FedLPPA: Learning Personalized Prompt and Aggregation for Federated Weakly-Supervised Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Federated learning (FL) addresses data silos but struggles with multi-center medical data heterogeneity.
• Weakly-supervised learning is crucial for medical image segmentation due to high annotation costs.
• Existing FL paradigms do not adequately support diverse annotation formats across sites.

Purpose of the Study:

• To propose a novel personalized federated learning framework (FedLPPA) for leveraging heterogeneous weak supervision in medical image segmentation.
• To enable uniform handling of diverse annotation formats within a federated learning setting.
• To improve deep model training by mitigating data heterogeneity and annotation costs.

Main Methods:

• FedLPPA utilizes a learnable universal knowledge prompt and personalized data distribution prompts.
• A dual-attention mechanism integrates prompts with sample features for adaptive decoders.
• A dual-decoder strategy and adaptable aggregation method enhance pseudo-label generation and customize task decoders.

Main Results:

• FedLPPA demonstrated superior performance across four distinct medical image segmentation tasks.
• The framework effectively accommodates heterogeneous data distributions and diverse weak supervision formats.
• Performance closely matched fully supervised centralized training, validating its efficacy.

Conclusions:

• FedLPPA offers a robust solution for federated medical image segmentation under weak supervision and data heterogeneity.
• The proposed framework successfully addresses the challenge of diverse annotation formats in FL.
• This approach holds significant promise for reducing annotation costs and improving model generalization in medical AI.