A pathological image classification method based on dyeing perception intelligent feature modeling
By using multi-source decoupled edge attribute embedding and metadata feature modulation, combined with hierarchical cross-staining interaction and gating feedback mechanisms, the problems of feature confusion and incomplete data in multi-staining pathological images are solved, achieving more accurate and robust pathological image classification.
Patent Information
- Application Number
- CN202610801772.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-04
- Publication Date
- 2026-08-28
AI Technical Summary
Existing graph neural network models cannot effectively distinguish the features of different staining modalities when processing multi-stained pathological images, ignoring the essential differences between modalities and leading to confusion in feature representation. At the same time, the models cannot dynamically adapt to the input dimension when processing incomplete data, resulting in a decline in classification performance.
We adopt a colorimetric-aware intelligent feature modeling method, which dynamically adjusts features to adapt to different colorimetric modalities through multi-source decoupling edge attribute embedding and metadata feature modulation mechanism, and improves the perceptual accuracy of the model through hierarchical cross-colorimetric interaction and gated feedback mechanism. At the same time, we introduce a robust conditional prototyping learning strategy to handle incomplete data.
It achieves accurate classification of multimodal pathological images, improves the stability and classification performance of the model under incomplete data, enhances the ability to discriminate tissue microenvironments, and maximizes the utilization value of incomplete data.