Spectral Tomographic Image Material Separation Using Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current spectral tomographic image processing technologies are limited in diagnostic efficiency and accuracy due to the inability to effectively separate materials in images, which restricts direct improvements in diagnostic efficiency and the ability to predict tumors or lesions.
Innovation Solution
A tomographic image processing apparatus and method that utilizes machine learning to separate materials in spectral tomographic images by training a material separation model and providing the results through a user interface, with data augmentation to reduce the burden of collecting training data and improve performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If spectral tomographic images are displayed with limited information (monochromatic images at each energy step or effective-z values), then the image display is simple and easy to process, but diagnostic efficiency and accuracy are limited
Solution Approach 1:
The patent segments the spectral tomographic image into multiple material components (e.g., soft tissue, bone, fat, calcification) by separating materials based on their unique attenuation characteristics at different energy levels. This segmentation allows diagnostic information to be extracted from specific material regions, thereby improving diagnostic efficiency without requiring complex full-image processing.
Solution Approach 2:
The patent extracts material separation information from the spectral tomographic image by identifying and isolating specific materials based on their attenuation profiles. The system extracts quantitative information about material concentration and distribution, making diagnostic data more accessible and improving efficiency without significantly increasing processing complexity.
2Measurement precision
If material separation is performed using traditional methods, then the processing is computationally simple, but the accuracy of material separation is insufficient for reliable diagnosis
Solution Approach 1:
The patent employs iterative optimization algorithms that use feedback from the spectral data to refine material separation results. The system adjusts material concentration estimates based on the discrepancy between predicted and actual attenuation measurements, progressively improving accuracy through multiple iterations without requiring excessive computational power.
Solution Approach 2:
The patent utilizes the energy-dependent attenuation parameters of different materials to achieve accurate separation. By analyzing how attenuation coefficients vary with photon energy for each material type, the system can precisely distinguish between materials with similar densities, significantly improving measurement precision through parameter-based differentiation.
3Measurement precision
If a material separation model is trained using machine learning, then the accuracy of material separation is improved, but the burden of collecting training data increases
Solution Approach 1:
The patent creates synthetic training data by generating virtual spectral tomographic images from phantom objects with known material compositions and attenuation characteristics. These synthetic copies of real patient data provide abundant training examples without requiring actual patient scans, thereby improving model accuracy while avoiding the burden of collecting extensive real-world training data.
Solution Approach 2:
The patent performs preliminary training of the material separation model using synthetic data from phantoms before applying it to real patient images. This preliminary action establishes the model's material differentiation capabilities in advance, allowing it to achieve high accuracy on actual diagnostic data without requiring large quantities of labeled patient images for training.
4Measurement precision
If more training data is collected to improve model performance, then the accuracy of material separation improves, but the time and resources required for data collection increase
Solution Approach 1:
The patent generates synthetic training data by computationally simulating spectral tomographic images from phantom objects with known material properties. These virtual copies can be generated rapidly without requiring physical data collection, providing abundant training examples that improve model accuracy while minimizing time investment compared to collecting real patient data.
Solution Approach 2:
The system uses phantom objects with known material compositions to self-generate training data through spectral imaging. The phantoms serve as self-contained training targets where the ground truth material distribution is predetermined, allowing the system to create its own training data autonomously without external data collection efforts, thereby reducing time and resource requirements.
Data Source
AI summary
A tomographic image processing apparatus including a display, an input interface configured to receive an external input, a storage storing an input tomographic image of an object, and at least one processor configured to control the display to display the input tomographic image, determine a material combination to be separated from the input tomographic image, and control the display to display material separation information corresponding to the determined material combination for a region of interest selected in the input tomographic image based on the external input. The input tomographic image is a spectral tomographic image having a plurality of tomographic images respectively corresponding to a plurality of energy levels.