Related Experiment Video
Updated: Jul 14, 2025
07:13
Author Spotlight: An Efficient and Robust Software for Automated Fusion of Multiple Preclinical Imaging Modalities
Published on: October 27, 2023
1.2K
Learning with limited target data to detect cells in cross-modality images.
Fuyong Xing1, Xinyi Yang1, Toby C Cornish2
1Department of Biostatistics and Informatics, University of Colorado Anschutz Medical Campus, 13001 E 17th Pl, Aurora, CO 80045, USA.
Medical Image Analysis
|October 6, 2023
Summary
This study introduces a novel unsupervised domain adaptation framework using generative adversarial networks (GANs) to improve cell detection in microscopy images, even with limited target data. The method enhances performance across different imaging modalities without needing extensive labeled target datasets.
Area of Science:
- Medical Image Analysis
- Computational Biology
- Artificial Intelligence
Background:
- Deep neural networks excel at cell quantification in microscopy but struggle with cross-modality data.
- Unsupervised domain adaptation (UDA) using generative adversarial networks (GANs) shows promise for cross-modality medical image quantification.
- Existing GAN-based UDA methods often require substantial target data, which is frequently unavailable in real-world applications.
Purpose of the Study:
- To address the challenge of limited unlabeled target data in unsupervised domain adaptation for cell identification in microscopy images.
- To develop a robust GAN-based UDA framework that enhances cell detection performance across diverse imaging modalities.
- To investigate the effectiveness of task-specific modeling and data augmentation within a unified UDA framework.
Main Methods:
- Enhanced a dual GAN with task-specific modeling to provide additional supervision for generator learning.
- Explored single-directional and bidirectional task-augmented GANs for domain adaptation.
- Introduced a differentiable, stochastic data augmentation module to mitigate discriminator overfitting.
- Investigated source-, target-, and dual-domain data augmentation, combined with joint task and data augmentation.
Main Results:
- The proposed framework significantly improved cell detection performance compared to baseline methods.
- The method achieved performance comparable to or exceeding fully supervised models trained with real target annotations.
- Outperformed recent state-of-the-art UDA approaches across multiple datasets with varying imaging conditions.
Conclusions:
- The developed GAN-based UDA framework effectively handles limited target data for cell detection in microscopy.
- The integration of task-specific modeling and advanced data augmentation enhances robustness and performance in cross-modality imaging.
- This approach offers a practical solution for cell quantification in scenarios where labeled target data is scarce.
