Reversed domain adaptation for nuclei segmentation-based pathological image classification.

Area of Science:

• Digital pathology
• Medical image analysis
• Deep learning

Background:

• Digital pathology offers a new paradigm in medicine, but deep learning models struggle with insufficient annotations and poor generalization.
• Weak generalization limits model performance in domains lacking adequate labeled data.

Purpose of the Study:

• To enhance deep learning model generalization in digital pathology using domain adaptation.
• To improve prediction accuracy on target domain data by leveraging source domain labels and nuclei segmentation.

Main Methods:

• Implemented a reversed unsupervised domain adaptation strategy to generate target-like results in the source domain.
• Integrated nuclei segmentation to provide additional diagnostic information for the classifier.
• Developed a unified framework for joint training of segmentation and classification modules.

Main Results:

• The reversed domain adaptation effectively reduced nuclei segmentation disparities between source and target domains without target labels.
• Significantly improved image classification performance in the target domain.
• Outperformed existing general domain adaptation methods in target domain classification.

Conclusions:

• Reversed unsupervised domain adaptation is a robust strategy for improving digital pathology image classification.
• The proposed method enhances model generalization and reduces reliance on extensive target domain annotations.
• Jointly trained segmentation and classification modules yield superior performance in challenging domains.