CT Image Denoising Training Data via Photon-Count Splitting

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing methods for generating image denoising training data for deep learning neural networks in computed tomography (CT) images face challenges such as excessive radiation exposure, misregistration, and correlated noise due to repeat scans or noise injection, making it difficult to obtain clean ground-truth images for training.

Innovation Solution

Utilizing a photon-counting-detector CT scanner to generate sinograms, which are split into reduced-photon-count sets via photon-wise binomial selection, allowing for the creation of independent noisy training inputs and outputs from a single scan, thereby avoiding repeat scans and correlated noise.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If repeat scans are performed to obtain clean ground-truth images for training, then image quality is improved, but radiation exposure increases

Engineering Contradiction:

Improveimage qualityVSAvoidradiation exposure

Core Design Contradiction:

Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent segments the photon counts from a single scan into multiple synthetic datasets by applying different noise models and photon-count splitting ratios. This creates multiple training samples (noisy input images and corresponding clean ground-truth images) from one physical scan, eliminating the need for repeat scans and reducing radiation exposure while maintaining image quality for training purposes

Inventive Principle:

Principle #1Segmentation

2Measurement precision

If repeat scans are performed to obtain training data, then image quality is improved, but misregistration occurs

Engineering Contradiction:

Improveimage qualityVSAvoidregistration accuracy

Core Design Contradiction:

Measurement precisionVSReliability

Solution Approach 1:

The patent performs preliminary processing of the sinogram data from a single scan by applying photon-count splitting and noise injection before image reconstruction. This ensures that all training images (noisy and clean) are derived from the same original projection data, guaranteeing perfect spatial alignment and eliminating misregistration issues that would occur with repeat scans

Inventive Principle:

Principle #10Preliminary action

3Productivity

If noise injection is used to generate training data, then training dataset is created, but correlated noise is introduced

Engineering Contradiction:

Improvetraining dataset generationVSAvoidnoise independence

Core Design Contradiction:

ProductivityVSLoss of information

Solution Approach 1:

The patent extracts independent noise realizations by applying photon-count splitting to separate photons into different bins, where each bin represents an independent Poisson process. This extraction method ensures that the noise in different training samples is statistically independent, unlike simple noise injection which creates correlated noise artifacts

Inventive Principle:

Principle #2Taking out (Extraction)

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

This approach enables effective image denoising performance comparable to using clean ground-truths while reducing radiation exposure and misregistration, improving the training of deep learning neural networks for CT image denoising.

Implementation Method 1

a set of sinograms generated by a photon-counting computed tomography scanner

Methodology Applied

Scientific EffectPhoton counting: Photoelectric Effect

Implementation Method 2

split the set of sinograms into a first reduced-photon-count set of sinograms and a second reduced-photon-count set of sinograms

Methodology Applied

Scientific EffectPhoton-wise binomial selection:

Implementation Method 3

convert, via image reconstruction, the first reduced-photon-count set of sinograms into at least one training input image and the second reduced-photon-count set of sinograms into at least one training output image

Methodology Applied

Scientific EffectImage reconstruction: Tomography

Data Source

PatentUS12462347B2Generation of image denoising training data via photon-count splitting

Publication Date: 2025.11.04 GE PRECISION HEALTHCARE LLC

AI summary

Systems/techniques that facilitate generation of image denoising training data via photon-count splitting are provided. In various embodiments, a system can access a set of sinograms generated by a photon-counting computed tomography scanner. In various aspects, the system can split the set of sinograms into a first reduced-photon-count set of sinograms and a second reduced-photon-count set of sinograms. In various instances, the system can convert, via image reconstruction, the first reduced-photon-count set of sinograms into at least one training input image and the second reduced-photon-count set of sinograms into at least one training output image. In various cases, the system can train a deep learning neural network based on the at least one training input image and the at least one training output image.