CT Voxel Segmentation Using Autoencoder Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CT inspection systems face challenges in providing accurate segmentation and surface determination due to reconstruction artefacts and noise in CT voxel data, limiting their application in metrology tasks and requiring more expensive and complex hardware.
Innovation Solution
The use of trained machine learning models, such as autoencoder convolutional neural networks, to denoise and remove artefacts from CT voxel data by adding artificial noise or using artefact-free training data, allowing for improved binarization and segmentation of CT voxel data to facilitate reliable surface determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional CT inspection systems are used for segmentation and surface determination, then the system structure is simpler, but the measurement precision and segmentation accuracy deteriorate due to reconstruction artefacts and noise
Solution Approach 1:
The patent replaces traditional mechanical/image processing methods with machine learning models (autoencoders and segmentation networks) to denoise and segment CT data. The ML-based denoising network processes CT images to remove noise and artifacts, while the segmentation network performs accurate material segmentation, significantly improving measurement precision without requiring complex hardware modifications
Solution Approach 2:
The patent introduces machine learning models as intermediary components between CT data acquisition and final segmentation/surface determination. The denoising network acts as an intermediary to clean the data before segmentation, and the segmentation network serves as another intermediary to extract meaningful information, thereby improving accuracy without direct hardware changes
2Reliability
If traditional CT inspection systems are used, then hardware costs are lower, but the reliability and confidence of surface determination deteriorate due to noise and artefacts
Solution Approach 1:
The patent substitutes expensive hardware improvements with software-based machine learning solutions. Instead of upgrading to more expensive CT hardware to reduce noise and improve reliability, the system uses ML denoising and segmentation networks that process the existing CT data to achieve high reliability and confidence in surface determination at lower hardware cost
3Measurement precision
If ML denoising and segmentation is applied, then measurement precision improves, but processing time and computational complexity increase
Solution Approach 1:
The patent applies denoising as a preliminary action before segmentation and surface determination. By cleaning the CT data first through the denoising network, subsequent segmentation and surface extraction operations become more accurate and efficient. This preliminary denoising step prevents noise from interfering with later processing stages, reducing overall processing time despite the added ML computation
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 enhances the accuracy and confidence of surface determination in CT scanning, enabling more precise metrology and reducing the need for costly hardware by effectively reducing noise and artefacts in CT voxel data.
Implementation Method 1
training an autoencoder using the patches
Implementation Method 2
emitting, by an x-ray source, x-rays towards an object; capturing, by a detector, the x-rays, wherein at least one of the x-rays passes through the object prior to being captured
Data Source
AI summary
Examples described herein provide a method that includes creating two-dimensional (2D) slices from a plurality of computed tomography (CT) voxel data sets. The method further includes adding artificial noise to the 2D slices to generate artificially noisy 2D slices. The method further includes creating patches from the 2D slices and the artificially noisy 2D slices. The method further includes training an autoencoder using the patches.