Diversity	31.2 $\pm$ 3.1	28.7 $\pm$ 3.4	18.8 $\pm$ 3.2	15.7 $\pm$ 4.6	0.655 $\pm$ 0.071	0.133 $\pm$ 0.042

TABLE II: Spatial reliability at a fixed 20% annotation budget on the in-house calcification dataset. Global Dice is the pooled micro-Dice (

\times

100) over the test set; Mean Bin Dice is the mean pooled Dice over the

7{\times}7

grid bins, Random held-out split; Anat-Zone Worst is the lowest pooled Dice over the seven lung zones (Fig. 7 ); Dead-Bin% is the share of GT-bearing

7{\times}7

bins with zero pooled overlap. SI is the Spatial Inconsistency of Eq. ( 6 ) (

\downarrow

) and PC the Performance Coverage (

\uparrow

). Mean

\pm

std over five shared seeds; ties are broken by the lower std. Bold = best in column, Underline = second best.

Fig. 3: Acquisition behavior and its spatial consequences (in-house; rows: Uncertainty vs. SPARCL ; columns: rounds 5/20/30). (a) Queries over the pool lesion density ( $\bullet$ : earlier rounds; $\star$ : this round) beside the $7{\times}7$ per-bin test Dice grid (red $\rightarrow$ green as Dice $0\rightarrow 1$ ). The upper-lung-zone bins (2nd–3rd rows), where pool lesions are densest, remain “blind spots” under Uncertainty (Dice $0.76$ – $0.86$ at round 30), whereas SPARCL raises them to $0.89$ – $0.93$ . (b) The red dashed circle marks a lesion-free region in the right upper lung zone of a test case: at round 30 Uncertainty places a false positive there (Dice $0.739$ , SI $0.338$ ) while SPARCL leaves it empty (Dice $0.956$ , SI $0.066$ ).

Fig. 4: Qualitative detection comparison on VinDr-CXR. Ground-truth boxes ( green ) and each method’s predictions ( red ; confidence $>0.5$ , top-5 shown). On the same cases, SPARCL localises the dispersed abnormalities more tightly and produces fewer spurious boxes than other baselines.

Fig. 5: Modality generalization to color fundus (DDR, microaneurysms). Rows: five test cases; columns: acquisition methods. Ground truth in green , predictions in blue , each true-positive overlap boxed in red . SPARCL recovers more of the tiny scattered lesions (more red boxes, higher per-image Dice) than Random , Uncertainty , and Coreset .

	VinDr-CXR					ChestX-Det10
Method	mAP † $\uparrow$	AP 50 † $\uparrow$	FROC † $\uparrow$	AULC $\uparrow$	Budget@95% $\downarrow$	mAP † $\uparrow$	AP 50 † $\uparrow$	FROC † $\uparrow$	AULC $\uparrow$	Budget@95% $\downarrow$
Random	7.1 $\pm$ 0.3	19.9 $\pm$ 0.6	25.0 $\pm$ 0.5	9.3 $\pm$ 0.3	49 $\pm$ 8%	8.5 $\pm$ 0.4	25.8 $\pm$ 0.5	26.5 $\pm$ 0.3	10.5 $\pm$ 0.2	75 $\pm$ 4%
Area-Random	7.3 $\pm$ 0.2	20.7 $\pm$ 0.2	25.7 $\pm$ 0.7	9.6 $\pm$ 0.1	42 $\pm$ 4%	8.2 $\pm$ 0.2	26.1 $\pm$ 0.2	26.4 $\pm$ 0.2	9.5 $\pm$ 0.1	82 $\pm$ 1%
Uncertainty	6.7 $\pm$ 0.4	17.5 $\pm$ 1.1	22.9 $\pm$ 0.5	8.6 $\pm$ 0.1	48 $\pm$ 2%	7.6 $\pm$ 0.3	24.4 $\pm$ 0.6	25.4 $\pm$ 0.5	9.9 $\pm$ 0.1	89 $\pm$ 4%
Area-Uncertainty	7.1 $\pm$ 0.1	20.4 $\pm$ 1.0	24.8 $\pm$ 0.7	9.3 $\pm$ 0.3	47 $\pm$ 6%	7.3 $\pm$ 0.3	23.3 $\pm$ 0.6	24.2 $\pm$ 0.5	10.0 $\pm$ 0.2	80 $\pm$ 2%
Coreset	6.9 $\pm$ 0.3	19.7 $\pm$ 0.7	25.5 $\pm$ 0.9	9.5 $\pm$ 0.3	45 $\pm$ 4%	8.6 $\pm$ 0.2	25.7 $\pm$ 0.3	26.2 $\pm$ 0.2	10.6 $\pm$ 0.1	73 $\pm$ 3%

TABLE III: Label efficiency and native object-detection performance on public detection benchmarks, mean

\pm

std over five shared random seeds. mAP † , AP 50 † , and FROC † are at a fixed

20\%

budget; AULC is the normalized area under the mAP-vs-budget curve; Budget@95% is the minimum budget to reach

95\%

of full-data mAP (VinDr-CXR 9.5

\pm

0.1 mAP / 25.8

\pm

0.2 AP 50 ; ChestX-Det10 13.2

\pm

0.4 mAP / 36.5

\pm

0.3 AP 50 at 100% budget). Bold = best in column, Underline = second best in column.

Method	Dice@20% $\uparrow$	AULC $\uparrow$	Final Dice $\uparrow$
Random	21.0 $\pm$ 0.29	25.9 $\pm$ 0.13	31.3 $\pm$ 0.21
Area-Random	23.4 $\pm$ 0.13	26.3 $\pm$ 0.07	31.5 $\pm$ 0.14
Uncertainty	21.4 $\pm$ 0.12	26.4 $\pm$ 0.05	31.6 $\pm$ 0.10
Coreset	22.9 $\pm$ 0.21	26.9 $\pm$ 0.13	32.2 $\pm$ 0.11
Diversity	15.4 $\pm$ 0.26	24.2 $\pm$ 0.21	29.4 $\pm$ 0.12
TAUDIS	16.5 $\pm$ 0.17	22.0 $\pm$ 0.15	30.5 $\pm$ 0.13

TABLE IV: Modality generalization to color fundus photography (DDR, microaneurysm segmentation, three shared seeds), showing the CXR pipeline transfers to a different imaging modality. Dice@20% and AULC use the annotation-budget fraction; Final Dice is at the last AL round.

Fig. 6: Qualitative segmentation on the in-house calcified-nodule cohort. Ground-truth contours in green , predictions in blue , and true-positive overlaps boxed in red , shown with per-image Dice. Baselines often miss these scattered calcifications (Dice $\approx 0$ ) while SPARCL recovers them.

Configuration	Dice@20% $\uparrow$	Budget@95% $\downarrow$	SI $\downarrow$
(A) Coverage state
SPARCL (cumulative $H_{0}$ , gate, $r{=}5$ )	43.4 $\pm$ 1.3	24.4 $\pm$ 1.4%	0.571 $\pm$ 0.038
reset ( $H_{0}\equiv 0$ ; within-batch coverage only)	41.2 $\pm$ 1.4	26.5 $\pm$ 1.6%	0.592 $\pm$ 0.012
coverage only ( $U(x)$ removed)	37.1 $\pm$ 1.6	28.7 $\pm$ 1.5%	0.624 $\pm$ 0.026
(B) Positional-diversity gate
w/o gate ( $\lambda\equiv\lambda_{\max}$ )	43.1 $\pm$ 1.5	25.1 $\pm$ 1.9%	0.599 $\pm$ 0.034

TABLE V: Ablation of SPARCL components and hyperparameters (in-house calcification,

7\times 7

grid, mean

\pm

std over five shared seeds). (A) Coverage state: cumulative

H_{0}

(Eq. ( 1 )) vs. resetting the state every round, and coverage without the uncertainty term. (B) Gate: removing it (

\lambda\equiv\lambda_{\max}

) and varying the ceiling

\lambda_{\max}

. (C) Uncertainty floor:

r{=}1

is uncertainty sampling (the batch is the top-

B

U(x)

); “full” runs the greedy step over the whole pool.

Fig. 7: Uniform grid vs. registration-free anatomy bins. Rows: two patients; columns: chest X-ray, fixed $7{\times}7$ grid, seven lung-anatomy zones (apex/mid/base per lung + central), and a finer 13-bin within-lung subdivision. Anatomy bins come per image from lung segmentation, so a given bin denotes the same anatomical region across patients despite body habitus, rotation, or view.

Ablation Category		Method	Budget	Dice( $\uparrow$ )	SI( $\downarrow$ )
Split Strategy	Grid Size( $K\times L$ )	Method	@95%( $\downarrow$ )	Dice( $\uparrow$ )	SI( $\downarrow$ )
Random	$M=6$ ( $3\times 2$ )	SPARCL	37%	0.433	0.421
	$M=25$ ( $5\times 5$ )	SPARCL	28%	0.447	0.493
	$M=49$ ( $7\times 7$ )	SPARCL	24%	0.463	0.414
	$M=64$ ( $8\times 8$ )	SPARCL	24%	0.443	0.455
Spatial Equal	$M=49$ ( $7\times 7$ )	Random	34%	0.445	0.511

TABLE VI: Ablation study on evaluation split strategies and grid size

M

using the In-house Calcification dataset. Dice,

SI

PC

, and Mean Bin Dice are measured at the final AL round and averaged over 5 seeds. SPARCL consistently maintains superior spatial consistency and label efficiency across all settings.

Bin scheme	Dice@20% $\uparrow$	SI $\downarrow$	PC $\uparrow$
Uniform grid ( $7{\times}7$ )	43.4 $\pm$ 1.3	0.571 $\pm$ 0.038	0.148 $\pm$ 0.009
Anatomy zones (7 lung zones)	41.8 $\pm$ 3.1	0.630 $\pm$ 0.041	0.147 $\pm$ 0.008
Anatomy zones (13 within-lung bins)	42.4 $\pm$ 2.2	0.597 $\pm$ 0.040	0.139 $\pm$ 0.013

TABLE VII: Bin definition: uniform grid vs. registration-free anatomy zones (in-house,

20\%

budget). Compares the standard

7{\times}7

grid against anatomical regions (Fig. 7 ). The simple grid maintains competitive spatial reliability without requiring an external lung segmenter. Acquisition uses the partition in each row; SI and PC are evaluated on the fixed 7×7 grid for all rows.

Gain $\psi(z)$	Budget@95% $\downarrow$	Dice@20% $\uparrow$	SI $\downarrow$	PC $\uparrow$
$\log(1+z)$	26.1 $\pm$ 2.1%	42.8 $\pm$ 1.9	0.613 $\pm$ 0.080	0.131 $\pm$ 0.043
$\sqrt{z}$ (ours)	24.4 $\pm$ 1.4%	43.4 $\pm$ 1.3	0.571 $\pm$ 0.038	0.148 $\pm$ 0.009
$z^{1/3}$	25.8 $\pm$ 1.5%	42.3 $\pm$ 1.8	0.608 $\pm$ 0.051	0.140 $\pm$ 0.015

TABLE VIII: Concave-gain (

\psi

) ablation (In-house Calcification, Random held-out split,

7\times 7

grid, mean

\pm

std over 5 seeds). Results are stable across concave functions, confirming the robustness of the diminishing-returns coverage reward.

Active Learning

Active learning selects the examples, labels, features, or experiments most useful for improving a model under limited budgets.

See all 296 papers

Unifying Active Learning and Semi-Supervised Learning for Medical Image Segmentation
Jul 27, 2026Bahram Jafrasteh, Cheng Wan, Heejong Kim +2Active Learning Semi-Supervised Medical Image Segmentation
Department of Radiology, Weill Cornell Medicine, New York, NY, USA · Department of Electrical and Computer Engineering, Cornell University, Ithaca, NY, USA
Efficient Chest X-ray Representation Learning via Semantic-Partitioned Contrastive Learning
Mar 7, 2026Wangyu Feng, Shawn Young, Lijian XuMedical Image Representation Learning Self-Supervised Pre-Training
Shenzhen University of Advanced Technology, Shenzhen, China
How Many Labels Are Enough? ALDA: Active Learning Deployment Advisor for Medical Image Classification
Aug 4, 2026Julia Machnio, Mads Nielsen, Mostafa Mehdipour GhaziActive Learning Medical Image Classification
Pioneer Centre for AI, University of Copenhagen, Copenhagen, Denmark

Jul 27, 2026J/K move · Enter open · S save

Unifying Active Learning and Semi-Supervised Learning for Medical Image Segmentation

Bahram Jafrasteh, Cheng Wan, Heejong Kim, Johannes C. Paetzold+1

Department of Radiology, Weill Cornell Medicine, New York, NY, USA · Department of Electrical and Computer Engineering, Cornell University, Ithaca, NY, USA

Active Learning Semi-Supervised Medical Image Segmentation Medical Image Segmentation

Open paper AskPDF arXiv

In practical settings, medical image segmentation models are often developed with limited annotated data rather than fully labeled datasets. Training frequently begins in ultra-low labeled regimes where only a small number of volumes are annotated. In such scenarios, practitioners must simultaneously decide which cases to annotate and how to best use the remaining unlabeled data. Although active learning (AL) and semi-supervised learning (SSL) both target annotation scarcity, they are typically designed and optimized independently, resulting in objective mismatch and unstable training during early-stage "cold start" conditions. We propose RegAL, a unified active semi-supervised framework governed by a shared topology-aware Pareto optimization that couples sample acquisition with unlabeled data utilization. RegAL evaluates images along three complementary axes, voxel-wise uncertainty, feature diversity, and a novel topological consistency metric, to select anatomically informative edge cases for annotation. On the other hand, the same criteria are used to identify geometrically stable atlas candidates for diffeomorphic registration-guided augmentation to train a self-supervised Mean Teacher segmentation network. Across BraTS 2021, dHCP, and ProstateX, RegAL remains stable with few labeled volumes and consistently outperforms state-of-the-art AL, SSL, and active semi-supervised baselines across Dice and boundary-distance (ASD, HD95) metrics under extreme annotation scarcity.

Abstract

Active Learning

Explore similar work

Unifying Active Learning and Semi-Supervised Learning for Medical Image Segmentation

Efficient Chest X-ray Representation Learning via Semantic-Partitioned Contrastive Learning

How Many Labels Are Enough? ALDA: Active Learning Deployment Advisor for Medical Image Classification