Deep-learning based breast cancer detection for cross-staining histopathology images

Area of Science:

• Pathology
• Computational Biology
• Biomedical Imaging

Background:

• Hematoxylin and eosin (H&E) staining is standard in pathology but lacks specific nuclear and cytoplasmic labeling, hindering AI segmentation.
• Fluorescent staining offers advantages like multiplexing and 3D imaging but differs in color from H&E, complicating cross-analysis with AI.
• Variations in staining technologies impede the consistent application of artificial intelligence (AI) algorithms in digital pathology.

Purpose of the Study:

• To develop a computational workflow for harmonizing H&E and fluorescent staining data for AI analysis.
• To enable AI models trained on H&E images to accurately analyze fluorescently stained tissue samples.
• To improve breast cancer tumor recognition by overcoming staining-related variations.

Main Methods:

• Applied color normalization and nucleus extraction techniques to address staining variations.
• Developed a workflow to adapt an H&E-trained AI segmentation model for fluorescently stained images.
• Utilized the workflow for breast cancer tumor recognition tasks.

Main Results:

• Achieved 89.6% accuracy for H&E-stained images and 80.5% accuracy for fluorescently stained images in recognizing specific tumor features.
• Demonstrated that cross-staining inference using the developed workflow maintained high precision.
• Validated the feasibility of applying existing pathology AI models to different staining techniques.

Conclusions:

• The developed workflow effectively bridges the gap between H&E and fluorescent staining for AI-driven pathological analysis.
• This approach expands the utility of current AI models in digital pathology, enhancing their applicability across diverse staining methods.
• The study provides a foundation for more robust and versatile AI applications in cancer diagnostics.