UFPS: A unified framework for partially annotated federated segmentation in heterogeneous data distribution

Area of Science:

• Medical Image Analysis
• Machine Learning
• Computer Vision

Background:

• Partially supervised segmentation offers a label-efficient approach for medical datasets with incomplete annotations.
• Real-world deployment faces challenges due to data privacy concerns and heterogeneity, limiting practical application.
• Federated learning enables collaborative model training without sharing raw data, preserving privacy.

Purpose of the Study:

• To introduce federated partially supervised segmentation (FPSS) to overcome privacy and data heterogeneity barriers.
• To propose a unified framework (UFPS) that addresses class heterogeneity and client drift in FPSS.
• To enable accurate segmentation of all classes in partially annotated medical datasets within a privacy-preserving setting.

Main Methods:

• Developed the Unified Federated Partially Labeled Segmentation (UFPS) framework.
• Incorporated Unified Label Learning (ULL) to prevent class collision and ensure comprehensive pixel segmentation.
• Utilized Sparse Unified Sharpness Aware Minimization (sUSAM) for feature space unification and to combat client drift.

Main Results:

• Empirical studies revealed that traditional federated learning and partially supervised methods suffer from class collision when combined.
• The UFPS framework demonstrated superior deconflicting capabilities compared to existing approaches.
• Extensive experiments on real medical datasets validated the enhanced generalization performance of UFPS.

Conclusions:

• The proposed UFPS framework effectively addresses the challenges of class heterogeneity and client drift in federated partially supervised segmentation.
• UFPS provides a robust solution for privacy-preserving medical image segmentation using partially annotated data.
• The framework shows significant improvements in segmentation accuracy and generalization on real-world medical datasets.