EvoPS: Evolutionary Patch Selection in the Training Embedding Space of Whole Slide Images

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in healthcare

Background:

• Whole-Slide Images (WSIs) in computational pathology are gigapixel-sized, requiring decomposition into numerous patches.
• High-dimensional patch embeddings are computationally expensive and often contain uninformative regions.
• Current patch selection methods lack explicit optimization for representation compactness and diagnostic accuracy.

Purpose of the Study:

• To develop a novel framework, Evolutionary Patch Selection (EvoPS), for efficient and accurate patch selection in WSIs.
• To address the trade-off between the number of selected patches and downstream task performance.
• To generate a Pareto front of optimal solutions balancing computational cost and diagnostic accuracy.

Main Methods:

• Formulating patch selection as a multi-objective optimization problem within the training embedding space.
• Employing an evolutionary search to simultaneously minimize selected patch embeddings and maximize similarity search performance.
• Validating the EvoPS framework across four major cancer cohorts from The Cancer Genome Atlas (TCGA) using five histopathology foundation models.

Main Results:

• EvoPS significantly reduces the number of required training patches by over 90%.
• The framework consistently maintains or improves classification F1-score compared to state-of-the-art methods.
• EvoPS produces higher-quality training representations, reducing memory requirements and improving signal-to-noise ratio.

Conclusions:

• EvoPS offers a robust and principled method for creating efficient, accurate, and interpretable WSI representations.
• The framework enables users to optimize the balance between computational cost and diagnostic performance.
• EvoPS enhances the utility of computational pathology by improving WSI analysis efficiency and accuracy.