Disconnect to Connect: A Data Augmentation Method for Improving Topology Accuracy in Image Segmentation

Juan Miguel Valverde; Maja Østergaard; Adrian Rodriguez-Palomo; Peter Alling Strange Vibe; Nina Kølln Wittig; Henrik Birkedal; Anders Bjorholm Dahl

cs.CVMar 7, 2025

Disconnect to Connect: A Data Augmentation Method for Improving Topology Accuracy in Image Segmentation

Authors: Juan Miguel Valverde, Maja Østergaard, Adrian Rodriguez-Palomo, Peter Alling Strange Vibe, Nina Kølln Wittig, Henrik Birkedal, Anders Bjorholm Dahl

Organizations: DTU Compute, Technical University of Denmark, Denmark · A.I. Virtanen Institute, University of Eastern Finland, Finland · Department of Chemistry and iNANO, Aarhus University, Denmark

Abstract

Accurate segmentation of thin, tubular structures (e.g., blood vessels) is challenging for deep neural networks. These networks classify individual pixels, and even minor misclassifications can break the thin connections within these structures. Existing methods that focus on improving topology accuracy, such as topology loss functions, require very precise training labels with the correct topology, which is difficult to obtain. Low image resolution and contrast can further complicate the annotation by causing these structures to appear disconnected. We present CoLeTra, a data augmentation method that improves topology accuracy, even on datasets with topologically inaccurate training labels. CoLeTra achieves this by artificially breaking the structures in the images while maintaining the original labels, incorporating to the models the prior knowledge that structures can be continuous even if they appear disconnected. Our extensive experiments, involving different architectures, loss functions, and datasets, demonstrate that CoLeTra leads to segmentations topologically more accurate while often improving the Dice coefficient and clDice. We also release a dataset specifically suited for image segmentation methods with a focus on topology accuracy. CoLetra's code can be found at https://jmlipman.github.io/CoLeTra.

Figures & tables

Figure 1: Left: Annotations missing important areas, compromising the structure’s true topology. Right: CoLeTra improves topology accuracy.

Figure 2: Overview of CoLeTra data augmentation.

Figure 3: Comparison of baselines (four datasets, two architectures, six loss functions) with and without CoLeTra. Red : statistically significant improvement. Table with all performance metrics in Appendix E .

Figure 4: Example segmentations on four datasets and the best-performing loss functions in terms of Betti error with DynUNet architecture, with and without CoLeTra. Arrows: topological errors fixed by CoLeTra.

Figure 5: Betti error and number of small (

<

30 pixels) connected components achieved with CoLeTra on the DRIVE dataset with different number of patches (a,b) and patch sizes (c,d). (CC: connected components)

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Example of labels in the Narwhal dataset. Left shows orthogonal central slices from the volume, middle shows the labels, and right shows the two together.

Table 7

Table 8

			$\beta_{0e}$ (global)	$\beta_{1e}$ (global)	$\beta_{2e}$ (global)	$\beta_{0e}$ (local)	$\beta_{1e}$ (local)	$\beta_{2e}$ (local)
DRIVE	CE	baseline	114.2 (11.8)	35.5 (2.4)	—	1.3 (0.1)	0.2 (0.0)	—
		+ CoLeTra	108.6 (10.2)	33.6 (2.9)	—	1.3 (0.1)	0.2 (0.0)	—
	Dice	baseline	101.2 (10.8)	28.5 (3.3)	—	1.2 (0.1)	0.2 (0.0)	—
		+ CoLeTra	93.2 (9.2)	26.2 (3.6)	—	1.1 (0.1)	0.2 (0.0)	—
	RWLoss	baseline	77.0 (8.2)	29.1 (2.8)	—	1.0 (0.1)	0.2 (0.0)	—
		+ CoLeTra	73.8 (6.2)	28.2 (3.1)	—	1.0 (0.1)	0.2 (0.0)	—

Appendix

Table 5: Global and local Betti Errors (DynUNet)

			$\beta_{0e}$ (global)	$\beta_{1e}$ (global)	$\beta_{2e}$ (global)	$\beta_{0e}$ (local)	$\beta_{1e}$ (local)	$\beta_{2e}$ (local)
DRIVE	CE	baseline	121.0 (13.6)	34.9 (3.1)	—	1.4 (0.1)	0.2 (0.0)	—
		+ CoLeTra	118.9 (13.7)	32.2 (2.8)	—	1.4 (0.2)	0.2 (0.0)	—
	Dice	baseline	75.5 (19.4)	29.2 (3.4)	—	1.0 (0.2)	0.2 (0.0)	—
		+ CoLeTra	69.4 (16.7)	27.8 (3.4)	—	0.9 (0.2)	0.2 (0.0)	—
	RWLoss	baseline	58.7 (11.7)	29.9 (3.1)	—	0.8 (0.1)	0.2 (0.0)	—
		+ CoLeTra	56.8 (11.2)	29.9 (3.0)	—	0.8 (0.1)	0.2 (0.0)	—

Appendix

Table 6: Global and local Betti Errors (AttentionUNet)

Panel of Figure 7 — Figure 7: PumonaryVA dataset. Left: label. Center: clDice + CoLeTra . Right: clDice.

			DynUNet			AttentionUNet
			$\beta_{0e}$	Dice	clDice	$\beta_{0e}$	Dice	clDice
DRIVE	CE	baseline	114.2 (11.8)	72.6 (0.4)	70.0 (0.7)	121.0 (13.6)	73.1 (0.5)	70.7 (0.7)
		+ CoLeTra	108.6 (10.2)	72.5 (0.5)	70.1 (0.6)	118.9 (13.7)	73.1 (0.5)	70.9 (0.7)
	Dice	baseline	101.2 (10.8)	73.7 (0.4)	71.8 (0.5)	75.5 (19.4)	74.9 (0.6)	73.4 (0.9)
		+ CoLeTra	93.2 (9.2)	73.7 (0.5)	72.0 (0.6)	69.4 (16.7)	74.7 (0.5)	73.4 (0.6)
	RWLoss	baseline	77.0 (8.2)	74.4 (0.4)	73.0 (0.6)	58.7 (11.7)	75.0 (0.3)	73.8 (0.4)

Appendix

Table 7: Performance comparison between baseline models (trained with extensive data augmentation) and our CoLeTra on the two architectures, three datasets, and six loss functions.

			DynUNet		AttentionUNet
			FPR	Precision	FPR	Precision
DRIVE	CE	baseline	2.1 (0.1)	76.0 (0.8)	2.1 (0.1)	76.1 (0.8)
		+ CoLeTra	2.1 (0.1)	75.8 (0.9)	2.2 (0.1)	75.2 (1.0)
	Dice	baseline	2.3 (0.1)	75.1 (0.6)	2.2 (0.1)	76.0 (0.7)
		+ CoLeTra	2.4 (0.1)	74.2 (0.7)	2.4 (0.1)	75.0 (0.9)
	RWLoss	baseline	2.2 (0.1)	76.0 (0.9)	2.1 (0.1)	76.6 (0.8)

Appendix

Table 8: False positive rate (FPR) and Precision.

			DynUNet			AttentionUNet
			$\beta_{0e}$	Dice	clDice	Betti	Dice	clDice
DRIVE	CE	baseline	[107.6,120.7]	[71.6,73.6]	[68.8,71.3]	[112.8,129.2]	[72.2,73.9]	[69.5,71.9]
		+ CoLeTra	[103.2,114.0]	[71.5,73.5]	[68.9,71.2]	[112.2,125.6]	[72.2,74.0]	[69.7,72.0]
	Dice	baseline	[93.3,109.0]	[72.7,74.6]	[70.5,73.0]	[68.1,83.0]	[74.1,75.6]	[72.2,74.6]
		+ CoLeTra	[86.1,100.3]	[72.7,74.7]	[70.7,73.2]	[62.9,76.0]	[73.8,75.6]	[72.2,74.6]
	RWLoss	baseline	[70.9,83.2]	[73.5,75.4]	[71.7,74.2]	[53.0,64.4]	[74.2,75.8]	[72.6,74.9]

Appendix

Table 9: Confidence intervals (95%)

			DynUNet			AttentionUNet
			Betti	Dice	clDice	Betti	Dice	clDice
DRIVE	CE	baseline	0.239	0.022	-0.005	0.074	-0.024	-0.035
		+ CoLeTra	0.239	0.022	-0.005	0.074	-0.024	-0.035
	Dice	baseline	0.277	-0.008	-0.047	0.226	0.038	-0.014
		+ CoLeTra	0.277	-0.008	-0.047	0.226	0.038	-0.014
	RWLoss	baseline	0.144	0.004	0.002	0.09	0.001	-0.015

Appendix

Table 10: Effect size (Cohen’s d).

Topology-Preserving Image Segmentation

Topology-preserving image segmentation models image regions while maintaining structural properties such as connectivity, branches, and holes.

See all 43 papers

Towards High-Quality Image Segmentation: Improving Topology Accuracy by Penalizing Neighbor Pixels
Mar 19, 2026Juan Miguel Valverde, Dim P. Papadopoulos, Rasmus Larsen +1Topology-Preserving Image Segmentation Semantic Segmentation
Technical University of Denmark
Continuous-tone Simple Points: An $\ell_0$ -Norm of Cyclic Gradient for Topology-Preserving Data-Driven Image Segmentation
Apr 30, 2026Wenxiao Li, Faqiang Wang, Yuping Duan +3Image Segmentation Topology-Preserving Image Segmentation
Laboratory of Mathematics and Complex Systems (Ministry of Education), School of Mathematical Sciences, Beijing Normal University, Beijing, 100875, People’s Republic of China · State Key Laboratory of Remote Sensing Science, Faculty of Geographical Science, Beijing Normal University, Beijing, 100875, People’s Republic of China
FMS $^2$ : Unified Flow Matching for Segmentation and Synthesis of Thin Structures
Mar 14, 2026Babak Asadi, Peiyang Wu, Mani Golparvar-Fard +2Thin-Structure Segmentation Topology-Preserving Image Segmentation
University of Illinois at Urbana-Champaign

Mar 19, 2026J/K move · Enter open · S save

Towards High-Quality Image Segmentation: Improving Topology Accuracy by Penalizing Neighbor Pixels

Juan Miguel Valverde, Dim P. Papadopoulos, Rasmus Larsen, Anders Bjorholm Dahl

Technical University of Denmark

Topology-Preserving Image Segmentation Semantic Segmentation

Open paper AskPDF arXiv

Standard deep learning models for image segmentation cannot guarantee topology accuracy, failing to preserve the correct number of connected components or structures. This, in turn, affects the quality of the segmentations and compromises the reliability of the subsequent quantification analyses. Previous works have proposed to enhance topology accuracy with specialized frameworks, architectures, and loss functions. However, these methods are often cumbersome to integrate into existing training pipelines, they are computationally very expensive, or they are restricted to structures with tubular morphology. We present SCNP, an efficient method that improves topology accuracy by penalizing the logits with their poorest-classified neighbor, forcing the model to improve the prediction at the pixels' neighbors before allowing it to improve the pixels themselves. We show the effectiveness of SCNP across 13 datasets, covering different structure morphologies and image modalities, and integrate it into three frameworks for semantic and instance segmentation. Additionally, we show that SCNP can be integrated into several loss functions, making them improve topology accuracy. Our code can be found at https://jmlipman.github.io/SCNP-SameClassNeighborPenalization.

Abstract

Topology-Preserving Image Segmentation

Explore similar work

Towards High-Quality Image Segmentation: Improving Topology Accuracy by Penalizing Neighbor Pixels

Continuous-tone Simple Points: An ℓ0\ell_0ℓ0​-Norm of Cyclic Gradient for Topology-Preserving Data-Driven Image Segmentation

FMS2^22: Unified Flow Matching for Segmentation and Synthesis of Thin Structures

Continuous-tone Simple Points: An $\ell_0$ -Norm of Cyclic Gradient for Topology-Preserving Data-Driven Image Segmentation

FMS $^2$ : Unified Flow Matching for Segmentation and Synthesis of Thin Structures