Related Experiment Video
Updated: Jun 28, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Federated learning with knowledge distillation for multi-organ segmentation with partially labeled datasets.
Soopil Kim1, Heejung Park2, Myeongkyun Kang1
1Department of Robotics and Mechatronics Engineering, Daegu Gyeongbuk Institute of Science and Technology, Republic of Korea; Department of Psychiatry and Behavioral Sciences, Stanford University, CA 94305, USA.
This study introduces a novel federated learning approach with knowledge distillation to improve multi-organ CT segmentation. The method enhances accuracy and efficiency by regularizing local training with global and organ-specific models, overcoming catastrophic forgetting.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep learning for multi-organ CT segmentation requires large, curated datasets, which are often limited by partial labeling and multi-institutional variations.
- Federated learning addresses data privacy but suffers from catastrophic forgetting, leading to unreliable predictions after local model updates.
Purpose of the Study:
- To develop a robust federated learning framework for accurate and efficient multi-organ CT segmentation.
- To mitigate catastrophic forgetting in federated learning using knowledge distillation.
Main Methods:
- Implemented a multi-head U-Net architecture for shared embedding space learning across different organs.
- Employed knowledge distillation to regularize local training with global and pre-trained organ-specific models.
Main Results:
- The proposed method significantly outperformed state-of-the-art techniques in accuracy and inference time.
- Achieved superior performance on multi-institutional abdominal CT datasets for 7 different organs.
Conclusions:
- The novel federated learning approach with knowledge distillation effectively addresses challenges in multi-organ CT segmentation.
- The method offers a promising solution for reliable and efficient medical image analysis in diverse clinical settings.
