Synthetic Medical Image Generation via Physical Model Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The medical imaging domain faces a scarcity of labelled data for training deep neural networks, which is exacerbated by the manual generation of labels, limiting the availability of training data for tasks like image segmentation.
Innovation Solution
A method and system for automatically generating training datasets by using a predefined physical model to synthesize simulated images from quantitative maps acquired through imaging systems, where tunable simulation parameters are used to create realistic images with corresponding label information, allowing for efficient training of machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual label generation is used for training data, then label accuracy can be maintained, but the quantity of labelled data remains scarce and time-consuming to produce
Solution Approach 1:
The patent creates synthetic copies of medical images by simulating different imaging sequences and contrasts from a single quantitative map. These synthetic images serve as training data without requiring manual annotation, thus multiplying the available training data while maintaining consistency with real imaging characteristics.
Solution Approach 2:
The system uses the quantitative map itself to generate multiple simulated images with different contrasts and sequences. The quantitative map serves as a self-contained source that can generate its own training variations through physical model simulations, eliminating the need for external manual labeling processes.
2Reliability
If deep neural networks with many parameters are used, then model performance improves, but the amount of training data required increases significantly
Solution Approach 1:
The patent pre-computes quantitative maps from real images before generating synthetic training data. This preliminary processing extracts the essential physical properties that can then be used to generate multiple realistic variations, providing a solid foundation for training deep networks without requiring vast amounts of manually annotated data.
Solution Approach 2:
The system varies simulation parameters such as echo time, repetition time, and inversion time to generate diverse synthetic images from a single quantitative map. This parameter variation creates the diversity needed for training deep neural networks while starting from a limited set of real annotated images.
3Quantity of substance
If data augmentation techniques are applied to existing labelled data, then training data quantity increases, but the data variability and realism may be compromised
Solution Approach 1:
The patent replaces traditional geometric augmentation operations (rotation, flipping) with physics-based simulations that model actual MRI signal generation processes. This substitution ensures that synthetic images obey the same physical laws as real images, maintaining realism while providing diverse training examples.
Solution Approach 2:
The system dynamically generates images with varying contrasts and sequences by changing simulation parameters. Rather than applying static transformations, the quantitative map is re-simulated under different physical conditions, creating dynamically diverse yet physically consistent training data.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The present invention concerns a method (100) and a system (200) for automatically generating a training dataset for medical image analysis, the method comprising: - acquiring (101) at least one quantitative map of a biological object; - storing (102) in a memory each acquired quantitative map; - generating (103) a training dataset comprising simulated images of the biological object and for each simulated image, a corresponding label information; - storing (104) the training dataset; the method being characterized in that the image contrast of each simulated image is computed/synthetized from one or more of said acquired quantitative maps, wherein a processing unit uses a predefined physical model for automatically computing/synthetizing the simulated images.