Methods and apparatus for analyzing pathology patterns of whole-slide images based on graph deep learning

Patent metadata

Field	Value
Publication number	US-12586184-B2
Application number	US-202318179846-A
Country	US
Kind code	B2
Filing date	Mar 7, 2023
Priority date	Mar 8, 2022
Publication date	Mar 24, 2026
Grant date	Mar 24, 2026

Abstract

The present invention relates to a method and apparatus for analyzing pathology patterns of whole-slide images based on graph deep learning, which may include: a whole-slide image (WSI) compression step of compressing WSI into a superpatch graph; a graph neural networks (GNN) analysis step of embedding node features and context features into the superpatch graph through a GNN model and calculating contributions for each node and edge; a biomarker acquisition step of classifying and grouping the superpatch graph according to the contributions for each node, connecting the classified and grouped superpatch graph in units of groups to generate connected graphs, normalizing and clustering features of the connected graphs, and acquiring environmental graph biomarkers for each group; and a diagnostic information extraction step of extracting and providing diagnostic information on the WSI based on the environmental graph biomarker for each group.

First claim

What is claimed is:

1 . A method of analyzing pathology patterns of whole-slide images (WSI) using graph deep learning, the method comprising:
segmenting the WSI into a plurality of small patches and extracting features of each small patch using a pretrained model; constructing a superpatch graph by grouping spatially adjacent patches with similar features into superpatches and representing each superpatch as a node, and generating edges between superpatches based on spatial information; embedding node features and spatial context features of the superpatch graph using a graph neural network (GNN) model, wherein the spatial context features comprise distance and angle between superpatches embedded via a learnable lookup table; generating a diagnostic score for the WSI using the GNN output; computing contribution scores for each node and edge in the superpatch graph using an integrated gradients method or another attribution method for interpretability; clustering nodes with contribution scores to define subgraphs representing environmental graph biomarkers; and providing diagnostic information for the WSI based on the spatial position and graph structure of the environmental graph biomarkers.

2 . The method of claim 1 , wherein generating the superpatch graph comprises:
segmenting the WSI into N small patches; generating the superpatch graph according to a similarity of the small patches, and then using the superpatch as a node and connecting between the superpatches with an edge to generate the superpatch graph; and integrating features of the small patches in units of superpatches to calculate node features for each superpatch, and calculating edge features through spatial information between the superpatches.

3 . The method of claim 1 , wherein the GNN model is configured to:
calculate contributions for each edge between the superpatches based on context features that reflect heterogeneous surrounding environmental conditions; and incorporate spatial information between superpatches to promote location information transfer between the superpatches.

4 . The method of claim 2 or 3 , wherein the spatial information includes a distance and an angle between the superpatches.

5 . The method of claim 1 , wherein embedding the node features and the spatial context features using the GNN model:
reducing a vector dimension of the node features for each superpatch; embedding the node features and the context features in the superpatch graph through the GNN model; extracting the context features through the GNN model and calculating the diagnostic score; and calculating the contributions for each node, wherein a contribution between adjacent superpatches, including spatial information between the superpatches, is calculated to reflect a heterogeneous surrounding environment feature to the context feature.

6 . The method of claim 1 , further comprising:
generating and displaying subgraphs by connecting between the superpatches with the edge using the superpatch as a node and additionally guiding at least one of contributions for each node and contributions for each edge; classifying and grouping the subgraphs according to the contributions for each node, and then connecting the classified and grouped subgraphs in the units of groups to generate the connected graphs, and averaging all the node features in the connected graph to define connected graph features; and normalizing and clustering all the connected graph features within the same group to extract biomarkers of each group.

7 . The method of claim 6 , wherein, in the step of defining the connected graph features, only graphs having a predetermined number of nodes or more among largest connected graphs within the same group are defined as the connected graph.

8 . The method of claim 1 , wherein the diagnostic information is generated by:
detecting and notifying the risk in the WSI based on a connectivity analysis result of the connected graph; or stratifying a patient's risk grade according to a graphical analysis result of the connected graph.

9 . An apparatus for analyzing pathology patterns of whole-slide images (WSI) using graph deep learning, the apparatus comprising:
a WSI compression unit, comprising a processor and memory, configured to segment the WSI into a plurality of small patches, extract patch-level features using a pretrained model, and compress the patches into a superpatch graph by grouping spatially adjacent and feature-similar patches into superpatches, each represented as a node; a graph analysis unit, configured to embed node features and spatial context features of the superpatch graph using a graph neural network (GNN), wherein the context features include distance and angle between superpatches; a contribution analysis unit, configured to compute contribution scores for each node and edge using an integrated gradients method or another attribution method for interpretability; a graph grouping unit, configured to classify and group superpatches based on their contribution scores to identify connected subgraphs representing environmental graph biomarkers; a biomarker acquisition unit, configured to extract environmental graph biomarkers by analyzing spatial topology of high-contribution subgraphs; and a diagnostic information extraction unit, configured to generate and provide a diagnostic score, based on the identified graph biomarkers and their spatial distribution in the WSI.

Assignees

• Seoul Nat Univ R&Db Foundation
• Seoul Nat Univ Hospital

Inventors

• Kwon Sunghoon
• Lee Yongju
• Shin Kyoungseob
• Moon Kyung Chul
• Park Jeong Hwan
• Oh Sohee

Classifications

• A61B8/5223

  for extracting a diagnostic or physiological parameter from medical diagnostic data (for algorithms to analyse biomedical images G06T7/0012) · CPC title

• G16H50/00 Primary

  ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics · CPC title

• G16H30/40

  for processing medical images, e.g. editing · CPC title

• G06V20/695

  Preprocessing, e.g. image segmentation · CPC title

• G06V10/70

  using pattern recognition or machine learning (optical pattern recognition or electronic computations therefor G06V10/88) · CPC title