Related Experiment Video
Updated: Jun 26, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
392
Federated 3D multi-organ segmentation with partially labeled and unlabeled data
Zhou Zheng1, Yuichiro Hayashi2, Masahiro Oda2,3
1Graduate School of Informatics, Nagoya University, Furo-cho, Chikusa-ku, Nagoya, Aichi, Japan. zzheng@mori.m.is.nagoya-u.ac.jp.
Summary
This study introduces a new framework for multi-organ segmentation using federated learning, addressing challenges with privacy and limited labels in medical imaging datasets. The method effectively trains generalizable models from imperfect, distributed data.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Collecting large-scale, fully annotated medical datasets is hindered by privacy concerns and high annotation costs.
- Existing datasets often suffer from inconsistent or partial annotations across different institutions.
- Learning from distributed, partially labeled, and unlabeled medical data presents a significant unexplored challenge.
Purpose of the Study:
- To address the novel problem of multi-organ segmentation using distributed, imperfect medical datasets.
- To develop a generalizable model capable of learning from privacy-restricted, partially labeled, and unlabeled data.
- To unify federated learning (FL), partially supervised learning (PSL), and semi-supervised learning (SSL) for robust medical image analysis.
Main Methods:
- A federated learning framework was designed with distributed clients and a central server to learn a global model.
- The framework integrates specialized modules for partially supervised learning (PSL), semi-supervised learning (SSL), and federated learning (FL).
- The PSL module handles partially labeled samples, SSL extracts information from unlabeled data, and FL aggregates local client models.
Main Results:
- The proposed method achieved an average Dice score of 84.83% and 95HD of 41.62 mm for multi-organ segmentation (liver, spleen, stomach) on abdominal CT datasets.
- The model demonstrated strong generalization capabilities through successful transfer learning applications.
- Extensive evaluations confirmed the framework's effectiveness in leveraging distributed, imperfect datasets.
Conclusions:
- A novel approach for multi-organ segmentation was presented, tackling the challenge of distributed, partially labeled, and unlabeled medical imaging data.
- The unified framework effectively combines PSL, SSL, and FL to train generalizable models from imperfect datasets.
- The validated framework shows significant potential for advancing medical image segmentation in real-world scenarios.
