An image retrieval method and apparatus

By constructing a multi-granularity discrete semantic identifier index architecture and a shared encoder, the problem that existing pathological image retrieval methods cannot meet multi-level semantic queries is solved, and efficient cross-granularity retrieval from cells to whole slides is achieved, improving retrieval accuracy and response speed.

CN122240874APending Publication Date: 2026-06-19NAT HEALTH COMMISSION INST OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NAT HEALTH COMMISSION INST OF SCI & TECH
Filing Date
2026-01-30
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Existing deep learning-based pathological image retrieval methods rely on fixed-granularity feature representations, which makes it difficult to simultaneously meet the multi-level semantic query needs from microscopic cells to macroscopic tissues.

Method used

By constructing a multi-granularity discrete semantic label index architecture, and employing staining normalization, foreground segmentation, and a shared encoder, multi-granularity joint retrieval of cells, image patches, and whole slices is achieved. Features are extracted and aggregated using the shared encoder to form a unified multi-granularity feature space.

Benefits of technology

It enables cross-granularity retrieval from cells to whole slides, improving retrieval accuracy and response speed, enhancing the system's robustness and interpretability, and supporting efficient multi-granularity semantic retrieval without manual annotation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122240874A_ABST
    Figure CN122240874A_ABST
Patent Text Reader

Abstract

This application relates to an image retrieval method and apparatus. By constructing a multi-granularity joint query method for pathological image retrieval, this application overcomes the limitations of existing technologies based on single-granularity feature representations. Through staining normalization and foreground segmentation, the influence of differences in different scanning devices and staining is eliminated, ensuring consistency of cross-center data. In feature extraction, a shared encoder and a multi-granularity feature aggregation mechanism are employed to ensure semantic consistency of features at the cell level, image block level, and whole-slice level, forming a unified multi-granularity feature space. By discretizing features into semantic identifiers and constructing an index architecture, efficient joint retrieval between different granularities is achieved, significantly improving retrieval accuracy and response speed. Furthermore, this method requires no manual annotation, supports cross-granularity retrieval from cells to whole slices, and enhances the robustness and interpretability of the system.
Need to check novelty before this filing date? Find Prior Art