Synthetic Training Data Transfer for Contrast Enhancement Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems for image enhancement, particularly those with deep architectures, suffer from overfitting and bias due to limited and biased training data, leading to poor generalization and performance degradation, especially in contrast-enhanced medical imaging, which also poses health and economic risks from excessive use of contrast media.
Innovation Solution
A system for synthesizing training data by transferring image information between high and low quality imagery, denoising, and adding noise components, using generative adversarial networks to create paired training data, reducing the need for large sets of labeled clinical cases, and enabling effective training of machine learning models for image enhancement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If more training data is provided to machine learning systems, then model performance and generalization improve, but data acquisition cost and complexity increase
Solution Approach 1:
The patent creates synthetic training data by copying and transforming existing medical images through generative adversarial networks. Instead of acquiring new real patient data, the system generates artificial training samples that mimic real imaging data, thereby improving model generalization without the complexity of collecting and annotating additional clinical cases
Solution Approach 2:
The system performs preliminary data synthesis by pre-generating training datasets using GANs before actual model training. This preliminary action creates a robust training corpus in advance, eliminating the need for complex real-time data acquisition and annotation processes during model development
2Manufacturing precision
If contrast media is administered to improve image contrast, then diagnostic quality improves, but patient health risks and costs increase
Solution Approach 1:
The patent replaces the physical-chemical mechanism of contrast media administration with a computational mechanism using machine learning. Instead of introducing iodine-based contrast agents that carry health risks, the system uses trained neural networks to computationally enhance image contrast, achieving the same diagnostic quality improvement without harmful side effects
Solution Approach 2:
The trained machine learning model acts as an intermediary between raw medical images and diagnostic interpretation. This intermediary computationally adjusts image contrast and quality, serving as a substitute for physical contrast media while maintaining the enhanced visualization needed for accurate diagnosis
3Manufacturing precision
If deep architecture machine learning models are used for image enhancement, then image quality improvement increases, but overfitting and bias increase
Solution Approach 1:
The system performs preliminary generation of diverse synthetic training data using GANs before training deep architecture models. This pre-prepared, varied training corpus helps deep networks learn robust features without overfitting, maintaining high image quality enhancement capability while improving model reliability through exposure to synthetic variations during training
Data Source
AI summary
A system (DSS) and related method for synthesizing training data or machine learning, based on a set (TD) including two types of training imagery, high image quality, IQ, imagery and low IQ imagery. The system comprises a data synthesizer (DSY), configured to register the at least two types of imagery and to transfer i) image information from high IQ imagery to the registered low IQ imagery to obtain synthesized high IQ imagery, or ii) image information from low IQ imagery to the registered high IQ imagery to obtain synthesized low IQ imagery. The synthesized data may be used for improved training of machine learning models for IQ enhancement.