Trained Models Combining Reconstruction Methods for CT Artifact Reduction

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Medical imaging devices using compressed sensing in sparse-view CT or low-dose CT suffer from image quality degradations such as stair-step artifacts and loss of smooth density changes and textures.

Innovation Solution

A trained model generation program that performs machine learning using a combination of input images from different reconstruction methods, including compressed sensing and analytical reconstruction, to generate a trained model that mitigates these degradations by optimizing smoothing parameters and weights in a convolutional neural network.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If compressed sensing reconstruction method is used to reduce dose exposure and scanning time, then productivity is improved, but manufacturing precision deteriorates due to stair-step artifacts and loss of smooth density changes

Engineering Contradiction:

Improvescanning speedVSAvoidimage quality

Core Design Contradiction:

ProductivityVSManufacturing precision

Solution Approach 1:

A deep learning model is introduced as an intermediary between the compressed sensing reconstruction and the final image output. The model takes the reconstructed image containing artifacts as input and transforms it into an artifact-reduced image, effectively mediating the trade-off between fast scanning and high image quality

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system changes the parameter of smoothing strength in the compressed sensing reconstruction and uses the deep learning model to compensate. By optimizing the smoothing parameter and training the model with various smoothing strengths, the system achieves both fast reconstruction and high image quality

Inventive Principle:

Principle #35Parameter changes

2Use of energy by stationary object

If compressed sensing reconstruction method is used to reduce dose exposure, then use of energy by stationary object is reduced, but object-affected harmful factors increase due to stair-step artifacts and elimination of textures

Engineering Contradiction:

Improvedose exposureVSAvoidimage artifacts

Core Design Contradiction:

Use of energy by stationary objectVSObject-affected harmful factors

Solution Approach 1:

The deep learning model serves as a mediator that processes the artifact-containing images from low-dose compressed sensing reconstruction and outputs artifact-reduced images, allowing low-dose scanning without suffering from the harmful artifacts

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The system intentionally introduces smoothing during compressed sensing reconstruction, which creates artifacts, then uses the deep learning model to learn and remove these artifacts. The harmful artifacts are converted into training data that helps the model learn the relationship between artifact-containing and artifact-free images

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

3Manufacturing precision

If smoothing parameter is increased in compressed sensing to reduce artifacts, then image quality is improved, but loss of information increases due to elimination of smooth density changes and textures

Engineering Contradiction:

Improveimage qualityVSAvoiddensity changes

Core Design Contradiction:

Manufacturing precisionVSLoss of information

Solution Approach 1:

The deep learning model acts as an intermediary that can reverse the information loss caused by smoothing. It learns the mapping from smoothed images to original images, recovering the lost density changes and textures that were eliminated by the smoothing process

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

Instead of trying to preserve all details during smoothing (which is impossible), the system applies smoothing and then inverts the process using the deep learning model to recover the original information, effectively doing the reconstruction process in reverse

Inventive Principle:

Principle #13The other way round (Inversion)

Data Source

PatentUS12367620B2Trained model generation program, image generation program, trained model generation device, image generation device, trained model generation method, and image generation method

Publication Date: 2025.07.22 UNIV OF TSUKUBA

AI summary

A trained model generation program causes a computer to implement a learning execution function of inputting first input image data representing a first input image generated by a first reconstruction method using compressed sensing and second input image data representing a second input image generated by a second reconstruction method different from the first reconstruction method to a machine learning device to execute machine learning, the second reconstruction method being an analytical reconstruction method, and causing the machine learning device to generate a trained model, and a trained model acquisition function of acquiring trained model data indicating the trained model. Then, input image data representing an input image is input to the trained model to generate a reconstructed image with improved image quality.