Assessing Generalisation of Perivascular Space Segmentation Across Heterogeneous MRI Cohorts: The DoRA-PVS Challenge 2026

Dash S, del C. Valdés Hernández M, Pham W, Salah Khlif M, Brodtmann A, Joliot M, Boutinaud P, Nozais V, Ibarra S, Sleiman K, Gimesi A, Hamadache RE, Mouheb K, Lladó X, Cao Y, Coello RD, Mattern H, Neumann K, Arndt P, Pfister M, Dongmo C, Beck C, Mai HT, Fuchs E, Tang S, Tosun D, Besteher B, Rocktäschel T, Reuken PA, Stallmach A, Opel N, Gaser C, Walter M, Dörner M, Behme D, Piechowiak C, Lading Y, Müller P, Braun-Dullaeus R, Meuth SG, van der Thiel MM, Jansen JF, van Heese EM, Laansma MA, Wardlaw JM, Schreiber S, Bernal J (2026)

Publication Type: Conference contribution

Publication year: 2026

Journal

Publisher: Springer Science and Business Media Deutschland GmbH

Book Volume: 16877 LNCS

Pages Range: 391-404

Conference Proceedings Title: Lecture Notes in Computer Science

Event location: London, GBR

ISBN: 9783032353924

DOI: 10.1007/978-3-032-35393-1_28

Abstract

MRI-visible perivascular spaces (PVS) are increasingly recognised as markers of compromised brain health, but their quantitative use requires segmentation methods that remain reliable beyond the datasets on which they were developed. Whether current methods meet this requirement remains unclear. To address this gap, we organised the Domain Randomisation PVS (DoRA-PVS) Challenge, a no-data-shared benchmark designed to evaluate out-of-sample PVS segmentation in two settings: an open method track for previously published methods, and a domain randomisation track for new strategies trained exclusively on synthetic data. The testbed comprised 285 images from 12 cohorts, capturing substantial heterogeneity in scanner vendors, field strengths, sequences, imaging protocols, image quality, and participant characteristics. Four teams competed in the first DoRA-PVS Challenge: two in the open method track, one in the domain-randomisation track, and one in both tracks. Preliminary results show that external generalisation is achievable. Across tracks, most methods achieved AUPRC values above random-classifier performance on unseen data, indicating that they could segment PVS beyond their development cohorts. However, this generalisation was not consistent: performance varied substantially across cohorts and evaluation metrics. These preliminary findings shift the central question from whether PVS segmentation methods can generalise at all to whether they can generalise reliably across heterogeneous data. The DoRA-PVS Challenge therefore establishes a rigorous benchmark for assessing out-of-sample robustness and provides a framework for developing PVS segmentation methods that are more consistent across acquisition protocols, populations, and anatomical regions.

Involved external institutions

Universitat de Barcelona (UB) / University of Barcelona Spain (ES) Concordia University Canada (CA) Instituto Superior Técnico Portugal (PT) Center for Intervention and Research on adaptive and maladaptive brain Circuits underlying mental health (C-I-R-C) Germany (DE) Universitat de Girona Spain (ES) Erasmus University Medical Center (MC) Netherlands (NL) University of Edinburgh United Kingdom (GB) Zentrum für neurowissenschaftliche Forschung / Center for Behavioral Brain Sciences (CBBS) Germany (DE) Universitätsklinikum Magdeburg A.ö.R. Germany (DE) Universitätsklinikum Jena Germany (DE) Charité - Universitätsmedizin Berlin Germany (DE) Universitätsspital Zürich (USZ) Switzerland (CH) Monash University Australia (AU) Université de Bordeaux France (FR) Otto-von-Guericke-Universität Magdeburg Germany (DE) University of California San Francisco (UCSF) United States (USA) (US) Amsterdam University Medical Centers (Amsterdam UMC) / Amsterdam Universitair Medische Centra Netherlands (NL) Universitätsklinikum Münster Germany (DE) Maastricht University Netherlands (NL)

How to cite

APA:

Dash, S., del C. Valdés Hernández, M., Pham, W., Salah Khlif, M., Brodtmann, A., Joliot, M.,... Bernal, J. (2027). Assessing Generalisation of Perivascular Space Segmentation Across Heterogeneous MRI Cohorts: The DoRA-PVS Challenge 2026. In Hao Ni, Daniele Cafolla (Eds.), Lecture Notes in Computer Science (pp. 391-404). London, GBR: Springer Science and Business Media Deutschland GmbH.

MLA:

Dash, Sumeet, et al. "Assessing Generalisation of Perivascular Space Segmentation Across Heterogeneous MRI Cohorts: The DoRA-PVS Challenge 2026." Proceedings of the 3rd International Conference on Artificial Intelligence in Healthcare, AIiH 2026, London, GBR Ed. Hao Ni, Daniele Cafolla, Springer Science and Business Media Deutschland GmbH, 2027. 391-404.

BibTeX: Download