Federated Multi-Organ Segmentation With Inconsistent Labels.

Area of Science:

• Artificial Intelligence
• Medical Image Analysis
• Federated Learning

Background:

• Federated learning (FL) offers privacy-preserving, decentralized machine learning for medical imaging.
• Current FL methods require consistent data labels across all sites, limiting applications where annotations vary.
• Clinical sites often have partial or non-overlapping annotations for specific organs, posing a challenge for unified federated models.

Purpose of the Study:

• To develop a novel federated learning method capable of handling partially labeled medical image datasets.
• To address the clinical need for multi-organ segmentation in federated settings with heterogeneous annotations.
• To improve the performance of federated models trained on datasets with varying organ annotations.

Main Methods:

• Introduced a federated multi-encoding U-Net (Fed-MENU) architecture for decentralized multi-organ segmentation.
• Proposed a multi-encoding U-Net (MENU-Net) with organ-specific encoding sub-networks for feature extraction.
• Implemented an auxiliary generic decoder (AGD) to regularize training, ensuring informative and distinctive features.

Main Results:

• The Fed-MENU method effectively trained a federated model using partially labeled abdominal CT datasets.
• Experiments on six public datasets demonstrated superior performance compared to localized and centralized learning approaches.
• The proposed method successfully handled variations in organ annotations across different clinical sites.

Conclusions:

• Fed-MENU provides a robust solution for federated multi-organ segmentation with partially labeled data.
• The approach overcomes the limitations of label consistency in existing federated learning methods for medical imaging.
• This work enables more practical and scalable federated learning applications in clinical settings with diverse annotation practices.