Feature Importance in Nonlinear Embeddings (FINE): Applications in Digital Pathology

Area of Science:

• Digital Pathology
• Computational Biology
• Biostatistics

Background:

• Quantitative histomorphometry (QH) computationally models disease in digital pathology images using numerous features.
• High-dimensional feature spaces pose challenges for building robust and interpretable disease classifiers.
• Dimensionality reduction (DR) is often used, but quantifying original feature contributions post-DR is difficult.

Purpose of the Study:

• To extend a feature scoring method to nonlinear dimensionality reduction (NLDR) techniques.
• To introduce the Feature Importance in Nonlinear Embeddings (FINE) method for improved classifier interpretability.
• To identify key QH features for predicting breast and prostate cancer recurrence risk.

Main Methods:

• Extension of a principal components analysis (PCA)-based feature scoring method to kernel PCA (KPCA).
• Application of the FINE method to four digital pathology datasets.
• Comparison of FINE with t-test, Fisher score, and Gini index for feature selection.

Main Results:

• FINE was applied to identify key QH features for predicting breast and prostate cancer recurrence.
• Measures of nuclear and glandular architecture and clusteredness were identified as important predictors.
• FINE identified a stable set of features yielding good classification accuracy across datasets.

Conclusions:

• The FINE method provides enhanced interpretability for feature importance in nonlinear embeddings.
• Key quantitative histomorphometry features related to architecture and clusteredness are crucial for predicting cancer recurrence.
• FINE demonstrates superior performance in identifying stable, predictive features compared to traditional methods.