Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedA3I: Annotation Quality-Aware Aggregation for Federated Medical Image Segmentation against Heterogeneous Annotation Noise
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.
Identifies heterogeneous (Non-IID across pixels and clients) annotation noise in federated medical image segmentation, models it with a contour evolution model, and proposes FedA3I, which aggregates client models with GMM-estimated annotation-quality weights applied layer-wise on top of FedAvg quantity weights.
Problem
Federated medical image segmentation suffers from pixel-wise annotation noise that is Non-IID within clients and heterogeneous across clients, which existing FL and noisy-label methods do not handle.
Approach
A contour evolution model (CEM) generates pixel-dependent annotation noise, and a multi-source noise model assigns client-specific CEM parameters. FedA3I estimates each client's noise via inner/outer boundary-region losses clustered by a two-component GMM, then combines quality-based and FedAvg quantity-based weights layer-wise, favoring quantity in shallow layers and quality in deep layers.
Key results
- On SKIN under N_E noise with the CE loss, FedA3I reaches 78.81% Dice versus 75.76% for the best CE-based baseline (FedProx).
- On BREAST under N_E noise with the CE loss, FedA3I reaches 74.35% Dice versus 70.15% for ADELE.
- Component ablation shows IntraGW and InterGW each add gains over FedAvg, e.g. SKIN N_S with CE rises from 63.85% (FedAvg) to 72.69% (full).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedA3I with CE rather than the strongest CE baseline | FedA3I (CE) · SKIN N_S | Dice coefficient (%): 66.35 ± 4.62 -> 72.69 ± 1.28 | Helped |
| Use FedA3I with CE rather than the strongest CE baseline | FedA3I (CE) · SKIN N_E | Dice coefficient (%): 75.76 ± 0.59 -> 78.81 ± 0.75 | Helped |
| Use FedA3I with CE rather than the strongest CE baseline | FedA3I (CE) · BREAST N_S | Dice coefficient (%): 68.10 ± 1.09 -> 69.03 ± 1.95 | Helped |
| Use FedA3I with CE rather than the strongest CE baseline | FedA3I (CE) · BREAST N_E | Dice coefficient (%): 68.09 ± 1.06 -> 74.35 ± 1.90 | Helped |
| Use FedA3I with DC rather than the strongest DC baseline | FedA3I (DC) · SKIN N_S | Dice coefficient (%): 73.33 ± 1.38 -> 75.92 ± 0.99 | Helped |
| Use FedA3I with DC rather than the strongest DC baseline | FedA3I (DC) · SKIN N_E | Dice coefficient (%): 73.75 ± 0.68 -> 77.05 ± 0.56 | Helped |
| Use FedA3I with DC rather than the strongest DC baseline | FedA3I (DC) · BREAST N_S | Dice coefficient (%): 74.02 ± 0.64 -> 75.28 ± 0.53 | Helped |
| Use FedA3I with DC rather than the strongest DC baseline | FedA3I (DC) · BREAST N_E | Dice coefficient (%): 69.04 ± 1.92 -> 76.68 ± 0.49 | Helped |
| Add IntraGW to FedAvg (ablation) | FedA3I (CE) · SKIN N_S | Dice coefficient (%): 63.85 ± 2.41 -> 67.45 ± 1.55 | Helped |
| Add IntraGW to FedAvg (ablation) | FedA3I (CE) · BREAST N_E | Dice coefficient (%): 68.49 ± 0.67 -> 69.26 ± 1.80 | Helped |
| Add InterGW to FedAvg + IntraGW (ablation) | FedA3I (CE) · SKIN N_S | Dice coefficient (%): 67.45 ± 1.55 -> 72.69 ± 1.28 | Helped |
| Add InterGW to FedAvg + IntraGW (ablation) | FedA3I (DC) · BREAST N_E | Dice coefficient (%): 68.31 ± 0.46 -> 76.68 ± 0.49 | Helped |
Surprising findings
- FedDM drops sharply under generic noise (e.g. 42.44% Dice on SKIN N_S), since its bounding-box containment assumption fails.
- The inter-group balance coefficient r can make InterGW fail when set improperly, showing its optimal value is noise-dependent.
Limitations
- Noise is synthetic, generated by CEM rather than from real annotator errors.
- The method depends on a warm-up model and on the chosen balance coefficient r, which is noise-dependent.
- Evaluated only on binary segmentation with two medical datasets and 50 clients.
Experimental setup
Datasets: ISIC 2017 BUS BUSIS UDIAT
Benchmarks: ISIC 2017 (SKIN) BREAST (BUS, BUSIS, UDIAT)
Metrics: Dice coefficient
Baselines: GCE SCE ELR ADELE RMD NR-Dice FedProx FedMix FedDM FedCorr FedNoRo
Builds on: FedAvg FedCorr FedMix FedDM Dice loss
Closest papers by meaning
- Dual-calibrated Co-training Framework for Personalized Federated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.86)
- FedST: Federated Style Transfer Learning for Non-IID Image Segmentation (AAAI 2024 · similarity 0.85)
- Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.85)
- Federated Modality-Specific Encoders and Multimodal Anchors for Personalized Brain Tumor Segmentation (AAAI 2024 · similarity 0.84)
- Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 2024 · similarity 0.84)
- FedVCK: Non-IID Robust and Communication-Efficient Federated Learning via Valuable Condensed Knowledge for Medical Image Analysis (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.