Encoding histopathology whole slide images with location-aware graphs for diagnostically relevant regions retrieval

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Computational Biology

Background:

• Content-based histopathological image retrieval (CBHIR) aids pathologists by finding similar regions in large databases.
• Retrieving diagnostically relevant regions from whole slide images (WSIs) is crucial but challenging.
• Existing CBHIR systems often struggle with preserving structural and location information.

Purpose of the Study:

• To develop a novel framework for efficient and accurate region retrieval from WSI databases.
• To enhance CBHIR by incorporating location-aware graphs and deep hashing techniques.
• To improve sensitivity to tissue distribution and scalability for varying region sizes and shapes.

Main Methods:

• A novel framework utilizing location-aware graphs and deep hash techniques for WSI retrieval.
• Graph convolution and self-attention operations to preserve structural and global location information of regions of interest (ROIs).
• Free-curve query definition for pathologists and efficient retrieval via hashing for large-scale databases.

Main Results:

• Achieved mean average precision above 0.667 on an endometrium dataset and 0.869 on the ACDC-LungHP dataset for irregular region retrieval.
• Demonstrated superior performance compared to state-of-the-art methods in histopathological image retrieval.
• Exhibited an average retrieval time of 0.752 ms from a database of 1855 WSIs.

Conclusions:

• The proposed location-aware graph and deep hash framework significantly advances CBHIR for WSIs.
• The method offers improved accuracy, scalability, and efficiency for clinical applications.
• This approach provides a powerful tool for pathologists in analyzing large histopathological image datasets.