Research map / Federated learning and privacy / Personalized and heterogeneous FL
Think Twice Before Selection: Federated Evidential Active Learning for Medical Image Analysis with Domain Shifts
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 FEAL, a federated active learning method that uses Dirichlet-based evidential models to calibrate aleatoric uncertainty with epistemic uncertainty and adds diversity relaxation, evaluated on five multi-center medical imaging datasets under domain shift.
Problem
Federated active learning for medical imaging selects informative local data using softmax point-estimate uncertainty, which is overconfident and miscalibrated under domain shift across clients, and cannot separate aleatoric from epistemic uncertainty.
Approach
Place a Dirichlet prior over predictions in both local and global models, compute aleatoric uncertainty (expected entropy) in both, scale it by the global model's epistemic uncertainty (differential entropy of the Dirichlet), and add a local-neighborhood diversity relaxation when selecting samples; training uses an evidential loss with evidence regularization.
Key results
- On Fed-ISIC, FEAL reaches 68.46% BMA in round 5, beating the second-best method CoreSet (E) by 1.62% and exceeding full supervision by 0.84%.
- On Fed-Camelyon, FEAL attains 99.40% of fully supervised accuracy using about 3.43% of training samples annotated, versus 98.93% for KAFAL.
- On Fed-Polyp, FEAL achieves 80.18% Dice in round 5, 1.34% above the second-best method and 2.00% above full supervision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FEAL instead of the strongest listed baseline | Federated active learning · Fed-Polyp R5 | Dice (%): 78.84 -> 80.18 | Helped |
| Use FEAL instead of the strongest listed baseline | Federated active learning · Fed-Prostate R5 | Dice (%): 86.50 -> 87.42 | Helped |
| Use FEAL instead of the strongest listed baseline | Federated active learning · Fed-Fundus R5 | Average Dice (%): 90.25 -> 90.58 | Helped |
| Add global epistemic uncertainty to global and local aleatoric uncertainty (ablation) | FEAL · Fed-ISIC R5 | BMA (%): 66.70±0.28 -> 68.46±0.37 | Helped |
| Add local aleatoric uncertainty to global epistemic and global aleatoric uncertainty (ablation) | FEAL · Fed-ISIC R5 | BMA (%): 66.85±0.84 -> 68.46±0.37 | Helped |
| Add global aleatoric uncertainty to global epistemic and local aleatoric uncertainty (ablation) | FEAL · Fed-ISIC R5 | BMA (%): 65.92±1.95 -> 68.46±0.37 | Helped |
| Add both global and local aleatoric uncertainty to global epistemic uncertainty (ablation) | FEAL · Fed-ISIC R5 | BMA (%): 66.41±1.10 -> 68.46±0.37 | Helped |
| Add global epistemic calibration to global and local aleatoric uncertainty (ablation) | FEAL · Fed-ISIC R5 | BMA (%): 66.70±0.28 -> 68.46±0.37 | Helped |
| Train with evidential loss instead of cross-entropy (ablation) | FEAL · Fed-ISIC R5 | BMA (%): 67.40±0.22 -> 68.46±0.37 | Helped |
| Train with evidential loss instead of Dice loss (ablation) | FEAL · Fed-Polyp R5 | Dice (%): 79.48±0.62 -> 80.18±0.10 | Helped |
| Set trade-off weight λ to 1e−2 (ablation) | FEAL · Fed-ISIC R5 | BMA (%): 65.45±1.06 -> 68.46±0.37 | Helped |
| Set trade-off weight λ to 1e−2 instead of 5e−3 (ablation) | FEAL · Fed-ISIC R5 | BMA (%): 66.48±0.86 -> 68.46±0.37 | Helped |
Surprising findings
- Recent FAL methods LoGo and KAFAL, strong in simulated federated settings, underperform in real multi-center federated scenarios with domain shift.
- Combining global and local models for selection does not always beat using the local model alone unless the global model's uncertainty is properly calibrated.
Limitations
- Evaluated only on five datasets across two classification and three segmentation tasks, with a fixed small number of AL rounds (R=5).
- Relies on hyperparameters (similarity threshold tau, minimum neighbor size n, loss weight lambda) that were tuned on Fed-ISIC and then applied to other datasets.
Experimental setup
Datasets: Fed-ISIC Fed-Camelyon Fed-Polyp Fed-Prostate Fed-Fundus
Benchmarks: Fed-ISIC Fed-Camelyon Fed-Polyp Fed-Prostate Fed-Fundus
Metrics: BMA ACC Dice HD95
Baselines: Random Entropy TOD Gradnorm CoreSet BADGE LoGo KAFAL
Builds on: Dirichlet-based evidential deep learning subjective logic LoGo KAFAL BADGE
Closest papers by meaning
- FedVCK: Non-IID Robust and Communication-Efficient Federated Learning via Valuable Condensed Knowledge for Medical Image Analysis (AAAI 2025 · similarity 0.87)
- Overcoming Heterogeneous Data in Federated Medical Vision-Language Pre-training: A Triple-Embedding Model Selector Approach (AAAI 2025 · similarity 0.85)
- Dual-calibrated Co-training Framework for Personalized Federated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- FedPIA – Permuting and Integrating Adapters Leveraging Wasserstein Barycenters for Finetuning Foundation Models in Multi-Modal Federated Learning (AAAI 2025 · similarity 0.84)
- dFLMoE: Decentralized Federated Learning via Mixture of Experts for Medical Data Analysis (CVPR 2025 · similarity 0.84)
- A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis (NeurIPS 2024 · 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.