associated_data_Investigating the Effectiveness of clDice Loss for Road Crack Segmentation

Suprosanna Shit; Johannes C. Paetzold; Anjany Sekuboyina; Ivan Ezhov; Alexander Unger

doi:10.5281/zenodo.15696263

associated_data_Investigating the Effectiveness of clDice Loss for Road Crack Segmentation

Suprosanna Shit, Johannes C. Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien P. W. Pluim, Ulrich Bauer, Bjoern Menze

Unknown Journal (2025)

Method-ToolMethod-ToolPDF AvailableGrade Eligible⚠️ Moderate Risk Flags

Overall Assessment

Strong Methodological Quality

Assessment created by PaperScorers Medical AI v0.1.0 on Jan 14, 2026

B

75/100

Key Takeaways

•Introduces clDice metric and soft-clDice loss emphasising connectivity via skeleton-mask intersections.
•Provides theoretical homotopy-equivalence guarantee when clDice=1 for foreground/background.
•Differentiable soft-skeletonisation via pooling enables end-to-end training with modest overhead.
•Benchmarks on 5 datasets show improved topology/graph metrics vs soft-Dice across 2D/3D.

Conclusion

A solid, novel loss with theory and practical gains; transparency is good (open code, public data). Multiplicity and lack of preregistration temper statistical claims.

Quick Actions

Read Full Paper

Quality Dimensions

Bias & Integrity

COI handling, outcome switching, selective reporting

B75

Methods Rigour

Protocol clarity, sampling, controls

B-70

External Validity

Generalisability, setting, population

B78

Transparency & Reproducibility

Data/code/materials availability, preregistration

A-88

Significance & Novelty

Contribution vs prior art; context if known

A90

Statistical Validity

Models, assumptions, multiplicity, CI priority

C-55

Integrity & Transparency

Integrity checks

P-hacking risk?

?

None suspected

Signs of selective hypothesis testing or analysis choices that could inflate false positive rates.

Outcome switching?

?

Review recommended

Evidence that primary outcomes were changed during or after the study, potentially distorting results.

Conflict of interest?

?

None suspected

Authors have financial or other relationships that could bias research findings.

Data integrity issues?

?

None suspected

Concerns about the accuracy, completeness, or authenticity of the reported data.

Open science signals

Research data

?

Open access

Raw data used in the analyses is publicly available or accessible upon request.

Analysis code

?

Open access

Code or scripts used to analyze the data are shared for reproducibility.

Study materials

?

Not applicable

Protocols, questionnaires, and other materials are publicly available.

Premise

Primary Research Question

Define and validate a topology-preserving similarity/loss (clDice/soft-clDice) for tubular structure segmentation.

Hypothesis

Optimising soft-clDice preserves connectivity (up to homotopy) and improves topology/graph metrics vs soft-Dice.

Hypothesis is Falsifiable

✓Yes

Benchmarks across datasets with predefined metrics (clDice, Betti errors, SMD, Opt-J).

PICO Framework

Population—Not specified

Algorithmic study, not patient/animal population.

Intervention—Not specified

No clinical intervention; proposes a loss function.

Comparison—Not specified

Comparators are baseline losses/architectures, not clinical controls.

Outcome✓Yes

Connectivity/topology metrics relevant for tubular-network segmentation tasks.

Literature Positioning

Literature Review Balanced

Well covered

Broad related work on topology-preserving segmentation and priors is cited and contrasted.

Evidence: Sec. 1.1 Related Literature

Research Gap Clearly Stated

Well covered

Explicit Q1–Q2 define need for topology-preserving metric and differentiable loss.

Evidence: Sec. 1, Q1–Q2

Stated Contribution Clear

Well covered

Defines clDice with topological guarantees and soft-clDice for training; benchmarked on 5 datasets.

Evidence: Abstract; Sec. 1.2; Sec. 5

Study Provenance

Peer review unclearGrant funding disclosedCOI not reportedAffiliations listed

Authors & Affiliations

Technical University of Munich; Eindhoven University of Technology.

Funding Statement

GCB and Translatum (TU Munich); TRABIT EU Grant 765148; acknowledgements list individual support.

Conflicts of Interest

Not reported

Peer Review

?arXiv preprint; peer-review status not stated.

Evidence: arXiv:2003.07311v7 (15 Jul 2022)

Methodological Assessment

All design modules are marked as not applicable. Use the accordion below to review their details.

Modules without applicable checklist items (10)

These designs were fully marked as not applicable but remain available for reference.

Abstract

Accurate segmentation of tubular, network-like structures, such as vessels,\nneurons, or roads, is relevant to many fields of research. For such structures,\nthe topology is their most important characteristic; particularly preserving\nconnectedness: in the case of vascular networks, missing a connected vessel\nentirely alters the blood-flow dynamics. We introduce a novel similarity\nmeasure termed centerlineDice (short clDice), which is calculated on the\nintersection of the segmentation masks and their (morphological) skeleta. We\ntheoretically prove that clDice guarantees topology preservation up to homotopy\nequivalence for binary 2D and 3D segmentation. Extending this, we propose a\ncomputationally efficient, differentiable loss function (soft-clDice) for\ntraining arbitrary neural segmentation networks. We benchmark the soft-clDice\nloss on five public datasets, including vessels, roads and neurons (2D and 3D).\nTraining on soft-clDice leads to segmentation with more accurate connectivity\ninformation, higher graph similarity, and better volumetric scores.\n

Quick Actions

Read Full Paper

Study Overview

Study Design

Method Tool

Primary classification used to evaluate this research.

Population

2D/3D images of tubular networks: retina vessels (DRIVE), roads (Massachusetts), neurons (CREMI), synthetic vessels, Vessap brain vasculature.

Brief summary of who or what was studied.

Pre-registration

Not registered

Indicates whether the protocol or trial was registered before data collection.

Primary Outcome

Primary outcome was not pre-specified.

Reviewer Notes

Methods paper proposing clDice loss; evaluated across multiple public imaging datasets, not a hypothesis-driven clinical study.

Publication Details

DOI

10.5281/zenodo.15696263

Published

June 19, 2025

Citations

345 citations

External Resources

DOI Link OpenAlex

Disclaimer: This assessment is generated by AI and should not be the sole basis for clinical or research decisions. Always review the original paper and consult with domain experts.

Suggested Papers

D+

A Blurred Vision of Health: Metabolic Syndrome as a Risk Factor for Glaucoma in a Large Taiwanese Population Study

Sheng Hao Chang et al.•2025

C

The genetic and functional basis of purine nucleotide feedback-resistant phosphoribosylpyrophosphate synthetase superactivity.

Michael A. Becker et al.•1995

No Report

CD4 CTL, a Cytotoxic Subset of CD4+ T Cells, Their Differentiation and Function

Arata Takeuchi et al.•2017

Suggested Papers

From Our Blog

Confidence Intervals: The Most Important Number in the Paper

P-values only tell you 'if'. Confidence Intervals tell you 'how much' and 'how sure'. Why you should ignore the p-value and look at the brackets.

Open Data: If You Hide It, We Don't Trust It

Trust, but verify. If a scientist refuses to share their raw data, what are they hiding?

Composite Endpoints: Mixing Apples and Oranges

Drug companies love to combine death, heart attacks, and chest pain into one score. It hides the truth.