KR20220144668A - Sparse-view ct reconstruction based on multi-level wavelet convolutional neural network - Google Patents

Abstract

A computed tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence is disclosed. The computerized tomography image reconstruction method for correcting artifacts in a computed tomography image of the present invention includes the steps of: acquiring a low-density CT image (sparse-view CT image); correcting and acquiring directional streak artifacts by inputting the sparse-view CT image to an artificial intelligence learning network of a multi-level wavelet-based U-Net structure; and removing the streak artifacts by subtracting the streak artifact image from the low-density imaging CT image (sparse-view CT image). A sparse-view method can remove linear patterns that occur when acquiring CT images, and thus it is possible to enhance diagnostic efficiency.

Description

A computerized tomography image reconstruction method for multi-level wavelet AI-based dose reduction {SPARSE-VIEW CT RECONSTRUCTION BASED ON MULTI-LEVEL WAVELET CONVOLUTIONAL NEURAL NETWORK}

The present invention relates to a computerized tomography image reconstruction method, and more particularly, after learning using an artificial intelligence learning network of a multi-level wavelet-based U-Net structure, linearity patterns such as line artifacts are removed. A computerized tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence that reconstructs sparse-view CT images.

In modern medicine, the use of CT has rapidly increased not only for accurate diagnosis of diseases, but also for improving efficiency during treatment or determining prognosis. In particular, thanks to the development of flat-panel X-ray detector technology, the application of the so-called cone-beam CT system has significantly increased.

Computed tomography (CT) is a useful medical imaging technique for non-invasively acquiring high-resolution cross-sectional images of the human body and observing internal tissues of the human body, such as bones, organs, and blood vessels.

Due to excessive CT scans or careless CT scan protocol management, it has been reported that there may be potential health risks due to the invasive nature of X-rays.

As a method of reducing the radiation dose, a low-dose irradiation method that physically reduces the X-ray radiation dose and a sparsely sampled method that reduces the number of X-ray radiation have been proposed. In addition, various studies have been conducted to reduce aliasing artifacts added to images reconstructed from sparsely sampled CT data.

Sparse-view CT is the simplest method to reduce the radiation exposure frequency and thereby lower the exposure dose to the patient.

However, if the exposure frequency is reduced, information on the amount of radiation transmitted to the patient is insufficient, and thus information loss occurs in the signal unit for receiving the image. Such information loss may cause a major problem in diagnosis or reading of an image, and may also lead to image artifacts having directionality when an image is reconstructed with insufficient signals and information.

That is, an insufficient projection view in the sparse view CT generates streaking artifacts that are severe enough to affect the diagnosis in the filtered-backprojection (FBP) reconstruction.

There is a need for a specific method for removing the artifacts generated by this and correcting the lost information.

To remove these image artifacts, interpolation and advanced image reconstruction methods (eg, adaptive-steepest-descent-projection-onto-convex-sets (ASD-POCS), Compressed sensing (CS)) are used, but complex The process of accurately finding and correcting an artifact having a structure is not easy, and there is a limit in restoring an accurate image.

In addition, since it is an iterative reconstruction method, there is a problem in that the time for image reconstruction is relatively large.

KR Patent Publication No. 10-2019-0102770 (2019.09.04)

The present invention to solve this problem is a computerized tomography image reconstruction method for dose reduction based on artificial intelligence based on multi-level wavelets that can more accurately and conveniently remove streaking artifacts of an image by using an artificial neural network. aims to provide

In addition, the present invention uses an artificial intelligence learning network of a multi-level wavelet-based U-Net structure that downsamples images acquired at high doses from sinograms (Full projections) and reconstructs sparse-view CT images to reduce the dose. Another object of the present invention is to provide a computerized tomography image reconstruction method for impression.

In addition, another object of the present invention is to provide a computerized tomography image reconstruction method for dose reduction based on multi-level wavelet AI, which has high temporal efficiency at 0.4 to 0.5 seconds per slice.

In addition, the present invention utilizes DWT and IWT, trains an image through an artificial neural network, corrects artifacts, uses the information of the original image, and can maintain resolution based on multi-level wavelet AI for dose reduction. It is another object to provide a computerized tomography image reconstruction method.

A computerized tomography image reconstruction method for correcting artifacts in a computed tomography image according to an embodiment of the present invention for solving these problems is (a) a low-density CT image (sparse-view CT image) acquiring, (b) inputting the sparse-view CT image into an artificial intelligence learning network of a multi-level wavelet-based U-Net structure to correct directional streaking artifacts to acquire, and (c) ) It can be achieved by configuring including the step of removing the line artefact by subtracting the line artefact image from the low-density imaging CT image (sparse-view CT image).

In addition, step (b) includes the steps of (b-1) down-sampling the CT image obtained at a high dose from a sinogram (full projection) and reconstructing it into a low-density CT image (sparse-view CT image), (b) -2) obtaining linearity patterns (streaking artifacts) that are directional images by subtracting the sparse-view CT image obtained in step (b-1) from the original image, and (b-3) the ( Using the image obtained in step b-2) as an output value, it may be configured to include the step of extracting the line artefact image according to the condition learned through the artificial intelligence learning network of the multi-level wavelet-based U-Net structure.

In addition, the artificial intelligence learning network of the multi-level wavelet-based U-Net structure removes noise from the linearity pattern (streaking artifacts) generated in step (b) with a convolutional neural network (CNN) and discrete wavelet transform (DWT; discrete). It is configured to be learned using wavelet transform) and inverse wavelet transform (IWT).

In addition, the multi-level wavelet-based U-Net structure of the artificial intelligence learning network can learn features of streaking artifacts appearing in sparse-view CT images using CNN, DWT, and IWT.

In addition, the multi-level wavelet-based U-Net structure of the artificial intelligence learning network generates the learning model that satisfies the frame condition through a convolution framelet-based mathematical analysis, and by the learning model It may include a trained neural network, and may include a multi-resolution neural network including a pooling layer and an unpooling layer.

Therefore, according to the computerized tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence of the present invention, the linear pattern that occurs when acquiring using the sparse-view method to achieve a low dose to a patient using an artificial neural network can be removed to increase the diagnostic efficiency.

In addition, according to the computer tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence of the present invention, a phenomenon caused by insufficient correction or malfunction occurring in a photon count-based detector is corrected by using an artificial neural network. It has the effect of improving the picture quality of

In addition, according to the computer tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence of the present invention, it is possible to not only remove linear artifacts using an artificial neural network but also to remove artifacts generated in various diagnostic devices. can

In addition, according to the computer tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence of the present invention, since an artificial neural network is used, a computed tomography image with a faster scan time than before is used to not only the chest but also the knee with a large area. It has an effect that can be used on other body parts such as joints, bones of the back of the hand, sinuses, and mammography.

In addition, according to the computerized tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence of the present invention, since it is a computerized tomography image system using an artificial neural network, it can be used as a database for entering the medical market, and can be used as a database for domestic medical treatment. It has the effect of contributing to the industrialization and technological development of devices.

In addition, according to the computer tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence of the present invention, the calculation time is about 60 to 100 seconds/slice in the case of using a GPU in the iterative reconstruction method to remove the existing artefacts On the other hand, the present invention achieves about 0.4 to 0.5 sec/slice, so that it can have high efficiency in time.

And, according to the computer tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence of the present invention, the existing linearity pattern removal method replaces the original value with the interpolated value of the surrounding pixel values, so information on the original image Although loss may occur, the present invention utilizes DWT and IWT compared to the existing method, and as the latest medical imaging technology, the image is trained through an artificial neural network to correct the artifact, and the information of the original image is used and the resolution is maintained. has the effect of making

1 is a diagram illustrating the framework of the Sparse-view CT reconstruction method of the present invention;
2 is a reference diagram for explaining the step of reconstructing a sparse-view CT image using full projections;
3 is a view related to the process of learning features of linear artifacts appearing in sparse-view CT images;
4 is a diagram illustrating an artifact feature extraction process using a 4-level wavelet;
5 is a view for explaining a method of obtaining an output value from an original image;
6 is a diagram illustrating an artificial intelligence architecture of a multi-level wavelet-based U-Net structure;
7 is a flowchart for explaining a computerized tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence;
8 is a block diagram of a computerized tomography image reconstruction apparatus for multi-level wavelet artificial intelligence-based dose reduction;
and
9 is a diagram illustrating a sparse-view CT image corrected by various methods.

The terms or words used in the present specification and claims are not to be construed as limited in their ordinary or dictionary meanings, and on the principle that the inventor can appropriately define the concept of the term in order to best describe his invention. It should be interpreted as meaning and concept consistent with the technical idea of the present invention.

Throughout the specification, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated. In addition, terms such as “…unit”, “…group”, “module”, and “device” described in the specification mean a unit that processes at least one function or operation, which is implemented by a combination of hardware and/or software can be

Throughout the specification, the term “and/or” should be understood to include all combinations possible from one or more related items. For example, the meaning of “the first item, the second item and/or the third item” means to be presented from the first, second, or third item as well as two or more of the first, second, or third items. A combination of all possible items.

In each step throughout the specification, identification symbols (eg, a, b, c, ...) are used for convenience of description, and identification codes do not limit the order of each step, and each step is Unless the context clearly dictates a specific order, the order may differ from the stated order. That is, each step may occur in the same order as specified, may be performed substantially simultaneously, or may be performed in the reverse order.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

First, according to the present invention, the sparse-view CT image can reduce the amount of radiation reaching the patient by reducing the frequency of radiation exposure, and has the advantage of reducing the image acquisition time. Because streaking artifacts can cause problems in diagnosis, one feature is to learn it with artificial intelligence and then acquire a reconstructed image with line artifacts removed from the sparse-view CT image.

7 is a flowchart for explaining a computer tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence. Referring to the drawings, a straight line of a computed tomography image according to an embodiment of the present invention A computerized tomography image reconstruction method for correcting a sexual pattern includes: acquiring a low-density CT image (sparse-view CT image) (S110); The linear pattern is obtained by correcting and obtaining directional streaking artifacts by input to the intelligent learning network (S120, S130) and subtracting the linearity pattern from the low-density CT image (sparse-view CT image). and acquiring the removed image (S140).

Step S110 acquires a low-density CT image (hereinafter, sparse-view CT image) for correcting a linear pattern such as a line artifact in a computed tomography image.

Sparse-view CT is the simplest method to reduce the radiation exposure frequency and thereby lower the exposure dose to the patient.

However, if the exposure frequency is reduced, information on the amount of radiation transmitted to the patient is insufficient, and information loss occurs in the signal unit that receives the image. This loss of information may cause a great problem in diagnosing or reading an image, and also, a linear pattern having directionality may occur when an image is reconstructed with insufficient signals and information.

Therefore, the sparse-view CT obtained in step 110 is input to the artificial intelligence learning network to correct the linearity pattern (S120).

In S120, it is input to the artificial intelligence learning network of a multi-level wavelet-based U-Net structure in order to extract the linearity patterns (streaking artifacts) in step S130.

The artificial intelligence learning network of the present invention is trained with an artificial neural network to extract linear patterns (streaking artifacts) from sparse-view CT.

An artificial intelligence learning method will be described with reference to the drawings.

1 is a framework of a sparse-view CT reconstruction method based on the present invention. First, a sparse-view CT image is reconstructed by down-sampling an image obtained at a high dose in sinograms (full projections).

This reconstructed image uses Convolutional Neural Network (CNN), Discrete wavelet transform (DWT), and Inverse wavelet transform (IWT) to learn features of linear patterns appearing in sparse-view CT images, and to learn the features of residual learning. It is possible to easily extract the linearity pattern for sparse-view CT data by using the neural network of the learning model learned by

Referring to the reference diagram for explaining the step of reconstructing the sparse-view CT image using the full projections of FIG. 2 , the CT image is sampled from the sinogram (full projections) using the full projections, and the sampled CT image is inversely filtered. Reconstruction using projected Filtered BackProjection (FBP) is shown.

The CT image reconstructed as described above proceeds with the artifact feature extraction process.

4 is a diagram illustrating an artifact feature extraction process using a 4-level wavelet, and FIG. 3 is a diagram related to a process of learning features of a linear artifact appearing in a sparse-view CT image.

Referring to the drawings, it relates to a process of extracting a feature of a linear pattern, which is a line artifact, using 4-level wavelet transformation. Convolutional Neural Network (CNN), Discrete wavelet transform (DWT) ) and inverse wavelet transform (IWT) to learn the features of linear artifacts appearing in sparse-view CT images. It is suitable for extracting features for

A convolutional neural network (CNN) technique is known as an artificial neural network modeling technique for modeling a neural network, and is particularly useful for recognizing and estimating patterns of two-dimensional input data such as images.

Discrete wavelet transform (DWT) is performed on the original image to separate it into quadruple images, and inverse wavelet transform (IWT) is performed on the separated image to restore the original image.

Referring to FIG. 3, if noise is removed from the reconstructed sparse-view CT image through a convolutional neural network (CNN), and the noise-removed 512ch original image is subjected to step 1 DWT, 4 As a screen, in step 2 DWT, it becomes 128ch 16 screens, in step 3, 64 screens in all directions, and in step 4, it becomes 256 screens of 32ch all around, and the linearity pattern features are extracted.

That is, if the sparse-view CT image is excluded from the original image, streaking artifacts are obtained as residuals, and it is learned as an output value and is easily used for AI learning convergence by using the directional feature ( see Fig. 5).

The streaking artifacts created in this way are improved by training the image through an artificial neural network of a multi-level wavelet-based U-Net structure.

Referring to the diagram illustrating the artificial intelligence architecture of the multi-level wavelet-based U-Net structure of FIG. 6, using a sparse-view CT image having a low dose as an input layer, using the built-in optimization algorithm (ADAM optimization algorithm) Using DWT and IWT for up and down sampling, an image with streaking artifacts is output to the output layer, and the resulting output image is subtracted from the image entered in the input layer to create an image with artifacts removed. there will be

That is, the artificial intelligence learning network of the multi-level wavelet-based U-Net structure of the present invention learns the characteristics of the streaking artifacts appearing in the sparse-view CT image using CNN, DWT, and IWT.

In other words, the artificial intelligence learning network of the multi-level wavelet-based U-Net structure of the present invention generates the learning model that satisfies the frame condition through a convolution framelet-based mathematical analysis, and A neural network trained by the model is included.

As shown in FIG. 6 , each network includes a convolution layer that performs a linear transform operation, a batch normalization layer that performs a normalization operation, and a nonlinear function operation. It includes a path connection with a rectified linear unit (ReLU) layer and concatenation that performs

In particular, each stage contains 4 sequential layers consisting of convolution with 4x4 kernels, batch normalization and ReLU layers. The last stage contains two sequential layers and the last layer, and the last layer contains only the convolution with 1 × 1 kernel. The number of channels for each convolutional layer is shown in Fig. 6, and the number of channels is doubled after each pooling layer.

The artificial intelligence learning network of the multi-level wavelet-based U-Net structure used for extracting linearity patterns such as line artifacts in the device of the present invention uses a convolution framelet-based mathematical analysis to determine the frame conditions. A satisfactory learning model is generated, and a neural network trained by the learning model may be included, and a multi-resolution neural network including a pooling layer and an unpooling layer may be included.

In addition, the neural network may include a bypass connection from the pooling layer to the unpooling layer.

Referring to the drawings exemplifying the sparse-view CT images corrected by various methods in FIG. 9 , images corrected by various methods are illustrated in the parse-view CT images reconstructed using 60 original images.

From the left image, reference image (Reference), 60view, interpolation, progressive image reconstruction method (eg, adaptive-steepest-descent-projection-onto-convex-sets (ASD-POCS), U-net, of the present invention The corrected image by MWCNN is illustrated.

8 is a diagram illustrating the configuration of a computerized tomography image reconstruction apparatus for dose reduction based on multi-level wavelet artificial intelligence according to an embodiment of the present invention.

Referring to the drawings, the computer tomography image reconstruction apparatus 100 according to an embodiment of the present invention includes a receiving unit 110 , an image processing unit 120 , and an output unit 130 for outputting an image from which the linear pattern is removed. do.

The receiver 110 receives sparse-view computed tomography data.

The image processing unit 120 satisfies a preset frame condition and uses a neural network for a convolution framelet-based learning model to generate a linear pattern for sparse-view CT imaging data. It is to reconstruct the image from which the linearity pattern has been removed by finding it and subtracting it from the original image.

This is because it is easy to converge AI learning because the linear pattern has a direction.

To this end, the image processing unit 120 learns to extract features for a linear pattern through an artificial intelligence learning network of a multi-level wavelet-based U-Net structure.

That is, the multi-level wavelet-based U-Net structure of artificial intelligence learning network downsamples the high-dose images from sinograms (full projections) and reconstructs them into low-density CT images (sparse-view CT images), By subtracting the reconstructed sparse-view CT image from the original image, it acquires directional patterns (streaking artifacts), and learns the acquired image as an output value using the characteristics of the directional linear pattern. will do

For the specific method, refer to the above-mentioned contents.

As a result, the image processing unit 120 satisfies the frame condition and can easily extract a linearity pattern for the sparse view computed tomography data by using the neural network of the learning model learned by residual learning.

Although the description of the device of FIG. 8 is omitted, each component constituting FIG. 8 may include all the contents described with reference to FIGS. 1 to 6 , which is apparent to those skilled in the art.

Although the present invention has been described in detail with respect to the described embodiments, it is obvious to those skilled in the art that various changes and modifications are possible within the scope of the technical spirit of the present invention, and it is natural that such variations and modifications belong to the appended claims.

100: computer tomography image reconstruction device
110: receiver
120: image processing unit
130: output unit

Claims (10)

A method for reconstructing a computed tomography image for correcting artifacts in a computed tomography image, the method comprising:
(a) obtaining a low-density CT image (sparse-view CT image);
(b) inputting the sparse-view CT image into an artificial intelligence learning network of a multi-level wavelet-based U-Net structure to correct and obtain directional streak artifacts; And
(c) removing the line artefact by subtracting the line artefact image from the low-density imaging CT image (sparse-view CT image);
A computerized tomography image reconstruction method for dose reduction based on artificial intelligence-based multi-level wavelet comprising a.
The method according to claim 1,
Step (b) is
(b-1) down-sampling the CT image acquired at high dose from a sinogram (full projection) and reconstructing it into a low-density CT image (sparse-view CT image);
(b-2) subtracting the sparse-view CT image obtained in step (b-1) from the original image to obtain streak artifacts, which are directional images; and
(b-3) extracting a line artefact image according to the conditions learned through an artificial intelligence learning network of a multi-level wavelet-based U-Net structure using the image obtained in step (b-2) as an output value;
A computerized tomography image reconstruction method for dose reduction based on artificial intelligence-based multi-level wavelet comprising a.
The method according to claim 1,
Step (c) is
After filtering the sparse-view CT image, the image is reconstructed using filtered backprojection (FBP), and the resulting streak artifacts are subtracted from the original image. Multi-level wavelet AI-based dose reduction A computerized tomography image reconstruction method for
The method according to claim 1,
The multi-level wavelet-based U-Net structure of artificial intelligence learning network is
Using a convolutional neural network (CNN), discrete wavelet transform (DWT), and inverse wavelet transform (IWT) to remove noise from the streaking artifacts generated in step (b) Multi-level wavelet artificial intelligence-based computerized tomography image reconstruction method for dose reduction.
The method according to claim 1,
The multi-level wavelet-based U-Net structure of artificial intelligence learning network is
A computerized tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence that learns the characteristics of streaking artifacts in sparse-view CT images using CNN, DWT and IWT.
The method according to claim 1,
The multi-level wavelet-based U-Net structure of artificial intelligence learning network is
Multi-level wavelet artificial intelligence based, characterized in that it generates the learning model that satisfies the frame condition through a convolution framelet-based mathematical analysis, and includes a neural network trained by the learning model A computerized tomography image reconstruction method for dose reduction.
The method according to claim 1,
The multi-level wavelet-based U-Net structure of artificial intelligence learning network is
A computer tomography image reconstruction method for multi-level wavelet AI-based dose reduction, comprising a multi-resolution neural network including a pooling layer and an unpooling layer.
The method according to claim 1,
The multi-level wavelet-based U-Net structure of artificial intelligence learning network is
A computerized tomography image reconstruction method for dose reduction based on multi-level wavelet AI that outputs streaking artifacts using 4-level wavelet transformation.
The method according to claim 1,
The multi-level wavelet-based U-Net structure of artificial intelligence learning network is
Image with sparse-view CT image with low dose as input layer and linearity pattern (streaking artifacts) as output layer using DWT and IWT with up and down sampling using the built-in optimization algorithm (ADAM optimization algorithm) A computerized tomography image reconstruction method for dose reduction based on AI-based multi-level wavelet that outputs
10. The method of claim 9,
A computerized tomography image reconstruction method for dose reduction based on multi-level wavelet artificial intelligence that generates and outputs an image from which artifacts are removed by subtracting the image having the linearity pattern (streaking artifacts) from the image that has entered the input layer.

Images