Twelve cases of the VALDO 2021 challenge, Task 1 (T1, T2 and FLAIR; enlarged perivascular spaces annotated by two raters). Every case is predicted by a classifier that never saw it (leave-one-case-out). Research and teaching use only: not a medical device, not for clinical decisions.
| Step | What happens |
|---|---|
| Detector A | DRIPS 3D U-Net on T2, posterior ≥ 0.10 (sensitive) |
| Detector B | MedNet-PVS on T1, posterior ≥ 0.50 (precise) |
| Candidate pool | 26-connected components of A ∪ B |
| Second stage | 29 features per component (size, shape, membership, posteriors, z-scored intensities inside and in a 1-voxel shell, SynthSeg tissue groups); HistGradientBoosting, threshold 0.4 |
| Metric | component-level F1 (a predicted component counts when it touches a reference component), bootstrap over cases |
Pooled over the 12 cases against rater 1: 903 detected, 335 false alarms, 878 missed (pooled F1 0.60, mean over cases 0.57). The best single detector reaches 0.44; the ceiling set by the two raters against each other is F1 0.29 on the six cases both annotated.
The images and reference annotations are derived from the public training data of the challenge:
Sudre, Carole Hélène; Van Wijnen, Kimberlin; Dubost, Florian; de Groot, Marius; de Bruijne, Marleen (2021). Training dataset for the VALDO 2021 challenge – Vascular Lesion Detection, Task 1. Zenodo. DOI: 10.5281/zenodo.4687995
Challenge paper: Sudre CH et al., "Where is VALDO? VAscular Lesions Detection and segmentatiOn challenge at MICCAI 2021", Medical Image Analysis 2024 (arXiv 2208.07167). Task 1 data come from the SABRE study and the Rotterdam Scan Study.
Funding acknowledged as requested by the data providers: Wellcome Trust (082464/Z/07/Z), British Heart Foundation (SP/07/001/23603, PG/08/103, PG/12/29/29497 and CS/13/1/30327), Erasmus MC University Medical Center, the Erasmus University Rotterdam, the Netherlands Organization for Scientific Research (NWO) Grant 918-46-615, the Netherlands Organization for Health Research and Development (ZonMW), the Research Institute for Disease in the Elderly (RIDE), and the European Union Seventh Framework Programme (FP7/2007–2013) under grant agreement No. 601055, VPHDARE@IT, the Dutch Technology Foundation STW.
Cohort references: Tillin T, et al. Southall And Brent REvisited: Cohort profile of SABRE. Int J Epidemiol 2012, doi:10.1093/ije/dyq175 · Jones S, et al. Cohort profile update: SABRE. Int J Epidemiol 2020, doi:10.1093/ije/dyaa135 · Ikram MA, et al. The Rotterdam Scan Study: design update 2016 and main findings. Eur J Epidemiol 2015, doi:10.1007/s10654-015-0105-7.
Licence: the source data are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License. The slice images, outlines and these pages are adaptations of that material and are shared under the same licence, CC BY-NC 4.0 (creativecommons.org/licenses/by-nc/4.0): non-commercial use only, give credit as above. Changes made: conversion of axial T2 slices to 8-bit images with a 1–98 % window, drawing of component outlines.
Detectors used as published: DRIPS (Bernal et al.) and MedNet-PVS; tissue groups from SynthSeg (Billot et al., Med Image Anal 2023).