Geometric Artifact Evaluation Method for CT Images Based on Residual Network

Abstract

The present invention provides a method for evaluating geometric artifacts of CT images based on a residual network. The method comprises: step 1: based on the characteristics of CT images, a matching training sample data set is prepared; step 2: using the residual network Resnet50 as the basic network, a geometric artifact evaluation network is obtained by reducing the convolution residual block step size and adding an attention module design; step 3: using the training sample data set to train the geometric artifact evaluation network; step 4: inputting the CT image to be evaluated into the trained geometric artifact evaluation network to obtain its geometric artifact level. The present invention can evaluate the degree of geometric artifacts of CT images with high accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of CT image processing, and in particular to a CT image geometric artifact evaluation method based on a residual network. Background Art

[0002] The deviation between the actual system geometric parameters and the ideal geometric parameters leads to geometric artifacts in the CT reconstructed images. Geometric artifacts seriously affect the quality of CT reconstructed images, resulting in blurred images and ghosting at the edges. It is one of the key factors hindering the development of CT towards intelligence and high precision. Geometric artifact correction is a prerequisite for the CT system to obtain high-quality three-dimensional reconstructed images. The performance of the geometric artifact correction method determines the accuracy of the CT image. Evaluating the severity of image geometric artifacts can, on the one hand, characterize the degree of mismatch of the system geometric parameters and be used for CT system quality control. On the other hand, it can fairly and objectively evaluate the accuracy of the geometric artifact correction method and screen the geometric artifact correction method. Summary of the invention

[0003] In order to characterize the degree of mismatch of system geometric parameters for CT system quality control, the present invention provides a CT image geometric artifact evaluation method based on a residual network, which can use a deep network to quickly and accurately complete the CT image geometric artifact evaluation. Geometric artifact evaluation is to classify images according to the degree of artifacts in the image. Its essence is a multi-classification problem. The present invention is a CT image geometric artifact evaluation method based on a residual network. By constructing a large-scale data sample, the degree of geometric artifacts is divided into three levels: no, slight and severe geometric artifacts. Then, a reasonable deep residual network model is designed to achieve accurate grading of the degree of geometric artifacts in CT reconstructed images.

[0004] The present invention provides a method for evaluating CT image geometric artifacts based on a residual network, comprising the following steps:

[0005] Step 1: Create a matching training sample dataset based on CT image characteristics;

[0006] Step 2: Using the residual network Resnet50 as the basic network, the geometric artifact evaluation network is designed by reducing the convolution residual block step size and adding an attention module;

[0007] Step 3: Using the training sample data set to train the geometric artifact evaluation network;

[0008] Step 4: Input the CT image to be evaluated into the trained geometric artifact evaluation network to obtain its geometric artifact level.

[0009] Furthermore, the step 1 specifically includes:

[0010] By changing Δu0, CT images with different degrees of geometric artifacts are reconstructed; where u0 is the projection abscissa of the X-ray source on the flat panel detector, and Δu0 is the deviation value of u0;

[0011] The CT images with different degrees of geometric artifacts are divided into three categories: no geometric artifacts, slight geometric artifacts and severe geometric artifacts, and category labels are set for the three categories of images respectively;

[0012] Each CT image and its category label constitute a training sample, and all training samples constitute a training sample dataset.

[0013] Furthermore, the step 2 specifically includes:

[0014] Using the residual network Resnet50 as the base network, the step size of the second convolution residual block in the original residual network Resnet50 is reduced from 2 to 1;

[0015] An attention module is added to each convolution residual block. The attention module contains two independent sub-modules: a channel attention module and a spatial attention module. The two sub-modules are used for channel and spatial attention, respectively.

[0016] In the geometric artifact evaluation network, the intermediate feature map F is given as the input of the attention module, and the attention module calculates the one-dimensional channel attention map M in turn. c (F) and the 2D spatial attention map M s (F'); connect the two attention maps in series for overall attention and obtain the output feature map F", which is specifically Among them, F' is the intermediate operation feature map, is element-wise multiplication.

[0017] Furthermore, the channel attention module uses global average pooling and global maximum pooling to aggregate the spatial information of the input feature map F to obtain feature maps and feature map Then the feature map and feature map Send it to the multi-layer perceptron to get the intermediate operation feature map and Then the feature map and conduct After the operation, the sigmoid function outputs a one-dimensional channel attention map M c (F); wherein, Represents element-wise summation.

[0018] Furthermore, the spatial attention module generates the global average pooling features and the global maximum pooling features of the input feature map F' by using the global average pooling and the global maximum pooling on the channel axis, and then performs channel splicing on the global average pooling features and the global maximum pooling features, and then sequentially passes through a 7×7 convolution layer and a sigmoid function to generate a two-dimensional spatial attention map M s (F').

[0019] Beneficial effects of the present invention:

[0020] The CT image geometric artifact evaluation method based on residual network provided by the present invention first innovatively constructs a data set containing different degrees of geometric artifacts by changing sensitive geometric parameters, and then designs a CT image geometric artifact evaluation network. The geometric artifact evaluation network uses the Resnet50 network as the basic structure, and designs the geometric artifact evaluation network by reducing the step size of the convolution residual block and adding an attention module. The residual block structure is used to extract higher-dimensional image features; an attention module is added to the convolution residual module to mine channel-level and spatial-dimensional features, further focus on image edge features, and improve the hierarchical evaluation effect of the geometric artifact evaluation network. The trained geometric artifact evaluation network can evaluate the degree of CT image artifacts with high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A schematic diagram of a flow chart of a method for evaluating geometric artifacts in CT images based on a residual network provided in an embodiment of the present invention;

[0022] Figure 2 The overall structure diagram of the geometric artifact evaluation network provided by the embodiment of the present invention;

[0023] Figure 3 A structural diagram of a Conv Block residual block in a geometric artifact evaluation network provided by an embodiment of the present invention;

[0024] Figure 4 A structural diagram of an ID Block residual block in a geometric artifact evaluation network provided by an embodiment of the present invention;

[0025] Figure 5 Attention module structure diagram provided for an embodiment of the present invention

[0026] Figure 6 A structural diagram of a channel attention module provided in an embodiment of the present invention;

[0027] Figure 7 A structural diagram of a spatial attention module provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] like Figure 1 As shown, an embodiment of the present invention provides a method for evaluating geometric artifacts of CT images based on a residual network, comprising the following steps:

[0030] S101: Based on the characteristics of CT images, a matching training sample data set is produced;

[0031] Specifically, the magnitude of the system geometric parameter deviation reflects the severity of the geometric artifacts of the reconstructed image. In order to obtain geometric artifacts of different severity, the embodiment of the present invention selects the most sensitive parameter u0 among the geometric parameters, and generates images with different degrees of geometric artifacts by changing the magnitude of its deviation value Δu0. u0 is the horizontal coordinate of the projection of the X-ray source on the flat panel detector, and the deviation of its value has the greatest impact on the reconstructed image.

[0032] Theoretically, the larger Δu0 is, the more severe the geometric artifacts are, the lower the resolution of the reconstructed image is, and the more information is lost. The performance of the trained network is closely related to the training data, and high-quality data is crucial for network optimization. In order to obtain high-quality geometric artifact images with different degrees of severity, the present invention uses the structural similarity (Structure SIMilarity, SSIM) indicator to evaluate the similarity of two images to further verify that as Δu0 increases, the severity of the geometric artifacts of the reconstructed image increases. The SSIM indicator comprehensively measures the similarity of two images in terms of brightness, contrast and structure.

[0033] By comparing the SSIM of the image without geometric artifacts when Δu0=0 and the image with geometric artifacts when Δu0=0.01mm, 0.02mm, 0.03mm, 0.04mm, 0.05mm, 0.06mm, 0.07mm, and 0.08mm, it is found that as Δu0 increases, the SSIM value between the reconstructed image and the image without geometric artifacts gradually decreases, indicating that the image similarity decreases and the severity of geometric artifacts gradually increases. This proves that geometric artifact images with different severity levels can be generated by changing the value of Δu0. However, the SSIM values ​​calculated for the same Δu0 in different phantom data are different, indicating that the artifact levels of different phantom data are different under the same parameter deviation. At the same time, in each phantom data, the Δu0 corresponding to the same SSIM value at different slice positions is different, indicating that the parameter deviations of different phantom data are different when the artifact level is the same.

[0034] Therefore, when constructing a training sample data set, the embodiment of the present invention mainly includes the following sub-steps:

[0035] S1011: reconstructing a CT image with different degrees of geometric artifacts by changing Δu0; wherein u0 is the projection abscissa of the X-ray source on the flat panel detector, and Δu0 is the deviation value of u0;

[0036] S1012: combining human experience and the deviation of Δu0 to classify the CT images with different degrees of geometric artifacts into three categories: no geometric artifacts, slight geometric artifacts, and severe geometric artifacts, and setting category labels for the three categories of images;

[0037] Specifically, no geometric artifacts are manifested as clear images, and no geometric artifacts can be observed visually. Slight geometric artifacts correspond to certain deviations in geometric parameters, but the artifacts are not obvious visually. Severe geometric artifacts correspond to large deviations in geometric parameters of images, and the artifacts are obvious visually.

[0038] It should be noted that, in order to preserve the structural information of the CT image, each CT image is saved in matrix form when constructing the training sample data set.

[0039] S1013: Each CT image and its category label constitute a training sample, and all training samples constitute a training sample data set.

[0040] S102: Using the residual network Resnet50 as the basic network, a geometric artifact evaluation network is designed by reducing the convolution residual block step size and adding an attention module;

[0041] Specifically, reducing the convolution residual block step size can help the network extract more comprehensive features, and adding an attention module can enable the network to further explore the features of the channel level and spatial dimension, more effectively extract edge feature information in CT images, and improve the accuracy of artifact grading evaluation. The overall structure of the geometric artifact evaluation network designed in the embodiment of the present invention is as follows: Figure 2 shown.

[0042] As an implementable method, the structural design of the geometric artifact evaluation network includes the following processes:

[0043] Using the residual network Resnet50 as the base network, the step size of the second convolution residual block in the original residual network Resnet50 is reduced from 2 to 1;

[0044] An attention module is added to each convolution residual block, such as Figure 5 As shown, the attention module includes two independent sub-modules, namely: a channel attention module and a spatial attention module; the two sub-modules are used for channel and spatial attention respectively;

[0045] In the geometric artifact evaluation network, the intermediate feature map F is given as the input of the attention module, and the attention module calculates the one-dimensional channel attention map M in turn. c (F) and the 2D spatial attention map M s (F'); connect the two attention maps in series for overall attention and obtain the output feature map F", which is specifically Among them, F' is the intermediate operation feature map, is element-wise multiplication;

[0046] Among them, Figure 6 As shown, the channel attention module uses global average pooling and global maximum pooling to aggregate the spatial information of the input feature map F, and obtains the feature maps and feature map Then the feature map and feature map Send it to the Multi-Layer Perceptron (MLP) to get the intermediate operation feature map and Then the feature map and conduct After the operation, the sigmoid function outputs a one-dimensional channel attention map M c (F); wherein, It represents element-by-element summation;

[0047] like Figure 7As shown in FIG. 1 , the spatial attention module uses global average pooling and global maximum pooling on the channel axis to generate the global average pooling features and global maximum pooling features of the input feature map F', and then concatenates the global average pooling features and the global maximum pooling features through channels, and then sequentially passes through a 7×7 convolution layer and a sigmoid function to generate a two-dimensional spatial attention map M s (F'); wherein the feature map F' is obtained by a one-dimensional channel attention map M c (F) and the input feature map F After the operation, Represents element-wise multiplication.

[0048] As an implementation method, the structure of the Conv Block residual block (also called convolution residual block) in the geometric artifact evaluation network is as follows: Figure 3 As shown, Figure 3 The “CBAM” in the figure is the added attention module.

[0049] As an implementation method, the structure of the ID Block residual block in the geometric artifact evaluation network is as follows: Figure 4 shown.

[0050] It should be noted that since the constructed training sample data set is saved in matrix form, the data set size is relatively large (can reach more than 10G). The geometric artifact evaluation network constructed in the embodiment of the present invention fully considers the size of the data set, the selected network depth is moderate, and the evaluation effect is better.

[0051] S103: training the geometric artifact evaluation network using the training sample data set;

[0052] Specifically, training and testing can be done under the Tensorflow deep learning framework on an AMAX workstation, which has two CPUs, Intel Xeon E5-2640 v4, 64GB of memory, and four GPUs, Geforce GTX 1080Ti, 11GB of video memory.

[0053] S104: Input the CT image to be evaluated into the trained geometric artifact evaluation network to obtain its geometric artifact level.

[0054] The CT image geometric artifact evaluation method based on residual network provided by the present invention first innovatively constructs a data set containing different degrees of geometric artifacts by changing sensitive geometric parameters, and then designs a CT image geometric artifact evaluation network. The geometric artifact evaluation network uses the Resnet50 network as the basic structure, and designs the geometric artifact evaluation network by reducing the step size of the convolution residual block and adding an attention module. The residual block structure is used to extract higher-dimensional image features; an attention module is added to the convolution residual module to mine channel-level and spatial-dimensional features, further focus on image edge features, and improve the hierarchical evaluation effect of the geometric artifact evaluation network. The trained geometric artifact evaluation network can evaluate the degree of CT image artifacts with high accuracy.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating geometric artifacts in CT images based on residual networks, characterized in that: include: Step 1: Create a matching training sample dataset based on CT image characteristics; Specifically include: By changing Δu0, CT images with different degrees of geometric artifacts are reconstructed; where u0 is the projection abscissa of the X-ray source on the flat panel detector, and Δu0 is the deviation value of u0; The CT images with different degrees of geometric artifacts are divided into three categories: no geometric artifacts, slight geometric artifacts and severe geometric artifacts, and category labels are set for the three categories of images respectively; Each CT image and its category label constitute a training sample, and all training samples constitute a training sample dataset; Step 2: Using the residual network Resnet50 as the basic network, the geometric artifact evaluation network is obtained by reducing the step size of the convolution residual block and adding an attention module; specifically, the residual network Resnet50 is used as the basic network, and the step size of the second convolution residual block in the original residual network Resnet50 is reduced from 2 to 1; An attention module is added to each convolution residual block. The attention module contains two independent submodules: a channel attention module and a spatial attention module. The two submodules are used for channel and spatial attention, respectively. In the geometric artifact evaluation network, the intermediate feature map F is given as the input of the attention module, and the attention module calculates the one-dimensional channel attention map M in turn. c (F) and the 2D spatial attention map M s (F'); connect the two attention maps in series for overall attention and obtain the output feature map F", specifically: Among them, F' is the intermediate operation feature map, is element-wise multiplication; Step 3: Using the training sample data set to train the geometric artifact evaluation network; Step 4: Input the CT image to be evaluated into the trained geometric artifact evaluation network to obtain its geometric artifact level.

2. The method for evaluating geometric artifacts of CT images based on residual networks according to claim 1, characterized in that: The channel attention module uses global average pooling and global maximum pooling to aggregate the spatial information of the input feature map F and obtain the feature maps and feature map Then the feature map and feature map Send it to the multi-layer perceptron to get the intermediate operation feature map and Then the feature map and conduct After the operation, the sigmoid function outputs a one-dimensional channel attention map M c (F); wherein, Represents element-wise summation.

3. The method for evaluating CT image geometric artifacts based on residual network according to claim 1 or 2, characterized in that: The spatial attention module uses global average pooling and global maximum pooling on the channel axis to generate the global average pooling features and global maximum pooling features of the input feature map F', and then concatenates the global average pooling features and the global maximum pooling features through channels, and then passes through a 7×7 convolution layer and a sigmoid function in sequence to generate a two-dimensional spatial attention map M s (F').

Images