Neural Network Medical Image Processing for Low-Dose CT Noise Reduction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing medical image processing technologies face challenges in improving low-dose CT image quality, particularly in reducing noise and artifacts, while also reducing computational time and hardware costs.

Innovation Solution

The use of deep learning (DL) networks, specifically convolutional neural networks (CNNs), to process reconstructed images and sinogram data, allowing for noise reduction and artifact mitigation in medical images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If model based iterative image reconstruction or sinogram restoration is used to improve low-dose CT image quality, then image quality is improved, but computational time increases and hardware costs increase

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

Solution Approach 1:

The patent uses a trained neural network model that has learned the mapping from low-dose to high-dose images during the training phase. During actual low-dose CT imaging, the pre-trained model directly processes the projection data or reconstructed images to produce high-quality images without requiring iterative reconstruction computations, thus dramatically reducing computational time while maintaining image quality.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The neural network model is trained in advance using paired high-dose and low-dose CT data to learn the degradation patterns and restoration strategies. This preliminary training phase prepares the model so that during actual clinical use, only fast forward propagation is needed to restore low-dose images, avoiding the need for time-consuming iterative reconstruction at the time of imaging.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If model based iterative image reconstruction or sinogram restoration is used to improve low-dose CT image quality, then image quality is improved, but hardware costs increase

Engineering Contradiction:
Improveimage qualityVSAvoidhardware costs
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces complex iterative reconstruction algorithms that require sophisticated hardware implementations with a neural network-based approach. The neural network, once trained, can be deployed on standard computing hardware or even specialized AI accelerators, significantly reducing the hardware complexity and cost compared to implementing model-based iterative reconstruction systems.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Productivity

If deep learning networks are used to process medical images, then image quality is improved and processing speed is increased, but system complexity increases

Engineering Contradiction:
Improveprocessing speedVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent extracts the complex image restoration task into a separate, pre-trained neural network module. This modular approach allows the main CT reconstruction pipeline to remain relatively simple while delegating the complex denoising and artifact reduction tasks to the specialized neural network, thereby managing system complexity through functional separation.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3467766B1Medical image processing apparatus and medical image processing system
Publication Date: 2025.04.30 CANON MEDICAL SYST CORP
  • EP3467766B1 patent drawingFigure 1A
  • EP3467766B1 patent drawingFigure 1B
  • EP3467766B1 patent drawingFigure 2A

AI summary

A medical image processing apparatus according to an embodiment comprises a memory and processing circuitry. The memory is configured to store a plurality of neural networks corresponding to a plurality of imaging target sites, respectively, the neural networks each including an input layer, an output layer, and an intermediate layer between the input layer and the output layer, and each generated through learning processing with multiple data sets acquired for the corresponding imaging target site. The processing circuitry is configured to process first data into second data using, among the neural networks, the neural network corresponding to the imaging target site for the first data, wherein the first data is input to the input layer and the second data is output from the output layer.