SAUF-Net: Structure--Appearance Representation Learning with Uncertainty Feedback for Semi-Supervised Medical Image Segmentation
Qin Lu, Zheyang Jing, Yujie Yang, Jianwang Li, Chen Yi, Shaofeng Jiang
cs.CV · cs.AI
Quick verdict
SSL segmentation net with structure–appearance decomposition and uncertainty feedback; gains over AdaptFRCNet modest, evaluation limited to 2D binary datasets.
Summary
SAUF-Net is a semi-supervised framework for 2D medical image segmentation that regularizes learning at the feature level rather than only at the prediction level. A Structure–Appearance Decomposition Module splits the SegFormer-B4 bottleneck into structural and appearance representations, which a Disentangled Guidance Module injects into decoding via serial deformable-convolution guidance. An Appearance-Swapped Consistency branch replaces a sample's appearance features with a shuffled batchmate's and demands the fused prediction still match the pseudo-label, while a dual-head discriminator (validity plus uncertainty), trained against Gaussian feature generators, provides feature-level uncertainty feedback weighted by a reliability map computed from the main and auxiliary predictions. On ISIC-2016 the method reports 91.88% Dice at 10% labels and 92.64% at 20%; on Kvasir-SEG, 90.53% and 91.98%, exceeding AdaptFRCNet and most of eleven other SSL baselines; per-image gains over AdaptFRCNet are reported significant (Wilcoxon, p<0.01). The evidence is thin in places: the gain at 20% on ISIC is only +0.11 Dice, no standard deviations or multiple-comparison handling are given, the ablation covers one dataset and one ratio, and the reliability map's second factor rewards branch disagreement — an unexplained design choice. No code release is stated.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 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
- Decomposing bottleneck features into structural and appearance representations (SADM) plus DGM decoding guidance lifts the baseline SegFormer-B4 from 82.16 to 89.91 Dice on Kvasir-SEG at 20% labels
- The proposed reliability map Mrel yields better pseudo-label supervision than conventional confidence thresholding (91.16 vs 89.91 Dice, Kvasir 20%)
- Appearance-swapped consistency and feature-level uncertainty feedback add further gains, and per-image Dice improvements over AdaptFRCNet are statistically significant (two-sided Wilcoxon, p<0.01) on both datasets at all evaluated ratios
- SAUF-Net outperforms state-of-the-art semi-supervised methods, especially in low-label settings, on ISIC-2016 and Kvasir-SEG
Strengths
Genuinely feature-level angle on a prediction-level-dominated field
The combination of SADM decomposition, DGM deformable guidance, appearance-swapped consistency, and dual-head discriminator uncertainty is a coherent attempt to address pseudo-label error accumulation, a real failure mode acknowledged in §1
Broad, current baseline table
Table 1 compares 12 methods including BCP (CVPR'23), ARCO/DAW (NeurIPS'23), and AdaptFRCNet (MedIA 2025) at 10% and 20% ratios on both datasets, avoiding a stale-baseline critique
Component-wise ablation plus hyperparameter sensitivity
Per-image statistical testing
Honest scope statement
Weaknesses
Statistical support is too thin for the margin
Reliability map's second factor is counterintuitive and unvalidated
Ablation and low-label claims rest on one dataset/one ratio
Uncertainty feedback alone degrades Dice
No code release and ambiguous baseline protocol
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 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 SAUF-Net?
It is a semi-supervised framework for 2D medical image segmentation built on a SegFormer-B4 backbone. It decomposes bottleneck features into structural and appearance representations (SADM), injects them into decoding (DGM), enforces consistency when appearance features are swapped across batch samples (ASC), and adds feature-level uncertainty feedback from a dual-head discriminator guided by a reliability map.
What results does SAUF-Net report and against which baselines?
On ISIC-2016 it reports 90.91/91.88/92.64 Dice at 5/10/20% labels; on Kvasir-SEG, 89.71/90.53/91.98 Dice at the same ratios. It outperforms AdaptFRCNet by +0.51 Dice (ISIC, 10%), +1.18 (Kvasir, 10%), +0.11 (ISIC, 20%), and +1.27 (Kvasir, 20%), with per-image improvements reported significant (Wilcoxon, p<0.01).
Is SAUF-Net's code available and can it be reproduced?
The paper states no code or repository release. It uses only public datasets (ISIC-2016: 900 train/379 test; Kvasir-SEG: 1000 images, 80/20 split), and gives most hyperparameters, but several reproduction details (ramp-up schedule, feature-generator architecture, wavelet transform) are absent, so independent replication requires reimplementation.
What datasets and tasks does SAUF-Net cover?
Only two 2D binary RGB segmentation benchmarks: ISIC-2016 (skin lesion dermoscopy) and Kvasir-SEG (polyp endoscopy). No 3D volumes, no multi-class organs, and no cross-dataset or cross-modality (CT/MRI) evaluation; the authors state this scope limit in the conclusion.
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.
- Table 2(a) component checkmarks are garbled in the PDF extraction (columns run together); the ablation row ordering above was reconstructed from the §3.4 narrative, which quotes the same numbers
- The metadata abstract is cut mid-sentence ('outperforms sta'), but the full-text abstract is complete, so no truncation affects this review
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.