Related Experiment Video
Updated: May 5, 2026
11:57
Three-dimensional Super Resolution Microscopy of F-actin Filaments by Interferometric PhotoActivated Localization Microscopy iPALM
Published on: December 1, 2016
10.7K
pFedBCC: Personalizing Federated multi-target domain adaptive segmentation via Bi-pole Collaborative Calibration
Huaqi Zhang1, Pengyu Wang2, Jie Liu3
1Department of Information Management, The National Police University for Criminal Justice, Baoding Hebei, China.
Computer Methods and Programs in Biomedicine
|February 16, 2025
Summary
This study introduces a novel federated learning framework for unsupervised medical image segmentation, achieving superior performance in multi-type immunohistochemistry analysis while preserving data privacy.
Area of Science:
- Artificial Intelligence
- Medical Image Analysis
- Computational Biology
Background:
- Multi-target domain adaptation (MTDA) is crucial for unsupervised segmentation but faces data privacy challenges in medical settings.
- Federated learning (FL) offers a solution for handling private, cross-institutional medical data.
- Existing FL and domain adaptation methods have limitations in unsupervised medical image segmentation.
Purpose of the Study:
- To propose a personalized Federated Bi-pole Collaborative Calibration (pFedBCC) framework for unsupervised multi-type immunohistochemically (IHC) image segmentation.
- To address client-side and server-side drift challenges in federated multi-target domain adaptation (FedMTDA).
- To leverage unannotated private client data and a public source-domain model for robust segmentation.
Main Methods:
- Developed the pFedBCC framework applying FL to MTDA for medical image segmentation.
- Introduced Semantic-affinity-driven Personalized Label Calibration (SPLC) to mitigate client-side prediction drift by aligning features and generating personalized pseudo-labels.
- Implemented Source-knowledge-oriented Consistent Gradient Calibration (SCGC) to reduce server-side aggregation drift using gradient clipping guided by source-domain information.
Main Results:
- pFedBCC achieved state-of-the-art performance on private and public IHC benchmarks, including 88.8% PA on the MT-IHC dataset and 88.4% PA on the LYON19 dataset.
- The framework demonstrated superior performance compared to all existing federated learning and domain adaptation methods.
- Ablation studies confirmed the significant contributions of SPLC and SCGC components to the overall performance.
Conclusions:
- The proposed pFedBCC framework effectively enables privacy-preserving, unsupervised multi-type IHC image segmentation.
- pFedBCC significantly outperforms existing methods in federated and domain adaptation scenarios for medical image analysis.
- The study introduced a new MT-IHC dataset with over 19,000 images across 10 types, facilitating further research in this area.
