Abstract
ISLES 2026 scores five quantities per case, only one of which is Dice. Our nnU-Net baseline reaches a cross-validated Dice of 0.645 against the challenge’s reported inter-rater agreement of $0.76\pm0.14$ but a lesion-wise Detection F1 of only 0.533, so the larger relative gap is in finding lesions rather than outlining them. Removing small predicted components is the standard remedy, and its threshold can be written in voxels or in physical units; both are in use, but we are not aware of a comparison at a matched magnitude. We provide one on a native-space cohort whose voxel volume spans $0.031$–$5.27$\,mm$^3$. Filtering at all, fitted to the challenge ranking rather than to Dice, raises Detection F1 to 0.590 and cuts lesion-count error from 2.08 to 1.85, for 0.001 of Dice and no training; lesion precision rises from 0.448 to 0.614 for a recall cost of 0.014. The choice of units is a second, far smaller effect: $+0.0036$ Detection F1 (95\% CI $[0.0016, 0.0058]$), roughly a twelfth of the first. It nonetheless holds on unseen sites in 99.5\% of 400 site splits and follows a dose-response. Refitting the threshold under nested cross-validation, so it never sees the fold it is scored on, puts the selection optimism at only $+0.0021$. In a thirteen-arm ablation calibrated against a repeated-seed control, only a longer schedule and a larger batch improve the ranked objective beyond that reference, and we report which widely-used ideas did not.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ISLES_052.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=RnGE9a4HCq
BibTex
@InProceedings{KolGov_Voxels_MICCAISAT2026,
author = { Kolli, Govinda AND Dukre, Adinath Madhavrao AND Razzak, Imran},
title = { { Voxels or Millimetres? A Controlled Comparison of Lesion-Size Filtering for Native-Space Stroke Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17255},
month = {pending},
page = {pending}
}