Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Computational methods enhance pathology workflows but lack clinical integration due to poor interpretability.
• Deep learning models offer high accuracy but often function as 'black boxes', hindering clinical trust and adoption.

Purpose of the Study:

• To develop an interpretable approach for predicting molecular phenotypes from whole-slide histopathology images.
• To leverage human-interpretable image features (HIFs) for enhanced understanding of tumor microenvironment characteristics.
• To enable accurate prediction of clinically relevant molecular data using explainable AI in digital pathology.

Main Methods:

• Trained deep learning models on >1.6 million pathologist annotations across >5700 whole-slide images.
• Developed 607 human-interpretable image features (HIFs) quantifying cell and tissue characteristics at micron-resolution.
• Combined cell and tissue classification outputs to generate comprehensive HIFs across five cancer types.

Main Results:

• HIFs demonstrated correlation with known tumor microenvironment markers.
• Predicted diverse molecular signatures, including immune checkpoint proteins and homologous recombination deficiency (AUROC 0.601-0.864).
• Achieved performance comparable to 'black-box' deep learning methods, highlighting interpretability without sacrificing accuracy.

Conclusions:

• The HIF-based approach provides a quantitative and interpretable method for analyzing tumor microenvironment composition and spatial architecture.
• This interpretable window aids in understanding histopathology images for improved clinical decision-making.
• Offers a viable alternative to 'black-box' models, facilitating clinical integration of computational pathology.