Image reconstruction method, system, electronic device and readable storage medium
By extracting style features from images favored by target users and generating filter kernel functions using convolutional neural networks, the problem of cumbersome adjustment of filter kernel functions in existing technologies is solved, achieving fast and accurate image reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI UNITED IMAGING HEALTHCARE
- Filing Date
- 2022-10-27
- Publication Date
- 2026-06-05
AI Technical Summary
In existing CT reconstruction techniques, the adjustment of the filtering kernel function depends on the developer's experience, resulting in a large workload and difficulty in meeting different needs, making it difficult to quickly obtain image reconstructions that meet the needs of clinicians.
By extracting style features from images that reflect the preferences of target users, and using a trained convolutional neural network to generate a target filtering kernel function, image reconstruction can be performed directly, reducing the tedious debugging process.
It enables the rapid acquisition of filtering kernel functions that meet different needs, reducing time and manpower costs and improving the efficiency and accuracy of image reconstruction.
Smart Images
Figure CN115496827B_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of image reconstruction technology, and in particular to an image reconstruction method, system, electronic device, and readable storage medium. Background Technology
[0002] Existing CT reconstruction techniques mostly employ FBP (Filtered Back Projection) for image reconstruction. During the filtering process, the choice of the kernel function determines the stylistic characteristics of the generated tomographic image. Essentially, the kernel function is a curve composed of filter coefficients of different frequencies.
[0003] In practical applications, different hospitals, doctors, or scanning sites often require different filtering kernel functions. Currently, filtering kernel functions primarily rely on developers adjusting existing functions based on experience. The adjusted kernel is used to generate a batch of sample images for clinicians to review, and their feedback is collected. The process is then repeated until the reconstructed images based on the adjusted kernel function meet the clinicians' needs. This process not only requires developers to have extensive experience, but the workload of repeatedly adjusting, generating sample images, and collecting feedback is enormous, making it difficult to obtain filtering kernel functions that meet diverse requirements. Summary of the Invention
[0004] To address the aforementioned technical problems, this disclosure provides an image reconstruction method, system, electronic device, and readable storage medium. It only requires extracting target style features from images favored by the target user and inputting these features into a trained network model. Based on these target style features, a target filtering kernel function can be directly inferred, enabling image reconstruction of the corresponding style features. This eliminates the need for repeated, tedious debugging and allows for the rapid acquisition of filtering kernel functions that meet different needs, significantly reducing the time and manpower costs required to generate filtering kernel functions.
[0005] In a first aspect, this disclosure provides an image reconstruction method, the method comprising:
[0006] Acquire the target sample image;
[0007] The style features of the target sample image are extracted to obtain the target style features;
[0008] The target style features are input into the first convolutional neural network to generate the corresponding target filtering kernel function;
[0009] Based on the target filtering kernel function, image reconstruction is performed on the data to be reconstructed to obtain the reconstructed image.
[0010] Preferably, the step of extracting style features from the target sample image to obtain target style features includes:
[0011] A second convolutional neural network, comprising multiple convolutional kernels, is used to extract image features from the target sample image to obtain features from multiple channels.
[0012] The target style features are obtained using the Gram matrix based on the image features of the multiple channels.
[0013] Preferably, the first convolutional neural network is obtained by training a neural network model based on a training dataset. The training process involves backpropagating the error according to the target loss function and iteratively updating the model parameters of the neural network model until the target loss function is less than a preset threshold. The training dataset includes several sample images and a sample filtering kernel function for reconstructing the sample images.
[0014] Preferably, the target loss function is determined based on a first loss function, which is obtained based on a first filtering kernel function and a sample filtering kernel function. The first loss function is used to characterize image style differences; wherein, the first filtering kernel function is generated based on the style features of the sample images.
[0015] Preferably, the target loss function is determined based on the first loss function and the second loss function, wherein the second loss function is the difference between the first image noise and the second image noise, and the second loss function is used to characterize the difference in image noise; the first image noise is the atmospheric image noise of the sample image; the second image noise is the atmospheric image noise of the image obtained by image reconstruction based on the actual acquisition and reconstruction conditions and CT data corresponding to the sample image using the first filtering kernel function; wherein the atmospheric image noise is the image noise of the region in the image that does not include the imaging object.
[0016] Preferably, the second image noise is obtained by adjusting the reference image noise using an image noise variation coefficient; the reference image noise is used to characterize the air image noise of the image obtained by reconstructing the image based on the CT data corresponding to the sample image using the first filtering kernel function according to preset acquisition and reconstruction conditions; the image noise variation coefficient is determined based on the difference between the actual acquisition and reconstruction conditions and the preset acquisition and reconstruction conditions.
[0017] Preferably, the reference image noise is the sum of the sampling frequency noise contribution values; the sampling frequency noise contribution value is the product of a preset noise contribution value and the corresponding amplitude value; wherein, the preset noise contribution value is the contribution value of different frequency components to the image noise under the preset acquisition and reconstruction conditions, and the frequency components and the corresponding amplitude values are obtained by sampling the first filtering kernel function.
[0018] Secondly, this disclosure provides an image reconstruction system, the system comprising:
[0019] The sample image acquisition module is used to acquire target sample images;
[0020] The style feature acquisition module is used to extract the style features of the target sample image to obtain the target style features;
[0021] The objective function generation module is used to input the target style features into the first convolutional neural network to generate the corresponding target filtering kernel function;
[0022] The image reconstruction module is used to reconstruct the image based on the target filtering kernel function to obtain the reconstructed image.
[0023] Preferably, the style feature acquisition module includes:
[0024] An image feature extraction unit is used to extract image features from multiple channels based on the target sample image using a second convolutional neural network including multiple convolutional kernels;
[0025] A style feature generation unit is used to obtain the target style features based on the image features of the multiple channels using a Gram matrix.
[0026] Preferably, the image reconstruction system further includes a model training module for training a neural network model based on a training dataset to obtain the first convolutional neural network. The training process involves backpropagating the error according to the target loss function and iteratively updating the model parameters of the neural network model until the target loss function is less than a preset threshold; wherein, the training dataset includes several sample images and sample filtering kernel functions for reconstructing the sample images.
[0027] Preferably, the target loss function is determined based on a first loss function, which is obtained based on a first filtering kernel function and a sample filtering kernel function. The first loss function is used to characterize image style differences; wherein, the first filtering kernel function is generated based on the style features of the sample images.
[0028] Preferably, the target loss function is determined based on the first loss function and the second loss function, wherein the second loss function is the difference between the first image noise and the second image noise, and the second loss function is used to characterize the difference in image noise; the first image noise is the atmospheric image noise of the sample image; the second image noise is the atmospheric image noise of the image obtained by image reconstruction based on the actual acquisition and reconstruction conditions and CT data corresponding to the sample image using the first filtering kernel function; wherein the atmospheric image noise is the image noise of the region in the image that does not include the imaging object.
[0029] Preferably, the second image noise is obtained by adjusting the reference image noise using an image noise variation coefficient; the reference image noise is used to characterize the air image noise of the image obtained by reconstructing the image based on the CT data corresponding to the sample image using the first filtering kernel function according to preset acquisition and reconstruction conditions; the image noise variation coefficient is determined based on the difference between the actual acquisition and reconstruction conditions and the preset acquisition and reconstruction conditions.
[0030] Preferably, the reference image noise is the sum of the sampling frequency noise contribution values; the sampling frequency noise contribution value is the product of a preset noise contribution value and the corresponding amplitude value; wherein, the preset noise contribution value is the contribution value of different frequency components to the image noise under the preset acquisition and reconstruction conditions, and the frequency components and the corresponding amplitude values are obtained by sampling the first filtering kernel function.
[0031] Thirdly, this disclosure provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the image reconstruction method described in the first aspect and any of its embodiments.
[0032] Fourthly, this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the image reconstruction method described in the first aspect and any of its embodiments.
[0033] The positive and progressive effects of this disclosure are as follows: The image reconstruction method, system, motor device, and readable storage medium provided by this disclosure only require extracting the target style features from the image preferred by the target user, inputting the target style features into the trained network model, and then directly inferring the target filtering kernel function based on the target style features, thereby performing image reconstruction with the corresponding style features. This eliminates the need for repeated and tedious debugging, and allows for the rapid acquisition of filtering kernel functions that meet different needs, greatly reducing the time and manpower costs required to generate filtering kernel functions. Attached Figure Description
[0034] Figure 1 This is a flowchart of the image reconstruction method in Embodiment 1 of this disclosure;
[0035] Figure 2 This is a schematic diagram of the process for obtaining the target loss function in Embodiment 1 of this disclosure;
[0036] Figure 3 This is a schematic diagram illustrating the principle of the image reconstruction method in Embodiment 1 of this disclosure;
[0037] Figure 4 This is a schematic diagram of the modules of the image reconstruction system in Embodiment 2 of this disclosure;
[0038] Figure 5 This is a schematic diagram of the structure of the electronic device in Embodiment 3 of this disclosure. Detailed Implementation
[0039] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.
[0040] Example 1
[0041] This embodiment provides an image reconstruction method.
[0042] like Figure 1 As shown, the image reconstruction method includes steps S1-S4:
[0043] Step S1: Obtain the target sample image;
[0044] Step S2: Extract the style features of the target sample image to obtain the target style features;
[0045] Step S3: Input the target style features into the first convolutional neural network to generate the corresponding target filtering kernel function;
[0046] Step S4: Reconstruct the image based on the target filtering kernel function to obtain the reconstructed image.
[0047] In step S1, a batch of CT images in the style desired by the target user are collected as target sample images.
[0048] In step S2, the style features of the target sample image are extracted as the target style features.
[0049] Style features are used to characterize the color, brightness, noise, texture, and other features of CT images.
[0050] In one feasible implementation, step S2 includes:
[0051] Step S21: Use a second convolutional neural network including multiple convolutional kernels to extract image features from multiple channels based on the target sample image;
[0052] Step S22: Use the Gram matrix to obtain target style features based on image features from multiple channels.
[0053] Specifically, the target sample image is input into a second convolutional neural network, which includes multiple convolutional kernels. Different convolutional kernels are used to output image features for different channels. That is, the target sample image is processed by different convolutional kernels to extract image features for multiple channels. These image features for multiple channels are then represented using a Gram matrix to obtain the target style features.
[0054] For example, the target sample image has a size of n*n, and the second convolutional neural network includes c convolutional kernels k, each kernel k having a size of m*m. The target sample image is processed by the second convolutional neural network to obtain image features I across multiple channels, which are calculated as follows:
[0055] I c,j,j =I n,n *k c,m,m ;
[0056] The target style feature IG is obtained by representing the image features of c channels using the Gram matrix, and its calculation method is as follows:
[0057]
[0058] In step S3, the first convolutional neural network infers based on the target style features to obtain the target filtering kernel function.
[0059] In one feasible implementation, the first convolutional neural network is obtained by training a neural network model based on a training dataset. The training dataset includes several sample images and sample filtering kernel functions for reconstructing the sample images.
[0060] The training process involves backpropagating the error based on the target loss function and iteratively updating the model parameters of the neural network until the target loss function is less than a preset threshold.
[0061] In one feasible implementation, the first convolutional neural network includes a series of convolutional layers. In another feasible implementation, the first convolutional neural network also includes a residual module.
[0062] In one feasible implementation, the target loss function is determined based on the first loss function.
[0063] Specifically, a first filtering kernel function is first generated based on the style features of the sample images. Then, a first loss function representing the style differences of the images is obtained based on the first filtering kernel function and the sample filtering kernel function. The first loss function is used as the target loss function.
[0064] For example, the first filtering kernel function generated based on the style features of the sample images is Kernel1, and the sample filtering kernel function is Kernel0. The first loss function L1 is the mean square error between the first filtering kernel function and the sample filtering kernel function, and its specific calculation method is as follows:
[0065]
[0066] In another feasible implementation, such as Figure 2 As shown, the target loss function is determined based on the first loss function and the second loss function.
[0067] Specifically, the first image noise is obtained by acquiring the atmospheric noise of the sample image. Then, the first filtering kernel function is used to reconstruct the image based on the actual acquisition and reconstruction conditions and CT data corresponding to the sample image. The atmospheric noise of the reconstructed image is used as the second image noise. Finally, the difference between the first image noise and the second image noise is obtained to obtain the second loss function. Here, atmospheric noise refers to the image noise in the region of the image that does not include the imaging object.
[0068] For example, if n square regions I are selected where the image edge does not include the imaged object, then the image noise in each square region is I. air The corresponding air image noise is for all square regions I air The image noise is the square root of the sample variance. Therefore, the first image noise is calculated as follows:
[0069]
[0070] In one feasible implementation, the second image noise is obtained by adjusting the reference image noise using an image noise variation coefficient.
[0071] Among them, the reference image noise is used to characterize the air image noise of the image obtained by image reconstruction based on the CT data corresponding to the sample image using the first filtering kernel function according to the preset acquisition and reconstruction conditions.
[0072] Furthermore, due to the differences between the actual acquisition parameters and the preset acquisition parameters, as well as the differences between the actual reconstruction parameters and the preset reconstruction parameters, there are certain differences between the actual acquisition and reconstruction conditions and the preset acquisition and reconstruction conditions.
[0073] The image noise variation coefficient is obtained by comparing the actual acquisition and reconstruction conditions with the preset acquisition and reconstruction conditions. This coefficient can be used to adjust the baseline image noise to obtain the second image noise. This way, the baseline image noise can be obtained without actually using the first filtering kernel function to reconstruct the image based on the CT data corresponding to the sample image according to the preset acquisition and reconstruction conditions.
[0074] In one feasible implementation, the reference image noise is the sum of the sampling frequency noise contributions of the first filtering and summing function.
[0075] Specifically, the first filtering kernel function is sampled to obtain the sampling frequency components and their corresponding amplitude values; the preset noise contribution value corresponding to the sampling frequency components is obtained, and the preset noise contribution value is the contribution value of different frequency components to image noise under preset acquisition and reconstruction conditions; the product of the preset noise contribution value and the corresponding amplitude value is calculated to obtain the sampling frequency noise contribution value; the sum of the sampling frequency noise contribution values is statistically calculated to obtain the reference image noise.
[0076] The method for calculating noise in the reference image is as follows:
[0077]
[0078] Where N is the number of sampling points after sampling the first filter kernel function at a unit of 1 lp / cm. i These are a set of parameters pre-calibrated by the machine, with an upper limit of lp / cm corresponding to the system's maximum resolution. i This characterizes the contribution of different frequency components of CT data to image noise. i It is the value sampled by the first filtering kernel function, representing the amplitude of each frequency component.
[0079] In CT scan acquisition parameters, kV value, mA value, and filter type affect image noise. Similarly, in CT image reconstruction parameters, image thickness also affects image noise. These factors have independent effects on image noise and can be calibrated and adjusted individually on the CT system. Therefore, the image noise variation coefficient includes the kV value change rate W. kV rate of change of mA value W mA and image thickness change rate W ImgThickness wait.
[0080] Specifically, the second image noise L2 is calculated as follows:
[0081] Noise2 = Noise0·W kV ·W mA ·W ImgThickness .
[0082] For example, the preset acquisition and reconstruction conditions are 120kV, 100mA, and an axial scan is performed using the filter plate corresponding to the head to reconstruct an image with a thickness of 0.5mm. The actual acquisition and reconstruction conditions of the sample image are 100kV, 100mA, and an axial scan is performed using the filter plate to reconstruct an image with a thickness of 1mm. Then W kV This will adjust the image noise change caused by changing KV from 120kV to 100kV; W mA No change, value is 1; W ImgThickness The image noise change caused by adjusting the thickness from 0.5mm to 1mm is then addressed. Ultimately, the resulting image, obtained using the first filtering kernel function based on the actual acquisition and reconstruction conditions of the sample image and CT data, contains the atmospheric noise.
[0083] After obtaining the first image noise and the second image noise, a second loss function can be obtained based on the change between the first image noise and the second image noise. This function is used to characterize the difference in image noise between the first filter kernel function generated by the first convolutional neural network and the image reconstructed using the corresponding sample filter kernel function under the same acquisition and reconstruction conditions.
[0084] For example, the second loss function is calculated as follows:
[0085] L2 = |Noise1 - Noise2|
[0086] Therefore, the target loss function can also be calculated as L = L1 + L2, or L = L1 / (L1 + L2).
[0087] In step S4, the target filtering kernel function is used to reconstruct the image based on the data to be reconstructed, and finally a reconstructed image that meets the requirements is obtained.
[0088] like Figure 3 As shown, the image reconstruction method provided in this embodiment only needs to extract the target style features from the image preferred by the target user, input the target style features into the trained network model, and then directly infer the target filtering kernel function based on the target style features, and then perform image reconstruction with the corresponding style features. It no longer requires repeated tedious debugging, and can quickly obtain filtering kernel functions that meet different needs, greatly reducing the time and manpower costs required to generate filtering kernel functions.
[0089] Example 2
[0090] This embodiment provides an image reconstruction system.
[0091] like Figure 4 As shown, the image reconstruction system includes:
[0092] Sample image acquisition module 21 is used to acquire target sample images;
[0093] Style feature acquisition module 22 is used to extract style features from the target sample image to obtain target style features;
[0094] The objective function generation module 23 is used to input the target style features into the first convolutional neural network to generate the corresponding target filtering kernel function;
[0095] Image reconstruction module 24 is used to reconstruct the image based on the target filtering kernel function to obtain the reconstructed image.
[0096] The sample image acquisition module 21 is used to collect a batch of CT images of the style desired by the target user as target sample images.
[0097] The style feature acquisition module 22 is used to extract the style features of the target sample image as the target style features.
[0098] Style features are used to characterize the color, brightness, noise, texture, and other features of CT images.
[0099] In one feasible implementation, the style feature acquisition module 22 includes:
[0100] The image feature extraction unit is used to extract image features from multiple channels based on the target sample image using a second convolutional neural network including multiple convolutional kernels.
[0101] The style feature generation unit is used to obtain target style features based on image features from multiple channels using a Gram matrix.
[0102] Specifically, the target sample image is input into a second convolutional neural network, which includes multiple convolutional kernels. Different convolutional kernels are used to output image features for different channels. That is, the target sample image is processed by different convolutional kernels to extract image features for multiple channels. These image features for multiple channels are then represented using a Gram matrix to obtain the target style features.
[0103] For example, the target sample image has a size of n*n, and the second convolutional neural network includes c convolutional kernels k, each kernel k having a size of m*m. The target sample image is processed by the second convolutional neural network to obtain image features I across multiple channels, which are calculated as follows:
[0104] I c,j,j =I n,n *k c,m,m ;
[0105] The target style feature IG is obtained by representing the image features of c channels using the Gram matrix, and its calculation method is as follows:
[0106]
[0107] The objective function generation module 23 is used to infer the objective style features using the first convolutional neural network, which can then obtain the objective filtering kernel function.
[0108] In one feasible implementation, the image reconstruction system further includes a model training module for training a first convolutional neural network based on a training dataset. The training dataset includes several sample images and sample filtering kernel functions for the reconstructed sample images.
[0109] The training process involves backpropagating the error based on the target loss function and iteratively updating the model parameters of the neural network until the target loss function is less than a preset threshold.
[0110] In one feasible implementation, the first convolutional neural network includes a series of convolutional layers. In another feasible implementation, the first convolutional neural network also includes a residual module.
[0111] In one feasible implementation, the target loss function is determined based on the first loss function.
[0112] Specifically, a first filtering kernel function is first generated based on the style features of the sample images. Then, a first loss function representing the style differences of the images is obtained based on the first filtering kernel function and the sample filtering kernel function. The first loss function is used as the target loss function.
[0113] For example, the first filtering kernel function generated based on the style features of the sample images is Kernel1, and the sample filtering kernel function is Kernel0. The first loss function L1 is the mean square error between the first filtering kernel function and the sample filtering kernel function, and its specific calculation method is as follows:
[0114]
[0115] In another feasible implementation, such as Figure 2 As shown, the target loss function is determined based on the first loss function and the second loss function.
[0116] Specifically, the first image noise is obtained by acquiring the atmospheric noise of the sample image. Then, the first filtering kernel function is used to reconstruct the image based on the actual acquisition and reconstruction conditions and CT data corresponding to the sample image. The atmospheric noise of the reconstructed image is used as the second image noise. Finally, the difference between the first image noise and the second image noise is obtained to obtain the second loss function. Here, atmospheric noise refers to the image noise in the region of the image that does not include the imaging object.
[0117] For example, if n square regions I are selected where the image edge does not include the imaged object, then the image noise in each square region is I. air The corresponding air image noise is for all square regions I air The image noise is the square root of the sample variance. Therefore, the first image noise is calculated as follows:
[0118]
[0119] In one feasible implementation, the second image noise is obtained by adjusting the reference image noise using an image noise variation coefficient.
[0120] Among them, the reference image noise is used to characterize the air image noise of the image obtained by image reconstruction based on the CT data corresponding to the sample image using the first filtering kernel function according to the preset acquisition and reconstruction conditions.
[0121] Furthermore, due to the differences between the actual acquisition parameters and the preset acquisition parameters, as well as the differences between the actual reconstruction parameters and the preset reconstruction parameters, there are certain differences between the actual acquisition and reconstruction conditions and the preset acquisition and reconstruction conditions.
[0122] The image noise variation coefficient is obtained by comparing the actual acquisition and reconstruction conditions with the preset acquisition and reconstruction conditions. This coefficient can be used to adjust the baseline image noise to obtain the second image noise. This way, the baseline image noise can be obtained without actually using the first filtering kernel function to reconstruct the image based on the CT data corresponding to the sample image according to the preset acquisition and reconstruction conditions.
[0123] In one feasible implementation, the reference image noise is the sum of the sampling frequency noise contributions of the first filtering and summing function.
[0124] Specifically, the first filtering kernel function is sampled to obtain the sampling frequency components and their corresponding amplitude values; the preset noise contribution value corresponding to the sampling frequency components is obtained, and the preset noise contribution value is the contribution value of different frequency components to image noise under preset acquisition and reconstruction conditions; the product of the preset noise contribution value and the corresponding amplitude value is calculated to obtain the sampling frequency noise contribution value; the sum of the sampling frequency noise contribution values is statistically calculated to obtain the reference image noise.
[0125] The method for calculating noise in the reference image is as follows:
[0126]
[0127] Where N is the number of sampling points after sampling the first filter kernel function at a unit of 1 lp / cm. i These are a set of parameters pre-calibrated by the machine, with an upper limit of lp / cm corresponding to the system's maximum resolution. i This characterizes the contribution of different frequency components of CT data to image noise. i It is the value after sampling by the first filter kernel function, representing the amplitude of each frequency component.
[0128] In CT scan acquisition parameters, kV value, mA value, and filter type affect image noise. Similarly, in CT image reconstruction parameters, image thickness also affects image noise. These factors have independent effects on image noise and can be calibrated and adjusted individually on the CT system. Therefore, the image noise variation coefficient includes the kV value change rate W. kV rate of change of mA value W mA and image thickness change rate W ImgThickness wait.
[0129] Specifically, the second image noise L2 is calculated as follows:
[0130] Noise2 = Noise0·W kV ·W mA ·W ImgThickness .
[0131] For example, the preset acquisition and reconstruction conditions are 120kV, 100mA, and an axial scan is performed using the filter plate corresponding to the head to reconstruct an image with a thickness of 0.5mm. The actual acquisition and reconstruction conditions of the sample image are 100kV, 100mA, and an axial scan is performed using the filter plate to reconstruct an image with a thickness of 1mm. Then W kV This will adjust the image noise change caused by changing KV from 120kV to 100kV; W mA No change, value is 1; W ImgThickness The image noise change caused by adjusting the thickness from 0.5mm to 1mm is then assessed. Ultimately, the atmospheric noise of the image reconstructed using the first filtering kernel function based on the actual acquisition and reconstruction conditions corresponding to the sample image and CT data is obtained.
[0132] After obtaining the first image noise and the second image noise, a second loss function can be obtained based on the change between the first image noise and the second image noise. This function is used to characterize the difference in image noise between the first filter kernel function generated by the first convolutional neural network and the image reconstructed using the corresponding sample filter kernel function under the same acquisition and reconstruction conditions.
[0133] For example, the second loss function is calculated as follows:
[0134] L2 = |Noise1 - Noise2|
[0135] Therefore, the target loss function can also be calculated as L = L1 + L2, or L = L1 / (L1 + L2).
[0136] The image reconstruction module 24 is used to reconstruct the image based on the data to be reconstructed using the target filtering kernel function, and finally obtain the reconstructed image that meets the requirements.
[0137] like Figure 3 As shown, the image reconstruction system provided in this embodiment only needs to extract the target style features from the image preferred by the target user, input the target style features into the trained network model, and then directly infer the target filtering kernel function based on the target style features, and then perform image reconstruction with the corresponding style features. There is no need to repeatedly perform tedious debugging, and the filtering kernel function that meets different needs can be obtained quickly, greatly reducing the time and manpower costs required to generate the filtering kernel function.
[0138] Example 3
[0139] Figure 5This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of this disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the image reconstruction method described in Embodiment 1. Figure 5 The electronic device 30 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments disclosed herein.
[0140] like Figure 5 As shown, the electronic device 30 can be represented in the form of a general computing device, such as a server device. The components of the electronic device 30 may include, but are not limited to: at least one processor 31, at least one memory 32, and a bus 33 connecting different system components (including memory 32 and processor 31).
[0141] Bus 33 includes a data bus, an address bus, and a control bus.
[0142] The memory 32 may include volatile memory, such as random access memory (RAM) 321 and / or cache memory 322, and may further include read-only memory (ROM) 323.
[0143] The memory 32 may also include a program / utility 325 having a set (at least one) of program modules 324, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.
[0144] The processor 31 executes various functional applications and data processing, such as the image reconstruction method described in Embodiment 1 of this disclosure, by running computer programs stored in the memory 32.
[0145] Electronic device 30 can also communicate with one or more external devices 34 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 35. Furthermore, the model-generating device 30 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 36. Figure 5 As shown, network adapter 36 communicates with other modules of the model-generated device 30 via bus 33. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the model-generated device 30, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.
[0146] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.
[0147] Example 4
[0148] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the image reconstruction method described in Embodiment 1.
[0149] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.
[0150] In a possible implementation, this disclosure can also be implemented as a program product comprising program code that, when the program product is run on a terminal device, causes the terminal device to execute and implement the image reconstruction method described in Embodiment 1.
[0151] The program code for executing this disclosure can be written using any combination of one or more programming languages. The program code can be executed entirely on a user device, partially on a user device, as a standalone software package, partially on a user device and partially on a remote device, or entirely on a remote device.
[0152] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.
Claims
1. An image reconstruction method, characterized in that, The method includes: Acquire the target sample image; The style features of the target sample image are extracted to obtain the target style features; The target style features are input into the first convolutional neural network to generate the corresponding target filtering kernel function; Based on the target filtering kernel function, image reconstruction is performed on the data to be reconstructed to obtain the reconstructed image; The first convolutional neural network is obtained by training a neural network model based on a training dataset. The training process involves backpropagating the error according to the target loss function and iteratively updating the model parameters of the neural network model until the target loss function is less than a preset threshold. The target loss function is determined based on a first loss function and a second loss function, wherein the second loss function is the difference between the first image noise and the second image noise, and the second loss function is used to characterize the difference in image noise. The first image noise is the air image noise of the sample image; The second image noise is the air image noise of the image obtained by image reconstruction based on the actual acquisition and reconstruction conditions and CT data corresponding to the sample image using the first filtering kernel function; The first filtering kernel function is generated based on the style features of the sample image; The air image noise refers to the image noise in areas of the image that do not include the imaged object.
2. The image reconstruction method according to claim 1, characterized in that, The step of extracting style features from the target sample image to obtain target style features includes: A second convolutional neural network, comprising multiple convolutional kernels, is used to extract image features from the target sample image to obtain features from multiple channels. The target style features are obtained using the Gram matrix based on the image features of the multiple channels.
3. The image reconstruction method according to claim 1, characterized in that, The training dataset includes several sample images and a sample filtering kernel function for reconstructing the sample images.
4. The image reconstruction method according to claim 1, characterized in that, The target loss function is determined based on a first loss function, which is obtained based on a first filtering kernel function and a sample filtering kernel function. The first loss function is used to characterize image style differences.
5. The image reconstruction method according to claim 1, characterized in that, The second image noise is obtained by adjusting the reference image noise using an image noise variation coefficient; The reference image noise is used to characterize the air image noise of the image obtained by reconstructing the image based on the CT data corresponding to the sample image using the first filtering kernel function according to preset acquisition and reconstruction conditions. The image noise variation coefficient is determined based on the difference between the actual acquisition and reconstruction conditions and the preset acquisition and reconstruction conditions.
6. The image reconstruction method according to claim 5, characterized in that, The noise in the reference image is the sum of the noise contribution values of the sampling frequency; The sampling frequency noise contribution value is the product of the preset noise contribution value and the corresponding amplitude value. The preset noise contribution value is the contribution value of different frequency components to image noise under the preset acquisition and reconstruction conditions. The frequency components and their corresponding amplitude values are obtained by sampling the first filtering kernel function.
7. An image reconstruction system, characterized in that, The system includes: The sample image acquisition module is used to acquire target sample images; The style feature acquisition module is used to extract the style features of the target sample image to obtain the target style features; The objective function generation module is used to input the target style features into the first convolutional neural network to generate the corresponding target filtering kernel function; The image reconstruction module is used to reconstruct the image based on the target filtering kernel function to obtain the reconstructed image; The model training module is used to train a neural network model based on a training dataset to obtain the first convolutional neural network. The training process involves backpropagating the error according to the target loss function and iteratively updating the model parameters of the neural network model until the target loss function is less than a preset threshold. The target loss function is determined based on a first loss function and a second loss function. The second loss function is the difference between the first image noise and the second image noise, and is used to characterize the difference in image noise. The first image noise is the atmospheric noise of the sample image. The second image noise is the air image noise of the image obtained by image reconstruction based on the actual acquisition and reconstruction conditions and CT data corresponding to the sample image using the first filtering kernel function; The first filtering kernel function is generated based on the style features of the sample image; The air image noise refers to the image noise in areas of the image that do not include the imaged object.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes a computer program, it implements the image reconstruction method as described in any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the image reconstruction method as described in any one of claims 1-6.