Deep learning features encode interpretable morphologies within histological images.

Area of Science:

• Computational pathology
• Machine learning in histopathology
• Digital pathology image analysis

Background:

• Convolutional neural networks (CNNs) are powerful for classifying phenotypes in whole slide images (WSIs) but lack interpretability.
• Existing interpretability methods often provide post hoc explanations without detailing contributing biological features.

Purpose of the Study:

• To investigate the interpretability of CNN features derived from H&E stained WSIs.
• To explore CNN features as abstract morphological genes ('mones') and their association with biological phenotypes.

Main Methods:

• Analysis of feature weights in the final layer of a transfer-learning-based CNN architecture.
• Empirical study of the properties and associations of CNN-derived features ('mones').
• Correlation analysis between 'mones' and gene expression data using bioinformatics approaches.

Main Results:

• CNN features ('mones') show strong, independent associations with biological phenotypes, acting like morphological genes.
• 'Mones' can be cancer-specific or shared across related cancer types, with robust correlations observed.
• Linear classifiers based on 'mones' achieve high accuracy in classifying tumor types and detecting tumors, and correlate with gene expression.

Conclusions:

• 'Mones' offer a novel, interpretable morphological decomposition of H&E images, analogous to gene expression.
• This approach enables linking histopathological features to molecular phenotypes without relying on classifiers as proxies.
• Findings support the use of 'mones' for understanding cancer biology and improving diagnostic accuracy in digital pathology.