Cross-Modality Learning for Predicting Immunohistochemistry Biomarkers from Hematoxylin and Eosin-Stained Whole Slide...

Area of Science:

• Computational pathology
• Artificial intelligence in histopathology
• Digital pathology

Background:

• Hematoxylin and eosin (H&E) staining is crucial for cancer diagnosis, but immunohistochemistry (IHC) provides specific protein insights.
• IHC staining is expensive, time-consuming, and requires specialized expertise, limiting its widespread application.

Purpose of the Study:

• To develop a novel deep learning framework, HistoStainAlign, capable of predicting IHC staining patterns directly from H&E whole slide images.
• To overcome the limitations of traditional IHC staining by offering a cost-effective and efficient computational alternative.

Main Methods:

• HistoStainAlign integrates paired H&E and IHC embeddings using a contrastive training strategy.
• The framework captures cross-modal features without requiring patch-level annotations or tissue registration.
• The model was validated on gastrointestinal and lung tissues for P53, PD-L1, and Ki-67 IHC stains.

Main Results:

• HistoStainAlign achieved weighted F1 scores of 0.735 for P53, 0.830 for PD-L1, and 0.723 for Ki-67.
• Embedding analyses confirmed the model's ability to capture meaningful cross-stain relationships.
• The contrastive learning approach demonstrated superior performance compared to a baseline model for stain pattern prediction.

Conclusions:

• HistoStainAlign shows significant potential as a computational prescreening tool in pathology.
• The framework can improve workflow efficiency by identifying cases that would benefit most from IHC staining.
• This approach offers a promising direction for enhancing diagnostic accuracy and treatment planning in oncology.