Deep Learning CT Noise Reduction via Synthetic Training Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current CT image enhancement techniques for reducing noise while minimizing radiation dose are either costly or require significant hardware and software upgrades, and existing deep learning methods lack effective training strategies for medical images, which are critical for safety and performance.

Innovation Solution

A deep learning-based method that involves generating and training multiple models using a CT image noise simulator to create composite high-noise images, allowing the models to learn noise reduction, and selecting the appropriate model based on input image characteristics to output a high-quality image with reduced noise.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If high quality CT systems with reduced radiation exposure are released, then radiation dose is minimized, but the price is higher than existing products

Engineering Contradiction:
Improveradiation doseVSAvoidprice
Core Design Contradiction:
Object-affected harmful factorsVSEase of manufacture

Solution Approach 1:

The patent uses deep learning models trained on synthetic noisy CT images to copy and replicate the noise reduction capabilities of expensive hardware systems. The model learns from simulated training data and applies learned patterns to real low-dose CT images, achieving high quality reconstruction without requiring costly hardware upgrades.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent replaces mechanical/hardware-based noise reduction mechanisms with a software-based deep learning approach. Instead of using complex hardware systems to achieve low-dose high-quality imaging, the system uses trained neural networks to process and reconstruct images, substituting computational methods for physical mechanisms.

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

2Object-affected harmful factors

If hardware and software upgrades are performed on existing CT products, then high quality CT images with reduced radiation exposure can be acquired, but the upgrade cost is considerable

Engineering Contradiction:
Improveradiation doseVSAvoidupgrade cost
Core Design Contradiction:
Object-affected harmful factorsVSDevice complexity

Solution Approach 1:

The patent segments the noise reduction function into a separate, independently trainable deep learning model. This allows the noise reduction capability to be developed and updated separately from the main CT hardware, enabling cost-effective deployment as a standalone software module that can be applied to existing systems without comprehensive upgrades.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The deep learning model is designed to be universally applicable to different CT systems and imaging scenarios. The same model architecture can process various types of CT images (different body parts, different scan protocols), making it a multi-functional solution that doesn't require system-specific customization or expensive targeted upgrades.

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

3Object-affected harmful factors

If deep learning is applied to CT image noise reduction, then high quality diagnostic images can be obtained with minimized radiation dose, but effective training strategies for medical images are required to ensure safety and performance

Engineering Contradiction:
Improveradiation doseVSAvoidtraining complexity
Core Design Contradiction:
Object-affected harmful factorsVSDevice complexity

Solution Approach 1:

The patent performs preliminary actions by generating synthetic noisy CT images through simulation before actual model deployment. The training data is prepared in advance using physics-based noise models that replicate real low-dose CT noise characteristics, allowing the model to learn from realistic synthetic examples before being applied to actual patient data, ensuring both safety and performance.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces a synthetic training data intermediary between the training process and real medical images. Instead of directly training on limited real low-dose images, the model is trained on synthetically generated images that serve as an intermediate representation, bridging the gap between clean simulated images and real noisy clinical images while maintaining data privacy and safety.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3869446B1Apparatus and method for deep learning-based CT image noise reduction
Publication Date: 2024.04.10 CLARIPI INC
  • EP3869446B1 patent drawingFigure 1~2
  • EP3869446B1 patent drawingFigure 3

AI summary

Provided is a deep learning based CT image noise reducing method. The deep learning based CT image noise reducing method includes extracting test information from an input CT image; selecting at least one deep learning model corresponding to the test information, among a plurality of previously trained deep learning models; and outputting a CT image with a reduced noise with respect to the input CT image with the input CT image as an input of the at least one selected deep learning model.