HIPI: Spatially resolved multiplexed protein expression inferred from H&E WSIs.

Area of Science:

• Oncology
• Computational Pathology
• Biomedical Imaging

Background:

• Solid tumors involve complex interactions within the tumor microenvironment, impacting diagnosis and prognosis.
• Quantifying these interactions typically requires expensive molecular techniques.
• Hematoxylin and eosin (H&E) staining is a cheap, routine method in pathology.

Purpose of the Study:

• To develop a computational method for predicting cell marker expression from H&E images.
• To establish a cost-effective approach for analyzing tumor-immune interactions.
• To leverage routine histopathology for enhanced diagnostic and prognostic capabilities.

Main Methods:

• Paired H&E and Cyclic Immunofluorescence (CyCIF) images from colorectal cancer serial sections were used for model training.
• A deep learning model, HIPI (H&E Image Interpretation and Protein Expression Inference), was developed.
• Model performance was validated on held-out tumor regions and new patient samples.

Main Results:

• HIPI accurately predicted the spatial distribution of key cell markers from H&E images.
• The model demonstrated effectiveness on both internal and external datasets.
• HIPI successfully inferred cell type colocalization using only tissue morphology from H&E images.

Conclusions:

• HIPI enables accurate prediction of protein expression and cell interactions from standard H&E stained tissue images.
• This method provides a cost-effective and scalable approach for tumor analysis.
• HIPI holds significant potential for clinical applications in cancer diagnostics and prognostics.