Khái niệm cốt lõi
A novel Federated Evidential Active Learning (FEAL) method that leverages both aleatoric and epistemic uncertainties in global and local models to effectively select informative samples for annotation in federated learning scenarios with domain shifts.
Tóm tắt
The content discusses a novel Federated Evidential Active Learning (FEAL) method for medical image analysis in the presence of domain shifts across multiple clients.
Key highlights:
- Federated learning enables collaborative learning across multiple medical institutions, but the expensive cost of data annotation remains a challenge.
- Existing federated active learning (FAL) methods focus on data from the same domain, making them unreliable in realistic medical scenarios with domain shifts.
- FEAL introduces a Dirichlet-based evidential model to capture both aleatoric (data) and epistemic (knowledge) uncertainties in global and local models.
- FEAL employs a Calibrated Evidential Sampling (CES) strategy that leverages the epistemic uncertainty in the global model to calibrate the aleatoric uncertainty in both global and local models, and also maintains data diversity.
- FEAL further introduces evidence regularization in the Evidential Model Learning (EML) scheme for accurate evidence representation and data assessment.
- Extensive experiments on five real multi-center medical image datasets demonstrate the superiority of FEAL over state-of-the-art active learning and FAL methods in the presence of domain shifts.
Thống kê
The energy score distributions across clients exhibit significant domain shifts, as indicated by the extremely low p-values in the cross-client kernel density estimation (KDE).
FEAL achieves a Balanced Multi-class Accuracy (BMA) of 68.46% on the Fed-ISIC dataset, outperforming the second-best method by 1.62% in the fifth active learning round.
On the large-scale Fed-Camelyon dataset, FEAL attains 99.40% of the fully supervised performance using only 3.43% of the total training samples.
For the segmentation tasks, FEAL yields Dice scores of 80.18%, 87.42%, and 90.58% on the Fed-Polyp, Fed-Prostate, and Fed-Fundus datasets, respectively, surpassing the fully supervised performance.
Trích dẫn
"Federated learning facilitates the collaborative learning of a global model across multiple distributed medical institutions without centralizing data."
"Existing methods mainly focus on all local data sampled from the same domain, making them unreliable in realistic medical scenarios with domain shifts among different clients."
"FEAL comprises two key modules, i.e., calibrated evidential sampling (CES) and evidential model learning (EML)."