High-Order Correlation-Guided Slide-Level Histology Retrieval With Self-Supervised Hashing.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Histopathological Whole Slide Images (WSIs) are vital for cancer diagnosis.
• Effective retrieval of similar WSIs is crucial for case-based diagnosis.
• Current methods often focus on patch-level retrieval, limiting slide-level performance.

Purpose of the Study:

• To develop an unsupervised method for accurate slide-level WSI retrieval.
• To address the limitations of existing methods in capturing slide-level information.
• To enhance the clinical utility of WSI retrieval in cancer diagnosis.

Main Methods:

• Proposed the High-Order Correlation-Guided Self-Supervised Hashing-Encoding Retrieval (HSHR) method.
• Utilized an attention-based hash encoder trained in a self-supervised manner for slide-level representation.
• Constructed a similarity-based hypergraph to explore high-order correlations for retrieval.

Main Results:

• HSHR generates representative, weighted slide-level hash codes.
• The hypergraph-guided retrieval module effectively explores multi-pairwise manifold correlations.
• Achieved state-of-the-art performance on multiple TCGA datasets (>24,000 WSIs).

Conclusions:

• HSHR significantly outperforms existing unsupervised WSI retrieval methods.
• The method demonstrates superior performance across various cancer subtypes.
• HSHR offers a more practical and intuitive approach for clinical WSI retrieval.