Graph Machine Learning for Disease Lesion Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging technologies lack an effective method to quantitatively estimate patient survival and therapy response for multi-focal diseases, relying heavily on radiologic and oncologic expertise, and struggle to analyze the spatial distribution and relationship of multiple disease lesions due to high input space dimensionality and missing image information between lesions.
Innovation Solution
A computer-implemented method using a graph machine learning model to analyze a graph representation of disease lesions, combining global and local features, which encodes the spatial relationships and characteristics of disease lesions without requiring complete medical images, allowing for improved prediction of disease progression, survival, and therapy response.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If complete medical images covering all tumor lesions are used as model input, then comprehensive disease information is captured, but input space dimensionality becomes enormous leading to poor model performance
Solution Approach 1:
The patent segments the complete medical image into multiple image patches, each containing one or more tumor lesions. These patches are then processed independently through the graph machine learning model. This segmentation reduces the input dimensionality from the entire image to manageable patches while preserving all disease-relevant information, resolving the contradiction between information completeness and computational feasibility.
Solution Approach 2:
The patent extracts only the relevant portions (image patches containing tumor lesions) from the complete medical image for model input. By taking out and processing only these critical regions rather than the entire image, the system maintains comprehensive disease information while significantly reducing input space dimensionality and improving model performance.
2Reliability
If conventional deep learning systems analyze full-body tumor burden, then comprehensive disease assessment is achieved, but missing image information between lesions and enormous input space dimensionality prevent effective analysis
Solution Approach 1:
The patent introduces a graph representation as an intermediary structure that connects extracted image patches containing tumor lesions. The graph captures spatial relationships and anatomical context between lesions, serving as a mediator that reconstructs the missing image information between patches. This allows comprehensive disease assessment without requiring complete continuous image data, resolving the contradiction between assessment accuracy and information completeness.
3Adaptability or versatility
If radiologic and oncologic expertise is used for survival estimation, then clinical judgment is applied, but the quality of estimation depends on individual expertise and lacks standardization
Solution Approach 1:
The patent creates a digital copy of the tumor lesion analysis process through the graph machine learning model. Instead of relying on individual radiologist or oncologist expertise, the system uses a standardized computational model that processes graph representations of tumor lesions. This copying of the analysis function into an automated system maintains the adaptability of clinical judgment while providing consistent, reproducible survival estimates that do not depend on individual human expertise, resolving the contradiction between flexibility and precision.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
A computer-implemented method for providing a clinical information, comprising receiving (105, 107) input data, wherein the input data comprises a graph representation of a plurality of disease lesions (209.1, 209.2, 209.3, 209.4) of a patient, applying (109) a trained function to the input data to generate the clinical information, wherein the trained function is based on a graph machine learning model, providing (111) the clinical information, wherein the clinical information comprises at least one information for the prediction of the disease progression, the survival or therapy response of the patient.