Neural Network Cone-Beam Artifact Reduction in CT Imaging

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Computed tomography (CT) scans, particularly cone-beam CT, suffer from imaging artifacts that degrade image quality, hindering clinical applications.

Innovation Solution

A medical image processing apparatus and method using a trained neural network to reduce cone-beam artifacts by separating images into low-frequency and high-frequency components, with the neural network trained on pairs of images with and without artifacts, and applying these models to correct images with cone-beam artifacts during scanning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If conventional CT reconstruction methods are used, then the imaging process is simple and fast, but imaging artifacts appear that degrade image quality

Engineering Contradiction:
Improveimage qualityVSAvoidprocessing complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The system performs preliminary actions by acquiring additional projection data at different cone angles during the scanning process, and pre-trains multiple neural network models for different anatomical structures and scanning protocols before actual image reconstruction, so that artifact correction can be applied efficiently during scanning

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces trained neural network models as intermediaries between the raw projection data and the final reconstructed image. These models act as mediators that process the projection data to eliminate cone-beam artifacts while preserving diagnostic image quality, resolving the contradiction between simple reconstruction and artifact-free images

Inventive Principle:
Principle #24Intermediary (Mediator)

2Manufacturing precision

If deep learning frameworks are integrated to correct cone-beam artifacts, then image quality improves, but computational time and hardware costs increase

Engineering Contradiction:
Improveimage qualityVSAvoidcomputational time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent segments the image correction task by training separate neural network models for different anatomical structures (head, body, pelvis) and different scanning protocols. This segmentation allows the system to apply only the necessary model for each specific case, reducing overall computational time compared to using a single comprehensive deep learning framework

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system changes parameters by selecting different pre-trained models based on the specific anatomical structure and scanning protocol being used. This parameter-based model selection optimizes computational efficiency by matching the correction algorithm to the specific imaging conditions, rather than applying a fixed complex deep learning framework to all cases

Inventive Principle:
Principle #35Parameter changes

3Manufacturing precision

If a single trained model is used for all scanning conditions, then the system is simple to operate, but accuracy decreases for specific anatomical structures or protocols

Engineering Contradiction:
Improvecorrection accuracyVSAvoidmodel selection complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent applies local quality by training separate neural network models optimized for specific anatomical structures (head, body, pelvis) and specific scanning protocols. Each model has specialized quality characteristics tailored to its designated application, ensuring high correction accuracy for each specific case rather than using a generic single model

Inventive Principle:
Principle #3Local quality

4Manufacturing precision

If multiple trained models are maintained for different parameters, then correction accuracy for specific cases improves, but device complexity and memory requirements increase

Engineering Contradiction:
Improvecorrection accuracyVSAvoidmemory storage
Core Design Contradiction:
Manufacturing precisionVSVolume of stationary object

Solution Approach 1:

The system achieves universality by designing a modular architecture where multiple trained models can be stored and selected based on the specific scanning conditions. The framework is designed to handle different anatomical structures and protocols through a unified interface, allowing the same system to serve multiple specialized functions without requiring separate hardware for each model

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3716214B1Medical image processing apparatus and method for acquiring training images
Publication Date: 2024.03.20 CANON MEDICAL SYST CORP
  • EP3716214B1 patent drawingFigure 1A~1B
  • EP3716214B1 patent drawingFigure 2A
  • EP3716214B1 patent drawingFigure 2B

AI summary

According to one embodiment, a medical image processing apparatus includes a processor and an acquirer. The processor inputs, to a trained model trained based on a first image acquired with a first x-ray beam having a first cone angle and a second image including a cone-beam artifact and acquired using simulation based on the first image, a third image including a cone-beam artifact and acquired with a second x-ray beam having a second cone angle larger than the first cone angle to generate a fourth image corresponding to the third image with a reduced cone-beam artifact. The acquirer acquires the third image.