Synthetic CT Image Generation via PSF Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating pairs of suitable training images and correct labels for machine learning in medical imaging, such as CT images, are inefficient, often requiring manual labor and may produce images that do not accurately represent real imaging conditions, leading to suboptimal classification accuracy.
Innovation Solution
A system that identifies the Point Spread Function (PSF) associated with existing CT images, transforms it, and applies a different PSF to generate new images with varying imaging conditions, ensuring the new images have the same correct labels as the original, thus enhancing data coverage and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual creation of correct labels by doctors is used, then classification accuracy can be improved, but productivity decreases due to increasing burden on doctors
Solution Approach 1:
The patent uses generative adversarial networks to create synthetic medical images that copy the characteristics of real medical images. These synthesized images serve as training data, replacing the need for manual label creation by doctors while maintaining classification accuracy. The system generates unlimited training samples without additional manual effort.
Solution Approach 2:
The system employs automated machine learning models that self-generate training data and labels without requiring continuous manual intervention from doctors. The generative model learns from existing labeled data and autonomously produces new training samples, freeing doctors from the burden of manual label creation.
2Quantity of substance
If image processing such as rotation, inversion, translation, scale change, color change, and contrast change is applied to training data, then quantity of training images increases, but the generated images may not accurately represent real imaging conditions
Solution Approach 1:
Instead of applying simple geometric transformations, the patent changes the fundamental parameters of image generation by using generative adversarial networks. The model learns the underlying distribution of real medical images and generates new samples with varied characteristics that naturally represent different imaging conditions, rather than artificially transforming existing images.
Solution Approach 2:
The patent replaces mechanical image processing operations (rotation, inversion, translation) with a learned generative model. Instead of applying predetermined transformations, the system uses neural networks to synthesize images that capture the true variability of medical imaging conditions, providing more realistic and diverse training data.
Data Source
AI summary
A non-transitory computer-readable storage medium storing a program that causes a computer to execute a process, the process includes identifying first filter processing applied to a first image that is training data used for machine learning; generating the first image from which characteristics of the identified first filter processing are removed; and generating, by applying second filter processing to the first image from which the characteristics are removed, a second image to be assigned a label identical to a label of the first image, the second image being used in the machine learning as the training data.