Related Experiment Video
Updated: Jul 1, 2025
10:39
A Label-Free Segmentation Approach for Intravital Imaging of Mammary Tumor Microenvironment
Published on: May 24, 2022
2.3K
Unsupervised domain adaptation for histopathology image segmentation with incomplete labels.
Huihui Zhou1, Yan Wang1, Benyan Zhang2
1Shanghai Key Laboratory of Multidimensional Information Processing, School of Communication and Electronic Engineering, East China Normal University, Shanghai 200241, China.
Computers in Biology and Medicine
|March 1, 2024
Summary
This study introduces a new Stain-Adaptive Segmentation Network with Incomplete Labels (SASN-IL) to improve deep learning segmentation in histopathology images. The method effectively corrects incomplete labels and adapts to stain variations, enhancing segmentation accuracy.
Area of Science:
- Digital pathology
- Computational imaging
- Machine learning in medicine
Background:
- Stain variations in histopathology images present significant challenges for deep learning segmentation.
- Existing unsupervised domain adaptation methods require extensive labeled source data, limiting their applicability.
- A novel scenario of unsupervised domain adaptation with incompletely labeled source data is addressed.
Purpose of the Study:
- To propose a Stain-Adaptive Segmentation Network with Incomplete Labels (SASN-IL) for robust segmentation in histopathology.
- To develop a method that overcomes the limitations of incomplete labels and stain variations.
- To improve the generalization of segmentation models across different staining conditions.
Main Methods:
- A two-stage approach: incomplete label correction and unsupervised domain adaptation.
- Incomplete label correction involves reliable model selection and rectifying false-negative regions.
- Unsupervised domain adaptation utilizes an adaptive stain transformation module tuned by segmentation performance.
Main Results:
- Significant improvements in segmentation accuracy on a gastric cancer dataset.
- A 10.01% increase in Dice coefficient compared to baseline methods.
- Competitive performance against existing state-of-the-art methods.
Conclusions:
- SASN-IL effectively handles incomplete labels and stain variations in histopathology image segmentation.
- The proposed adaptive stain transformation module enhances model generalization.
- This method offers a promising solution for challenging segmentation tasks with limited labeled data.
