WSISum: WSI summarization via dual-level semantic reconstruction

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Imaging

Background:

• Gigapixel whole slide images (WSIs) contain redundant data, causing high computational, storage, and transmission burdens.
• Automatic WSI summarization is crucial for efficient analysis and reducing these overheads.

Purpose of the Study:

• To introduce WSISum, a novel framework for automatic whole slide image summarization.
• To develop a method that extracts a compact, representative subset of patches from WSIs.

Main Methods:

• WSISum employs a dual-level semantic reconstruction approach.
• It integrates low-level patch reconstruction using clustering-based sparse sampling.
• High-level slide reconstruction is achieved through knowledge distillation from WSI foundation models.

Main Results:

• WSISum effectively reduces computational costs associated with WSI processing.
• The framework demonstrates satisfactory performance across various downstream tasks.
• Successful applications include cancer subtyping, biomarker prediction, and metastasis subtyping.

Conclusions:

• WSISum provides an efficient solution for whole slide image summarization.
• The dual-level reconstruction strategy enhances patch selection and representation.
• This approach significantly lowers computational demands while maintaining performance in critical diagnostic tasks.