Research map / Label-efficient and robust learning / Domain adaptation and generalization
AIF-SFDA: Autonomous Information Filter Driven Source-Free Domain Adaptation for Medical Image 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 AIF-SFDA, a source-free domain adaptation method for medical image segmentation that uses a learnable DCT-based frequency filter optimized with an information-bottleneck MI loss and self-supervised pseudo-label and consistency losses.
Problem
Source-free domain adaptation for medical image segmentation under domain shift, where decoupling domain-variant from domain-invariant information is hard without access to source data or labels.
Approach
A learnable frequency (DCT) attention filter that adaptively selects frequency components of target images, with an information-bottleneck mutual-information constraint reducing domain-variant information and pseudo-label plus feature-consistency self-supervision preserving domain-invariant information, using only target data.
Key results
- On retinal vessel segmentation, AIF-SFDA reaches 66.22% DSC on AVRDB versus 65.16% for the best baseline CS-CADA (a UDA method), and 76.68% DSC on STARE versus 75.33% for UPL-SFDA.
- On ultrasound cartilage, AIF-SFDA achieves 69.14% DSC on B (best SFDA baseline TSFCT 67.03%) and 55.16% DSC on C (SFODA 55.14%).
- Ablation on AVRDB shows the full module set reaches 66.22% DSC versus 58.62% for the source-only configuration without MI minimization, PL selection, or consistency.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MI minimization (ablation) | AIF-SFDA · AVRDB | DSC: 58.62 -> 62.34 | Helped |
| add feature consistency to MI minimization (ablation) | AIF-SFDA · AVRDB | DSC: 62.34 -> 65.25 | Helped |
| add confidence-aware pseudo-label selection to MI minimization (ablation) | AIF-SFDA · AVRDB | DSC: 62.34 -> 63.22 | Helped |
| add feature consistency to MI minimization and pseudo-label selection (ablation) | AIF-SFDA · AVRDB | DSC: 63.22 -> 66.22 | Helped |
| use adaptive information filtering rather than fixed frequency filters (ablation) | AIF-SFDA · AVRDB vessel segmentation | DSC | Helped |
| use AIF-SFDA instead of the strongest reported baseline | Retinal vessel segmentation · AVRDB | DSC: 65.16 -> 66.22 | Helped |
| use AIF-SFDA instead of the strongest reported baseline | Retinal vessel segmentation · CHASEDB1 | IoU: 47.48 -> 47.49 | Helped |
| use AIF-SFDA instead of the strongest reported baseline | Retinal vessel segmentation · DRHAGIS | IoU: 55.56 -> 55.16 | Hurt |
| use AIF-SFDA instead of the strongest reported baseline | Retinal vessel segmentation · LES-AV | DSC: 77.56 -> 78.08 | Helped |
| use AIF-SFDA instead of the strongest reported baseline | Retinal vessel segmentation · STARE | IoU: 62.20 -> 63.01 | Helped |
| use AIF-SFDA instead of the strongest reported baseline | Joint cartilage segmentation · B | DSC: 67.03 -> 69.14 | Helped |
| use AIF-SFDA instead of the strongest reported baseline | Joint cartilage segmentation · C | DSC: 55.14 -> 55.16 | Helped |
Surprising findings
- TSFCT and UPL-SFDA show negative adaptation on some target datasets (e.g., TSFCT 47.36% DSC on AVRDB, below the 54.34% source model).
Limitations
- Model selection for the source model uses the test set via early stopping, which may bias reported results.
- Evaluation on only two medical tasks (retinal vessel and ultrasound cartilage) with a private ultrasound dataset.
- Source-model training details and some hyperparameter sensitivity are not fully reported in the main text.
Experimental setup
Datasets: DRIVE AVRDB CHASEDB1 DRHAGIS LES-AV STARE private ultrasound joint cartilage datasets A, B, C
Benchmarks: AVRDB CHASEDB1 DRHAGIS LES-AV STARE ultrasound joint cartilage B and C
Metrics: DSC IoU
Baselines: Rolling-Unet DTMFormer DAMAN CS-CADA MAAL SFODA UPL-SFDA UBNA TSFCT
Builds on: information bottleneck source-free domain adaptation U-Net
Closest papers by meaning
- Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.85)
- Learning Generalized Medical Image Segmentation from Decoupled Feature Queries (AAAI 2024 · similarity 0.85)
- Supportive Negatives Spectral Augmentation for Source-Free Cross-Domain Segmentation (AAAI 2025 · similarity 0.85)
- MAPSeg: Unified Unsupervised Domain Adaptation for Heterogeneous Medical Image Segmentation Based on 3D Masked Autoencoding and Pseudo-Labeling (CVPR 2024 · similarity 0.84)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.84)
- A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · similarity 0.84)
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.