Histopathology image classification based on semantic correlation clustering domain adaptation.

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning

Background:

• Deep learning models excel in histopathology image classification but require extensive labeled data.
• Acquiring and annotating human pathological images (whole slide images - WSIs) is challenging, limiting model performance.
• Histopathology datasets from animal models are more accessible for training and annotation.

Purpose of the Study:

• To develop an unsupervised domain adaptation method for classifying human WSIs using animal model datasets.
• To address the data scarcity and annotation difficulties in human histopathology image analysis.
• To enable accurate classification and recognition of cancerous regions in human WSIs.

Main Methods:

• Proposed an unsupervised domain adaptation method based on semantic correlation clustering.
• Utilized multi-scale fused features, normalized and mapped to a new feature space.
• Employed cosine distance for semantic correlation, aligning domain and class centers.
• Applied multi-granular information for cross-domain knowledge transfer.
• Used probabilistic heatmaps for visualization and annotation of cancerous regions.

Main Results:

• Achieved high classification accuracy for whole slide images (WSIs).
• Demonstrated effective transfer of knowledge from animal model datasets to human WSIs.
• Generated annotations of cancerous regions that closely resemble manual annotations.

Conclusions:

• The proposed unsupervised domain adaptation method shows significant potential for clinical applications in histopathology.
• It effectively overcomes data limitations by leveraging animal model datasets for human WSI analysis.
• The method provides accurate classification and annotation, aiding in cancer diagnosis.