Synthetic Radiologic Image Generation With Iterative Contrast Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for generating synthetic medical images, such as those described in WO2019/074938A1 and WO2022184297A1, face challenges like instability, overfitting, and the generation of unrealistic images, particularly when reducing the amount of contrast agent used.
Innovation Solution
A pre-trained conditional generative model is used to generate synthetic representations of examination areas with reduced contrast enhancement by iteratively modifying embeddings and model parameters based on native and contrast-enhanced representations, quantifying deviations, and refining the output to match the desired contrast level.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If convolutional neural networks are used to generate synthetic medical images, then image generation capability is improved, but training stability and model reliability deteriorate
Solution Approach 1:
The patent applies preliminary action by using a pre-trained conditional generative model that has already been trained on a large dataset of reference representations. This pre-training step is performed beforehand, allowing the model to capture general patterns and relationships in medical imaging data. When generating synthetic images, the pre-trained model can be directly applied without requiring extensive re-training, thus maintaining both high image generation capability and training stability.
2Productivity
If convolutional neural networks are used to generate synthetic medical images, then image generation capability is improved, but manufacturing precision deteriorates
Solution Approach 1:
The patent implements feedback through an iterative refinement process where the generated synthetic representation is transformed and compared against the target contrast-enhanced representation. Deviations are quantified and used to modify the starting data, embeddings, or model parameters. This feedback loop continues for one or more iterations, progressively improving the realism and accuracy of the generated images while maintaining the efficiency of the pre-trained model.
3Loss of information
If contrast agent is administered to obtain contrast-enhanced representations, then diagnostic information is improved, but loss of substance increases
Solution Approach 1:
The patent applies copying by generating synthetic contrast-enhanced representations that replicate the diagnostic information found in real contrast-enhanced images. The conditional generative model learns the relationship between native and contrast-enhanced representations from training data, then uses this learned knowledge to synthesize realistic contrast-enhanced images from native representations. This allows obtaining diagnostic information without actually administering contrast agents to the new examination object.
4Manufacturing precision
If extensive hyperparameter tuning is performed to improve model performance, then manufacturing precision is improved, but loss of time increases
Solution Approach 1:
The patent eliminates extensive hyperparameter tuning by using a pre-trained model with fixed, optimized parameters. The pre-training phase is performed in advance on comprehensive datasets, capturing optimal parameter settings. During deployment, the pre-trained model can generate synthetic images directly without requiring time-consuming hyperparameter adjustment, thus maintaining high image quality while significantly reducing time loss.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
Systems, methods, and computer programs disclosed herein relate to generating synthetic representations, such as synthetic radiologic images.