Trained Models Using Synthetic X-Ray Projections for Multi-Element Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods are limited to processing a single specific image element and require actual CT image data to perform machine learning, making it difficult to efficiently process various and multiple image elements.
Innovation Solution
A method of creating a trained model by generating a superimposed image from a three-dimensional X-ray image data simulation, allowing machine learning without relying on actual CT data, and performing inter-image arithmetic operations to process multiple image elements separately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning is performed using actual CT image data including the image element to be extracted, then the training data reflects real medical conditions, but it requires preparation of specific CT data and limits processing to actually included elements
Solution Approach 1:
The patent creates synthetic training data by copying and combining actual medical image data with simulated image elements. The system generates artificial projections of image elements (such as bones, blood vessels, or medical devices) and superimposes them onto actual medical images, creating realistic training samples without requiring actual CT data containing every possible target element.
Solution Approach 2:
The system performs preliminary processing by pre-generating projection images of various image elements through simulation before they are needed for training. These pre-computed projections are stored and can be rapidly combined with actual medical images during the training process, enabling efficient creation of diverse training datasets.
2Ease of manufacture
If image processing is limited to one specific image element (specific part), then the processing is simple and focused, but it cannot process various and multiple image elements simultaneously
Solution Approach 1:
The patent develops a universal image processing system that can handle multiple types of image elements through a single trained model. The system extracts features common to various image elements (bones, blood vessels, medical devices, etc.) and processes them using the same neural network architecture, enabling multi-functional processing without requiring separate specialized models for each element type.
Solution Approach 2:
The system segments the image processing task by separately extracting and processing different image elements through the trained model. Each image element can be identified and processed independently, allowing the system to handle multiple element types in a single image while maintaining focused processing for each specific element.
3Measurement precision
If CT image data is required to perform machine learning, then the training is based on real anatomical structures, but it is difficult to efficiently perform learning on various image elements without actual CT data
Solution Approach 1:
The patent introduces simulated projection images as an intermediary between actual medical images and the training process. These simulations act as a bridge, providing realistic anatomical contexts from actual images while adding synthetic image elements that represent various target objects, thereby eliminating the need to obtain actual CT data containing every possible target element.
Solution Approach 2:
The system varies parameters in the simulated projections (such as position, orientation, size, and appearance of image elements) to create diverse training samples from a limited set of actual medical images. This parameter variation allows efficient generation of numerous training examples without requiring proportional increases in actual CT data.
Data Source
AI summary
In a creation method of a trained model, a reconstructed image (60) obtained by reconstructing three-dimensional X-ray image data (80) is generated. A projection image (61) is generated from a three-dimensional model of an image element (50) by a simulation. The projection image is superimposed on the reconstructed image to generate a superimposed image (67). A trained model (40) is created by performing machine learning using the superimposed image, and the reconstructed image or the projection image.