Research map / Label-efficient and robust learning / Domain adaptation and generalization
Supportive Negatives Spectral Augmentation for Source-Free Cross-Domain Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SNSA, which actively selects a few hard negative target fundus images for expert annotation and uses FFT amplitude-mixing spectral augmentation plus batch spectral shrinkage to adapt a source-free segmentation model.
Problem
Source-free domain adaptation for fundus optic disc and cup segmentation suffers from cross-domain shift, causing the adapted model to make wrong predictions on hard instances; annotation budgets are limited.
Approach
Select a few target-domain hard negatives via class-centroid cosine similarity of masked features, have experts annotate them, and augment them with multi-round Fourier amplitude mixing with unlabeled target images, combined with pseudo-label loss and batch spectral shrinkage.
Key results
- On Drishti-GS, average Dice 91.82% and ASSD 6.14 with full method in ablation (vs 90.47 / 7.01 baseline).
- On RIM-ONE-r3 optic disc, Dice 94.76% and ASSD 4.55 vs CBMT 93.36% and 6.20.
- On open compound setting, average Dice 88.98% vs PLPB 87.07% using only 4-5% annotated hard negatives.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SNSA instead of CBMT | source-free fundus segmentation · Drishti-GS optic cup | Dice: 84.33 -> 86.54 | Helped |
| Select hard negatives with SNSA rather than SALAD | source-free sample selection · RIM-ONE-r3 | Avg ASSD: 6.92 -> 5.99 | Helped |
| Add spectral data augmentation (ablation) | SNSA on Drishti-GS · Drishti-GS | Avg Dice: 90.47 -> 91.18 | Helped |
| Add BSS loss (ablation) | SNSA with data augmentation on Drishti-GS · Drishti-GS | Avg Dice: 91.18 -> 91.82 | Helped |
| Add BSS loss (ablation) | SNSA with data augmentation on Drishti-GS · Drishti-GS | Avg ASSD: 6.71 -> 6.14 | Helped |
| Increase spectral augmentation rounds from 1 to 5 (ablation) | SNSA on Drishti-GS · Drishti-GS | Avg Dice | Helped |
| Vary λ; λ = 0.5 performs best (ablation) | SNSA on Drishti-GS · Drishti-GS | Avg Dice | Mixed |
Surprising findings
- Random selection of the same annotation budget performs comparably to the proposed selection on RIM-ONE-r3 (avg Dice 87.22 vs 87.30).
Limitations
- Hard-negative selection relies on source-model pseudo-labels and centroid similarity, which may be unreliable under strong shift.
- Requires expert annotation of a few target images, so it is not fully source-free and target-free of labels.
- Evaluated only on fundus optic disc/cup segmentation with a single source dataset.
Experimental setup
Datasets: REFUGE RIM-ONE-r3 Drishti-GS
Benchmarks: RIM-ONE-r3 Drishti-GS REFUGE val
Metrics: Dice ASSD
Baselines: DPL CBMT SFDA-FSM TT-SFDA TENT AdvEnt SRDA DAE BEAL PLPB SALAD OCDA
Builds on: Batch Spectral Shrinkage DeepLabv3+ DPL SALAD
Closest papers by meaning
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.85)
- AIF-SFDA: Autonomous Information Filter Driven Source-Free Domain Adaptation for Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · similarity 0.85)
- SF(DA)$^2$: Source-free Domain Adaptation Through the Lens of Data Augmentation (ICLR 2024 · similarity 0.84)
- Learning Generalized Medical Image Segmentation from Decoupled Feature Queries (AAAI 2024 · similarity 0.83)
- Stable Neighbor Denoising for Source-free Domain Adaptive Segmentation (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.