AI Eye Condition Prediction from Bilateral Image Sets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current eye condition detection methods, such as confocal microscopy and tear film osmolarity, are invasive, time-consuming, and lack sufficient sensitivity and reproducibility, especially for mild cases, making it difficult to accurately diagnose conditions like corneal ectasia and graft rejection.
Innovation Solution
A system and method utilizing a neural network-based prediction model that analyzes a set of eye images to generate predictions related to anterior eye segment or corneal conditions, adjusting predictions based on clustering of eye locations and bilateral presence of conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If confocal microscopy is used to diagnose dry eye syndrome, then diagnostic capability is improved, but the procedure becomes time-consuming and invasive
Solution Approach 1:
The patent replaces mechanical/optical diagnostic systems (confocal microscopy, slit-lamp examination) with an AI-based prediction model that processes eye images to detect corneal conditions. This substitution eliminates the need for time-consuming manual procedures while maintaining or improving diagnostic accuracy through automated analysis of corneal thickness, curvature, and other parameters from standard eye images.
Solution Approach 2:
The patent creates a virtual model of the cornea by generating a representation from three-dimensional structural images (such as OCT scans). This digital copy allows for automated analysis of corneal properties without requiring physical contact or specialized microscopy procedures, thereby reducing procedure time while preserving diagnostic capability.
2Ease of operation
If slit-lamp examination is used to detect corneal conditions, then non-invasive detection is achieved, but magnification is limited and subclinical rejection episodes are missed
Solution Approach 1:
The patent transforms the detection approach by changing from direct optical examination parameters to AI-processed parameters derived from three-dimensional structural images. The system extracts multiple features including corneal thickness maps, curvature profiles, and layer-specific measurements, then applies prediction models to detect subtle changes indicative of subclinical rejection episodes, thereby improving detection sensitivity while maintaining non-invasive operation.
3Measurement precision
If endothelial cell count using specular microscopy is used, then corneal condition detection is performed, but reproducibility and sensitivity are insufficient for mild cases
Solution Approach 1:
The patent creates a universal AI-based prediction model that can detect multiple corneal conditions (rejection episodes, Fuchs' dystrophy, edema, ectasia) from the same set of three-dimensional structural images. This multi-functional approach improves reproducibility by applying consistent AI analysis across different conditions and cases, eliminating variability associated with manual microscopy interpretation while maintaining sensitivity for mild cases.
4Ease of operation
If central cornea thickness measurements are used, then corneal assessment is simplified, but the wide range of normal thickness complicates diagnosis of mild cases
Solution Approach 1:
The patent segments the corneal analysis into multiple independent features extracted from three-dimensional structural images, including but not limited to central corneal thickness. By dividing the diagnostic task into separate feature extractions (thickness maps, curvature, layer boundaries, volume measurements) and then integrating them through AI prediction models, the system maintains measurement simplicity while improving diagnostic accuracy for mild cases through multi-parameter assessment.
Data Source
Figure 1A
Figure 1B
Figure 2A~2B
AI summary
In some embodiments, a set of eye images related to a subject may be provided to a prediction model. A first prediction may be obtained via the prediction model, where the first prediction is derived from a first eye image and indicates whether an eye condition is present in the subject. A second prediction may be obtained via the prediction model, where the second prediction is derived from a second eye image and indicates that the eye condition is present in the subject. An aspect associated with the first prediction may be adjusted via the prediction model based on the second prediction's indication that the eye condition is present in the subject. One or more predictions related to at least one eye condition for the subject may be obtained from the prediction model, where the prediction model generates the predictions based on the adjustment of the first prediction.