Multi-Source Data Integration for Segmentation of Unannotated MRI Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks enhance magnetic resonance imaging (MRI) segmentation for clinical applications.
• Training these models requires extensive annotated data, which is time-consuming and costly.
• Variability in MRI data across patients, scanners, and protocols necessitates domain-specific retraining and expert annotation.

Purpose of the Study:

• To develop a method for unsupervised federated domain adaptation to overcome the need for persistent data annotation in MRI segmentation.
• To enable knowledge transfer from multiple annotated source domains to an unannotated target domain.
• To improve the generalizability and reduce the annotation burden of deep learning models for MRI segmentation.

Main Methods:

• Unsupervised federated domain adaptation using multiple annotated source domains.
• Minimizing pair-wise distribution distances between target and source domains in a latent embedding space.
• Employing an ensemble approach to integrate knowledge from all domains for improved segmentation.

Main Results:

• Demonstrated effectiveness of the proposed method on two experimental datasets.
• Successful transfer of knowledge from annotated to unannotated domains.
• Reduced reliance on manual annotation by expert radiologists for new domains.

Conclusions:

• The developed unsupervised federated domain adaptation method effectively addresses the data annotation challenge in MRI segmentation.
• The approach facilitates knowledge transfer across diverse MRI data domains, enhancing model applicability.
• Publicly available code enables further research and development in automated medical image segmentation.