Fundus Image CVD Risk Prediction Using Segmented CNNs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for predicting cardiovascular disease (CVD) risk using fundus images are inaccurate due to indirect predictors and failure to identify major contributors, such as blood pressure and glycaemic control, leading to many false positives and negatives.
Innovation Solution
A system and method utilizing convolutional neural networks (CNNs) for retinal image analysis, including Quality Assurance, eye-identification, and risk contributing factor sets to process fundus images, producing a feature vector for a CVD risk prediction neural network to determine overall CVD risk and the relative contribution of each risk factor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional statistical methods with regression models are used for CVD risk prediction, then the prediction can be made using available indirect predictors, but the accuracy is modest resulting in too many false positives and false negatives
Solution Approach 1:
The patent segments the CVD risk prediction task into multiple specialized CNN models, each trained to detect specific risk factors (blood pressure, glycaemic control, cholesterol, smoking) independently from retinal images. This segmentation allows each model to focus on specific features, improving detection accuracy and reducing false positives compared to conventional regression models that treat all predictors uniformly.
Solution Approach 2:
The patent introduces retinal images as an intermediary indicator to indirectly assess CVD risk factors. Instead of directly measuring blood pressure, glycaemic control, cholesterol, and smoking through traditional clinical methods, the system uses AI analysis of retinal vasculature changes as a mediator to infer these risk factors, providing a non-invasive and more accurate prediction method.
2Loss of information
If conventional CVD risk prediction equations are used, then the prediction can be made using indirect measures such as age, sex, ethnicity, smoking, diabetes duration, blood pressure, cholesterol, HbA1c, and ACR, but the available predictors fail to identify the major contributors of CVD risk
Solution Approach 1:
The patent divides the CVD risk assessment into separate specialized CNN models, each dedicated to detecting a specific risk factor (blood pressure, glycaemic control, cholesterol, smoking). This segmentation enables the system to identify and quantify the contribution of each major risk factor independently, rather than treating all predictors uniformly as in conventional regression models.
Solution Approach 2:
The patent analyzes changes in retinal vasculature characteristics (analogous to color changes) to detect risk factors. Specifically, it measures changes in vessel caliber, tortuosity, and other vascular features that reflect the impact of blood pressure, glycaemic control, cholesterol, and smoking on the retinal circulation, thereby identifying major contributors to CVD risk.
3Productivity
If retinal images are analyzed using AI deep learning algorithms trained against single labels such as chronological age or conventional CVD risk equations, then the analysis can be performed, but the outcome is a single number (perceived risk) that has proven to be inaccurate
Solution Approach 1:
Instead of using a single AI model trained on one label (such as chronological age or overall CVD risk), the patent employs multiple specialized CNN models, each trained to detect specific risk factors independently. This segmentation transforms the single-output approach into a multi-output system that provides detailed breakdown of risk contributors while maintaining efficient automated analysis.
Solution Approach 2:
The patent transitions from a single-dimensional output (one risk score) to a multi-dimensional output by analyzing multiple risk factors simultaneously through separate CNN models. Each model contributes a specific dimension of risk assessment (blood pressure effect, glycaemic control effect, cholesterol effect, smoking effect), creating a comprehensive risk profile rather than a single perceived risk number.
Data Source
AI summary
Systems and methods for predicting a risk of cardiovascular disease (CVD) from one or more fundus images are disclosed. Fundus images associated with an individual are processed to determine whether fundus images are of sufficient quality. The fundus images of sufficient quality are processed to identify fundus images belonging to a single eye. A plurality of risk contributing factor sets of CNNs (RCF CNN) are configured to output an indicator of probability of the presence of a different risk contributing factor in each of the one or more fundus images. At least one of the RCF CNNs is configured in a jury system model having a plurality of jury member CNNs, each being configured to output a probability of a different feature in the one or more fundus images. The outputs of the jury member CNNs are processed to determine the indicator of probability of the presence of the risk contributing factor output by the RCF CNN. An individual feature vector is produced based on meta-information for the individual, and the outputs of the RCF CNNs. The individual feature vector is processed using a CVD risk prediction neural network model to output a prediction of overall CVD risk for the individual. The model is configured to determine a relative contribution of each of the risk contributing factors to the prediction of overall CVD risk. The overall CVD risk is reported, together with the relative contribution of each of the risk contributing factors to the overall CVD risk.