Ophthalmic Image Feature Models for Rare Disease Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing global image-based machine learning models for ophthalmic disease detection face challenges in optimizing models for rare diseases due to the scarcity of training images and the need for retraining when different diagnostic rules are applied across regions.
Innovation Solution
Implementing machine learning models that identify specific features correlated with various ophthalmic and systemic diseases, allowing for easier training with partial images depicting these features and enabling adaptation to different diagnostic rules without full model retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If global image-based ML models are used for disease detection, then the model can directly indicate disease presence from images, but it requires numerous training images and is difficult to optimize for rare diseases
Solution Approach 1:
The patent segments the disease detection task into multiple independent feature detection models. Each model is trained to detect specific retinal features (e.g., hemorrhages, exudates, microaneurysms) rather than requiring a single global model to detect all diseases. This segmentation allows each feature model to be trained on fewer, more targeted images while maintaining high detection accuracy.
Solution Approach 2:
The patent extracts and detects individual retinal features separately from the full image before synthesizing disease diagnoses. By taking out specific features (hemorrhages, exudates, etc.) as independent detection targets, the system reduces the training data requirement for each feature detector compared to training a global model on all possible disease presentations.
2Adaptability or versatility
If global image-based ML models are trained to identify diseases, then the model can detect disease presence, but it requires full model retraining when different diagnostic rules are applied in different regions
Solution Approach 1:
The patent segments the diagnostic system into independent feature detection models and a separate disease synthesis component. The feature detection models remain consistent across regions, while only the synthesis rules need to be adjusted for different diagnostic guidelines. This segmentation eliminates the need to retrain entire global models when diagnostic rules change.
Solution Approach 2:
The patent implements dynamic adaptability by allowing the disease synthesis component to be reconfigured based on regional diagnostic rules without retraining the underlying feature detection models. The system dynamically adjusts which detected features contribute to which disease diagnoses according to the applicable diagnostic guidelines.
Data Source
AI summary
An example method includes identifying at least one image of an eye of a patient. The method further includes detecting, by a first computing model, at least one first feature in the at least one image and detecting, by a second computing model, at least one second feature in the at least one image. Further, using a third computing model that is different than the first computing model or the second computing model, the method includes identifying a likelihood that the patient has one or more diseases consistent with the at least one feature and the at least one second feature. A recommendation for care of the patient is generated based on the likelihood.