AI Echocardiogram Style Transfer for Bias-Resistant CHD Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing medical imaging systems for detecting cardiovascular anomalies, particularly congenital heart defects, suffer from bias due to training data that includes style information specific to certain imaging systems, leading to inaccurate anomaly detection.
Innovation Solution
A system and method that processes medical images using a style transfer generator to incorporate representative styles from multiple imaging systems, followed by an anomaly detection model to identify cardiovascular anomalies, reducing bias and improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a model is trained using image data from multiple imaging systems, then the model's adaptability to different imaging systems improves, but the model may still associate anomalies with specific imaging system styles rather than actual anatomical features
Solution Approach 1:
The patent introduces a style transfer generator as an intermediary component that transfers the visual style of images from one imaging system to match another. This mediator allows the model to learn from diverse imaging systems while maintaining consistency in style representation, thereby preventing the model from associating anomalies with specific system styles rather than actual anatomical features.
Solution Approach 2:
The patent changes the style parameters of images using a style transfer generator. By modifying visual style parameters (such as color distribution, texture, and overall appearance) while preserving anatomical content, the system enables the model to learn anomaly detection in a style-invariant manner, improving reliability across different imaging systems.
2Measurement precision
If style transfer is applied to normalize images from different imaging systems, then measurement precision of anomaly detection improves, but device complexity increases due to additional processing components
Solution Approach 1:
The patent applies style transfer as a preliminary action before the main anomaly detection process. By pre-normalizing the visual style of input images to match a target imaging system's style, the system eliminates style-related variability that would otherwise interfere with anomaly detection, thereby improving measurement precision without requiring complex modifications to the core detection algorithm.
Data Source
AI summary
Systems and methods are provided for processing image data generated by a medical imaging system such as an ultrasound or echocardiogram system using artificial intelligence and machine learning to determine a presence of one or more congenital heart defects (CHDs) and/or other cardiovascular anomalies in the image data in a manner that is agnostic to the type of imaging system, software, and/or hardware. Image data from various types imaging systems, software, and/or hardware, having various styles of imaging data generated may be processed to determine image styles. Input image data for analysis may then be processed together with representative styles of image data to generate styled input images for each style. The styled input images may be processed by an image analyzer to detect one or more cardiovascular anomalies in the styled image data, for example. Alternatively, training data may be styled and used to train the image analyzer.