Tomographic Lesion Image Synthesis for Realistic AI Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating training data for machine learning models to detect lesions in tomographic images from tomosynthesis imaging are inadequate, leading to low accuracy in lesion detection due to the use of two-dimensional image combinations and lack of realistic training data.
Innovation Solution
A method to generate tomographic images by combining lesion images with projection images based on geometrical relationships and radiation attenuation coefficients, creating accurate training data for machine learning models to detect lesions in tomographic images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If lesion images are added to two-dimensional normal images to generate training data, then the quantity of training data increases, but the accuracy of lesion detection in tomographic images deteriorates
Solution Approach 1:
The patent transitions from two-dimensional image addition to three-dimensional tomographic image synthesis. By incorporating depth information and reconstructing images in three dimensions using back-projection methods, the system generates training data that matches the actual tomographic imaging modality, thereby improving detection accuracy while maintaining sufficient data quantity
Solution Approach 2:
The patent changes the fundamental parameter of image dimensionality from 2D to 3D. By synthesizing tomographic images with embedded lesion information in three dimensions and using multiple projection angles, the training data more accurately represents real clinical images, resolving the contradiction between data quantity and detection accuracy
2Ease of manufacture
If simple combination of lesion images with normal images is used, then the ease of manufacture of training data improves, but the realism of training images deteriorates
Solution Approach 1:
The patent creates realistic copies of actual tomographic images by synthesizing images that replicate the physical imaging process. By using back-projection methods and incorporating geometric relationships between radiation sources and lesions, the system generates training images that are faithful reproductions of real clinical data characteristics
Solution Approach 2:
The patent transforms the image synthesis approach by changing parameters such as projection angles, radiation source positions, and reconstruction algorithms. This creates training images with realistic geometric and radiological properties while maintaining automated generation efficiency
3Device complexity
If two-dimensional images are used for training, then the device complexity is reduced, but the measurement precision of lesion detection in tomographic images deteriorates
Solution Approach 1:
The patent implements three-dimensional image reconstruction from multiple two-dimensional projections using back-projection algorithms. This adds the depth dimension to training data, enabling the machine learning model to learn features specific to tomographic images and improve lesion detection accuracy in the actual three-dimensional imaging modality
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enhances the accuracy of lesion detection in tomographic images by generating realistic training data, allowing for high-precision machine learning models to identify lesions effectively.
Implementation Method 1
combining the lesion image with a plurality of first projection images on a basis of a geometrical relationship between a plurality of radiation source positions and a position of the lesion virtually disposed in the object to derive a plurality of second projection images
Data Source
AI summary
A processor acquires a plurality of first projection images acquired by imaging an object at a plurality of radiation source positions and acquires a lesion image indicating a lesion. The processor combines the lesion image with the plurality of first projection images on the basis of a geometrical relationship between the plurality of radiation source positions and a position of the lesion virtually disposed in the object to derive a plurality of second projection images. The processor reconstructs the plurality of second projection images to generate a tomographic image including the lesion.