SADL: A Size-Adaptive Decision Layer for Chronic Stroke Lesion Segmentation in ISLES 2026

Marek Wodzinski; Gniewosz Drwiega; Wojciech Szymanski

cs.CVOct 9, 2026

SADL: A Size-Adaptive Decision Layer for Chronic Stroke Lesion Segmentation in ISLES 2026

Authors: Marek Wodzinski, Gniewosz Drwiega, Wojciech Szymanski

Organizations: Sano – Centre for Computational Personalised Medicine International Research Foundation, Krakow, Poland · AGH University of Krakow, Krakow, Poland

Abstract

Chronic ischemic stroke lesions are difficult to segment on T1-weighted MRI. Their appearance overlaps with healthy tissue, and their volumes span four orders of magnitude, so voxel-level and lesion level metrics often disagree. We build an eleven-member nnU-Net ensemble that covers five design axes: a baseline dice + top-K loss, heavy augmentation, per-lesion volume loss re-weighting, stochastic weight averaging, and a sampling-and-architecture block that pairs cohort-balanced sampling with a residual-encoder backbone. Weighted soft-map fusion of the members feeds a tuned decision layer, SADL, that estimates lesion volume, applies an adaptive per-lesion-size threshold, prunes components by voxel count and peak probability, and adds a never-empty rescue. All fusion weights and per-bucket parameters are fitted on out-of-fold (OOF) predictions, so tuning never touches the held-out test set. On a site-disjoint 91-case holdout, SADL reaches Dice 0.7322, lesion-F1 0.7461 and PR-AUC 0.8430 when scored online with the challenge's released evaluation code. A paired non-parametric bootstrap against a reference single ensemble member confirms the lesion-F1 gain at p=0.024, and the gain is largest in the small-lesion regime, where the rescue step recovers cases that a fixed threshold would silently discard.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of the pipeline. Eleven nnU-Net members spanning five design axes feed a weighted-mean soft fusion; the fused foreground field enters the SADL decision layer, which estimates the lesion volume, selects a per-bucket threshold, prunes low-confidence connected components, and rescues cases that would otherwise collapse to an empty prediction.

Configuration	Dice $\uparrow$	LF1 $\uparrow$	PR-AUC $\uparrow$	$\Delta$ Vol. mL $\downarrow$	$\Delta$ Count $\downarrow$
Baseline (dice + top-K)	0.7171	0.6746	0.8358	5.15	0.98
ResEnc	0.7118	0.6731	0.8352	5.33	1.02
DA5 (heavy augmentation)	0.7241	0.6743	0.8236	5.36	1.23
BucketWeighted ( $v_{\text{ref}}=3$ mL)	0.7260	0.6920	0.8351	5.35	0.96
BucketWeighted $\times$ SWA	0.7212	0.7018	0.8304	5.32	0.98
Uniform-fusion ensemble (top-11)	0.7219	0.6851	0.8416	5.13	0.98

Table 1: Overall ablation on the 91-case site-disjoint holdout, scored offline with the ISLES 2026 released evaluation code. Higher is better for Dice / lesion-F1 / PR-AUC, lower is better for the two absolute-difference metrics. Best per column in bold .

	Baseline		Bwt-3		Uniform ens.		SADL
Bucket	Dice	LF1	Dice	LF1	Dice	LF1	Dice	LF1
$<0.5$ mL	0.530	0.693	0.548	0.693	0.535	0.593	0.592	0.767
$0.5$ – $5$ mL	0.567	0.614	0.596	0.666	0.575	0.642	0.579	0.682
$5$ – $50$ mL	0.764	0.684	0.759	0.727	0.768	0.699	0.783	0.792
$\geq 50$ mL	0.873	0.731	0.874	0.675	0.875	0.737	0.877	0.756

Table 2: Per-lesion-size breakdown for Dice and lesion-F1. Bucket populations:

<0.5

mL: 5 cases,

0.5

–

5

mL: 29,

5

–

50

mL: 33,

\geq 50

mL: 24. Best per row and metric in bold .

Figure 2: Case at 5.4 mL: SADL produces a tighter, more complete mask than the baseline. GT contour and TP in green, FP in blue, FN in red.

Explore similar work

LightMedSeg-ISLES: Stroke Lesion Segmentation with 81x Fewer Parameters than nnU-Net
Sep 9, 2026Giorgi Nikvashvili, Hanxue Gu, Jie Bao +2Medical Image Segmentation
University of California, Berkeley, Berkeley, CA, USA · University of California, San Francisco, San Francisco, CA, USA
Native-Space 3D CarveMix for Multi-Site T1w Stroke Segmentation
Aug 24, 2026Dexter Wen Jie Teo, Kumaradevan PunithakumarMedical Image Segmentation 3D Medical Image Segmentation
College of Computing and Data Science, Nanyang Technological University, Singapore · Department of Radiology and Diagnostic Imaging, University of Alberta, Edmonton, Canada
Deep Learning Segmentation of Diffusion-Weighted MRI Acute Ischaemic Stroke: A Pragmatic Evaluation Across Three Datasets
Aug 26, 2026Atle Bjørnerud, Till Schellhorn, Thor H. Skattør +4Medical Image Segmentation U-Net
Computational Radiology & Artificial Intelligence Unit, Department of Physics and Computational Radiology, Oslo University Hospital, Oslo, Norway · Center for Lifespan Changes in Brain and Cognition, Department of Psychology, University of Oslo, Oslo, Norway · Division of Radiology and Nuclear Medicine, Oslo University Hospital, Oslo, Norway +5

Sep 9, 2026J/K move · Enter open · S save

LightMedSeg-ISLES: Stroke Lesion Segmentation with 81x Fewer Parameters than nnU-Net

Giorgi Nikvashvili, Hanxue Gu, Jie Bao, Kang Wang+1

University of California, Berkeley, Berkeley, CA, USA · University of California, San Francisco, San Francisco, CA, USA

Medical Image Segmentation

Open paper AskPDF arXiv

Large networks and ensembles often lead medical image segmentation challenges, but their storage and inference demands complicate deployment. We present LightMedSeg-ISLES, a 1.26-million-parameter pipeline for T1-weighted stroke lesion segmentation in ISLES'26. On a 146-case held-out cohort, flip test-time augmentation produces 0.618 mean Dice and 0.599 lesion-wise F1. A 102.35-million-parameter nnU-Net ResEnc-L produces 0.634 Dice and 0.544 lesion-wise F1 after size filtering. LightMedSeg therefore retains 97.5% of nnU-Net's Dice with 81.4

\times

fewer parameters while improving lesion-wise F1 by 0.055. Its four-pass TTA operating point requires 4.7

\times

fewer FLOPs per standardized patch than nnU-Net. It also slightly exceeds filtered UNETR++ and nnFormer. Longer training and stronger augmentation add 0.0358 Dice without increasing capacity, establishing a strong single-checkpoint alternative to much larger models.

Abstract

Explore similar work

LightMedSeg-ISLES: Stroke Lesion Segmentation with 81x Fewer Parameters than nnU-Net

Native-Space 3D CarveMix for Multi-Site T1w Stroke Segmentation

Deep Learning Segmentation of Diffusion-Weighted MRI Acute Ischaemic Stroke: A Pragmatic Evaluation Across Three Datasets