HistoKernel: Whole slide image level Maximum Mean Discrepancy kernels for pan-cancer predictive modelling.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in medicine

Background:

• Whole slide images (WSIs) often lack detailed labels, requiring weakly supervised learning for clinical tasks.
• Existing methods struggle to capture holistic distributional differences within WSIs, limiting pathological modeling.
• Accurate WSI-level scoring is crucial for tasks like cancer classification and survival prediction.

Purpose of the Study:

• Introduce HistoKernel, a novel WSI-level Maximum Mean Discrepancy (MMD) kernel.
• Quantify distributional similarity between WSIs using local feature representations.
• Enable diverse WSI-level applications and provide patch-level explainability.

Main Methods:

• Developed HistoKernel, a WSI-level MMD kernel for distributional similarity.
• Utilized local feature representations for WSI comparison.
• Implemented a perturbation-based method for patch-level explainability.

Main Results:

• HistoKernel matches or exceeds state-of-the-art performance on large pan-cancer datasets.
• Demonstrated effectiveness in WSI retrieval, drug sensitivity regression, mutation classification, and survival analysis.
• Achieved high performance across diverse tasks with large sample sizes.

Conclusions:

• HistoKernel provides a theoretically grounded framework for WSI analysis.
• Pioneers kernel-based methods for various WSI-level predictive tasks.
• Facilitates rapid prototyping and research on large, complex computational pathology datasets.