LightMedSeg-ISLES: Stroke Lesion Segmentation with 81x Fewer Parameters than nnU-Net
Giorgi Nikvashvili, Hanxue Gu, Jie Bao, Kang Wang, Yang Yang
cs.CV · cs.LG
Quick verdict
Compact 1.26M-param model keeps 97.5% of nnU-Net Dice on ISLES'26 stroke lesions with better lesion F1; solid ablations, but test-set-tuned thresholds temper it.
Summary
This paper presents LightMedSeg-ISLES, a compact 3D segmentation pipeline of about 1.26 million parameters for T1-weighted stroke lesion segmentation in the ISLES'26 setting. Built on the LightMedSeg backbone with GhostConv patch embedding, learned spatial anchors, a local structural prior module, multi-scale skip fusion, and FiLM conditioning on four clinical variables, it is trained on 1045 T1 MRI scans (ATLAS v2, SOOP, ISLES'26) with lesion-balanced patch sampling, Dice/CE/boundary loss, deep supervision, and a 1000-epoch strong-augmentation recipe. On a 146-case held-out cohort, flip test-time augmentation yields 0.618 mean Dice and 0.599 lesion-wise F1, versus 0.634 Dice and 0.544 F1 for a filtered 102.35-million-parameter nnU-Net ResEnc-L (raw nnU-Net reaches 0.646 Dice). The authors thus claim 97.5% of nnU-Net's Dice with 81.4× fewer parameters and 4.7× fewer FLOPs per standardized patch, also exceeding filtered UNETR++ and nnFormer. Main limitations: operating points (thresholds, filtering) were selected on the internal test cohort and protocols differ across baselines; results rest on single runs without confidence intervals (the paired Wilcoxon test gives p=0.095); the contribution of the backbone itself versus the training recipe is not isolated against a capacity-matched vanilla baseline; no code release or hidden challenge-set results are reported.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 17 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 1.26M-parameter LightMedSeg model with four-pass flip TTA reaches 0.618 mean Dice and 0.599 lesion-wise F1 on the 146-case held-out ISLES'26 cohort.
- It retains 97.5% of filtered nnU-Net ResEnc-L's Dice (0.6178 vs 0.6339) with 81.4× fewer parameters while improving lesion-wise F1 by 0.055.
- Its TTA operating point costs 4.7× fewer FLOPs per standardized 128³ patch than nnU-Net (305.9 vs 1434.8 GFLOPs).
- Training recipe alone (1000 epochs + aug-v2) adds 0.0358 Dice at fixed 1.26M capacity (0.5705 → 0.6063), making a single checkpoint competitive with far larger models.
- The residual gap to nnU-Net concentrates in small-to-medium lesions (Q2: 0.588 vs 0.645), while Q4 lesions are nearly tied (0.797–0.814).
Strengths
Controlled nine-variant ablation with honest negative results
Table 1 compares base/boundary-refined/padded-256³ designs at three sizes; the paper reports outright that boundary refinement (LMSBR) fails to improve the base design (e.g., 0.5705 vs 0.5511 Dice at large size), and §3.3 cleanly isolates the recipe effect at fixed 1.26M capacity (0.5705 → 0.5922 at 1000 epochs → 0.6007 with aug-v2 → 0.6063 at convergence).
Transparent reporting of unfavorable comparisons
§3.4 states that raw nnU-Net achieves the highest Dice of all (0.6464 vs 0.6178), that the paired Wilcoxon test is non-significant (p=0.095, 59/146 case wins), and §3.5 localizes the deficit to Q2 lesions (0.588 vs 0.645) instead of claiming global superiority.
Deployment-relevant, standardized cost accounting
Implementation-level pipeline disclosure
Weaknesses
Operating points selected on the test cohort itself
Baseline training regimes not matched
No uncertainty quantification
Headline framing anchored to the weaker nnU-Net variant
No code release and incomplete specification
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets ISLES'26: T1-weighted stroke lesion segmentation spanning 60+ centers and acute/subacute/chronic stages. It proposes LightMedSeg-ISLES, a 1.26M-parameter single-checkpoint pipeline, and claims: (1) 0.618 Dice / 0.599 lesion-wise F1 on a 146-case holdout with four-pass flip TTA; (2) 97.5% of a filtered 102.35M nnU-Net ResEnc-L's Dice with 81.4× fewer parameters and +0.055 lesion-wise F1; (3) 4.7× fewer operating-point FLOPs per standardized 128³ patch (305.9 vs 1434.8 GFLOPs); (4) +0.0358 Dice purely from a longer, stronger training recipe; (5) slight superiority over filtered UNETR++ (0.6166) and nnFormer (0.6142). Claims 1–4 are supported by the internal numbers; claim 5 rests on margins of 0.001–0.004 Dice with no variance estimates.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is LightMedSeg-ISLES?
A compact 3D segmentation model with about 1.26 million parameters for outlining stroke lesions on T1-weighted brain MRI, developed for the ISLES'26 challenge setting. It combines the LightMedSeg backbone with flip test-time augmentation and reaches 0.618 mean Dice and 0.599 lesion-wise F1 on a 146-case internal test set.
How much smaller is LightMedSeg-ISLES than nnU-Net?
The paper's nnU-Net ResEnc-L reference has 102.35 million parameters versus 1.26 million for LightMedSeg — an 81.4× reduction. Including four-pass test-time augmentation on a standardized 128³ patch, it needs 305.9 versus 1434.8 GFLOPs, about 4.7× less compute.
Does LightMedSeg-ISLES beat nnU-Net?
Not on all metrics. On the paper's 146-case cohort it keeps 97.5% of filtered nnU-Net's Dice (0.618 vs 0.634) and scores higher lesion-wise F1 (0.599 vs 0.544), but raw (unfiltered) nnU-Net achieves the best Dice overall (0.646), and the per-case Dice difference is not statistically significant (Wilcoxon p=0.095).
Is code or data available for LightMedSeg-ISLES?
The paper states no code release and does not publish its 146-case holdout split. Training data come from ATLAS v2, SOOP, and ISLES'26, of which ATLAS v2 and SOOP are publicly available, but exact result reproduction is not possible from the paper alone.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Stray non-paper text ('Ambient Trance / Psybient / Ultimae') precedes Table 2 in the extraction; it is an extraction artifact and was disregarded.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.