Dual low-dose attenuation correction method and system for pet / CT system

Abstract

A dual low-dose attenuation correction method for a PET / CT system, belonging to the field of medical positron emission tomography, and used to achieve the goal of reducing radiation dose while ensuring image quality of imaging. The technical solution is: performing multi-scale feature extraction on a low-dose ACCT image, to obtain a first feature map set {fj CT|j=1, 2, ..., N}, and performing multi-scale feature extraction on a low-dose non-attenuated PET image, to obtain a second feature map set {fj PET|j=1, 2, ..., N}, N being a set value; adaptively aligning spatial positions of the first feature map fj CT and the second feature map fj PET, and matching and fusing the two aligned feature maps, to obtain a third feature map set {fj CT-PET|j=1, 2, ..., N}; performing scale-invariant feature extractions on the third feature map f<sp / > CT-PET a number M of times, to obtain an attenuation-corrected feature map fN CT-PET, M being a set value; enlarging the feature map fj AC-PET to the same size as fj-1 CT-PET, and splicing fj-1 CT-PET ​​with fj AC-PET, to obtain an attenuation-corrected feature map fj-1 AC-PET, j=N, ..., 2; on the basis of the feature map f1 AC-PET, obtaining a visible standard-dose PET image.

Description

Dual low-dose attenuation correction method and system for PET / CT system Technical Field

[0001] This case involves medical positron emission tomography, and in particular, a dual low-dose attenuation correction method and system for PET / CT systems. Background Art

[0002] Positron emission tomography / computed tomography (PET / CT) is a medical system widely used in the management of cancer patients, including diagnosis, monitoring and follow-up treatment.

[0003] In PET imaging, attenuation correction is performed to improve visual interpretation and enable accurate quantitative analysis. CT-based methods are commonly used for PET attenuation correction, converting CT Hounsfield units into linear attenuation coefficients at 511 keV energy using a bilinear relationship. However, this leads to spatial misalignment between PET and CT and limitations of CT-based artifacts.

[0004] In addition, after sequential scans under PET / CT scanners, patients may face increased radiation risks due to CT dose and potential hazards of overexposure, especially children. While the new generation of whole-body PET / CT scanners has unprecedented image quality and quantitative accuracy, with approximately 40 times higher sensitivity, patients receive more radiation exposure during whole-body scanning compared to single-site examinations.

[0005] Summary of the Invention

[0006] In order to solve the above-mentioned problems existing in the prior art, the purpose of this case is to propose a dual low-dose attenuation correction method for PET / CT systems. During its implementation, the radiation sources in PET and CT imaging are taken into account, and the proposed dual dose reduction strategy is used to directly generate normal-dose AC PET images from low-dose PET and low-dose ACCT (CT images utilized for attenuation correction) images by reducing the injection dose and tube current, thereby reducing the radiation risk of the PET / CT system. The method can be used to obtain normal-dose whole-body PET images in whole-body PET / CT equipment.

[0007] In order to achieve the above technical objectives, the technical solution of this case is as follows.

[0008] First, this case proposes a dual low-dose attenuation correction method for a PET / CT system, the method comprising the following steps:

[0009] Perform multi-scale feature extraction on low-dose ACCT images to obtain the first feature map set Perform multi-scale feature extraction on low-dose unattenuated PET images to obtain the second feature map set N is the set value;

[0010] The first feature map Second feature map Perform adaptive spatial position alignment, match and fuse the two aligned feature maps to obtain the third feature map set

[0011] The third feature map Perform M times of scale-invariant feature extraction to obtain the feature map of attenuation correction M is the set value;

[0012] The feature map Enlarge it to the same size as Same size and and Splicing to get the attenuation correction feature map

[0013] Based on feature map Obtain visualizable standard-dose PET images.

[0014] In the above technical solution, one implementation of adaptive spatial position alignment includes the following steps:

[0015] Based on the first feature map Get the second feature map The first feature map of the same size For j = 1, the first feature map with the same size as the cropped low-dose unattenuated PET image is obtained.

[0016] Based on feature map Perform affine transformation to obtain affine transformation parameters (γ, β), j = 1, 2, ..., N;

[0017] Using affine transformation parameters (γ, β), the feature map Perform scaling and translation operations on the channel to obtain the second feature map

[0018] In the above technical solution, one implementation step of matching and fusion includes:

[0019] Based on feature map and Use Hadamard product operation to achieve matching and obtain feature map

[0020] Based on the first feature map Get the second feature map The first feature map of the same size For j = 1, the first feature map of the same size as the cropped low-dose unattenuated PET image is obtained.

[0021] The feature map and Fusion is achieved using Hadamard addition operation, j = 1, 2, ..., N.

[0022] Secondly, this case proposes a dual low-dose attenuation correction system for PET / CT systems. The system uses a dual low-dose attenuation calibration model to take low-dose ACCT images and low-dose unattenuated PET images as input, and outputs standard-dose attenuation-corrected PET images.

[0023] The dual low-dose attenuation calibration model includes a first encoder, a second encoder, a spatial position alignment module and a decoder, wherein:

[0024] The first encoder is configured to perform multi-scale feature extraction on the low-dose ACCT image to obtain a first feature map set N is the set value;

[0025] The second encoder is configured to perform multi-scale feature extraction on the low-dose unattenuated PET image to obtain a second feature map set

[0026] The spatial position alignment module is configured to realize the first feature map Second feature map Perform adaptive spatial position alignment, match and fuse the two aligned feature maps to obtain the third feature map set

[0027] The decoder is configured to implement the feature map Upsample to make it Same size and and Splicing to get the attenuation correction feature map Among them: feature map Based on the third feature map Perform M times of scale-invariant feature extraction to obtain it, where M is a set value.

[0028] In the above technical solution, one implementation of the first encoder, the second encoder and the decoder is that they are all composed of residual modules;

[0029] Each residual module that constitutes the encoder is composed of two 3D convolutional layers, with a batch normalization layer and an activation function layer set between the two convolutional layers;

[0030] Each residual module constituting the decoder is composed of a 3D transposed deconvolution layer and a convolution layer, with a batch normalization layer and an activation function layer between the transposed deconvolution layer and the convolution layer.

[0031] In one embodiment of the above technical solution, the loss function of the dual low-dose attenuation calibration model training is as follows:

[0032] Where: G represents the mapping relationship, θ i represents the network parameters, x i is a sample in the low-dose ACCT image dataset, y i is a sample in a low-dose unattenuated PET image dataset with various dose levels, z i is the sample in the PET image dataset after standard dose attenuation correction, and n is the total number of training samples.

[0033] In one embodiment of the above technical solution, the spatial position alignment module includes a cropping unit, an affine unit, and a scaling and translation unit; wherein:

[0034] The cropping unit is configured to be based on the first feature map Get the second feature map The first feature map of the same size For j = 1, the first feature map of the same size as the cropped low-dose unattenuated PET image is obtained.

[0035] Affine unit, configured based on feature map Perform affine transformation to obtain affine transformation parameters (γ, β), j = 1, 2, ..., N;

[0036] The scaling and translation unit is configured to use the affine transformation parameters (γ, β) to transform the feature map Perform scaling and translation operations on the channel to obtain the second feature map

[0037] In one implementation of the above technical solution, the affine unit obtains affine transformation parameters using a filter and a sigmoid function.

[0038] In one embodiment of the above technical solution, the spatial position alignment module further includes a matching unit and a fusion unit; wherein:

[0039] The matching unit is configured to be based on the feature map and Use Hadamard product operation to achieve matching and obtain feature map

[0040] The fusion unit is configured to combine the feature maps and Fusion is achieved using Hadamard addition, j = 1, 2, ..., N, where: feature map By getting it from the clipping unit again.

[0041] In one embodiment of the above technical solution, M residual modules are used to Perform M times of scale-invariant feature extraction to obtain the feature map of attenuation correction M is the set value.

[0042] Beneficial technical effects of this case:

[0043] Using a deep learning model to directly generate attenuation-corrected PET images from low-dose unattenuated PET images and low-dose ACCT images significantly reduces patient radiation risk and is particularly beneficial for the diagnosis, monitoring, and follow-up treatment of pediatric diseases. When used for whole-body imaging, this approach reduces CT scan time and minimizes respiratory artifacts during CT scans, thereby ensuring the quality of attenuation-corrected PET images. Therefore, this approach can both ensure image quality and reduce harmful radiation dose, and has important scientific significance and application prospects in the field of medical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0045] Figure 1 is a schematic diagram of the overall system framework in one embodiment;

[0046] FIG2 is a schematic diagram of image size changes during spatial alignment in one embodiment;

[0047] FIG3 is a PET image after attenuation correction in one embodiment;

[0048] FIG4 is a schematic diagram of a reference standard image in one embodiment;

[0049] FIG. 5 . An unattenuated PET image according to one embodiment. DETAILED DESCRIPTION

[0050] PET attenuation correction requires CT images to calculate the attenuation correction coefficient map, resulting in higher radiation doses for patients. Most current methods for low-dose whole-body PET / CT imaging ignore the actual physical characteristics of the equipment and scanning protocols, or fail to consider the dose issue of ACCT (CT images utilized for attenuation correction).

[0051] To this end, this solution, based on considerations of radiation sources in PET and CT imaging, utilizes a deep learning model to directly generate normal-dose AC PET images from low-dose PET and low-dose ACCT images, while achieving dual dose reduction and meeting clinical application requirements, thereby reducing the radiation risk of the PET / CT system. The technical solution of the present invention can be used for both localized (e.g., brain) PET attenuation correction and whole-body PET attenuation correction without worrying about increased radiation. Furthermore, the dual low-dose strategy shortens CT scan time, thereby suppressing respiratory artifacts introduced during whole-body CT scans. This, in turn, improves the quality of PET attenuation correction and the accuracy of quantitative analysis of PET images.

[0052] The following, combined with the accompanying drawings, provides a clear and complete description of how the technical solution of this case is implemented. Obviously, the described implementation methods are only part of the implementation methods of this case, not all of the implementation methods. Based on the implementation methods of this case, all other implementation methods obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0053] FIG1 is a schematic diagram of a dual low-dose attenuation calibration model used when implementing a dual low-dose attenuation correction system for a PET / CT system.

[0054] The dual low-dose attenuation calibration model includes a first encoder (first gray column), a second encoder (second green column), a spatial position alignment module, and a decoder (third blue column). Referring to Figure 1, the first encoder and the second encoder first expand the number of channels of the image to increase the dimension before extracting features from the image. Correspondingly, at the end of the decoder, the number of channels is restored to obtain a visible image. In the feature map processing part, illustratively, the first encoder, the second encoder, and the decoder are all composed of 5 residual modules.

[0055] The first encoder is used to extract features from the input low-dose ACCT image. Each residual module is implemented using two three-dimensional convolutional layers (using a filter size of 3×3×3). At the same time, a batch normalization layer and an activation function layer are added between these two convolutional layers, which can be expressed as: f j+1 =F j (f j )=g(w j f j +b j )

[0056] Where: f j+1 Represents the feature map obtained by the residual module, w j and b j Represents the weights and biases corresponding to the convolution operation.

[0057] The output of the encoder for low-dose ACCT images can be expressed as: CT =F L (...F1(x i ))

[0058] Where: f CT Represents the ACCT image output feature map, x i represents a low-dose ACCT image sample, F j Represents the residual module, L represents the number of residual modules, thereby obtaining multi-scale features, which are recorded as the first feature map set The schematic diagram of image size change is shown in the corresponding column of Figure 2.

[0059] Similarly, for low-dose PET images, another second encoder with the same design is introduced to extract features from the input low-dose NAC PET images, which can also be expressed as: PET =F L (...F1(y i ))

[0060] Where: f CT =F L (...F1(x i )) represents the output feature map of the low-dose unattenuated PET image, y i Represents a low-dose, unattenuated PET image sample. After feature extraction by the residual module, it is downsampled using a convolution with a kernel size of 3×3×3 and a step size of 2, so that the network can obtain features of different sizes. The obtained multi-scale features are recorded as The schematic diagram of image size change is shown in the corresponding column of Figure 2.

[0061] The input low-dose ACCT images (512×512) and NAC PET images (192×192) have different sizes. Therefore, a spatial position alignment module is designed to facilitate image feature matching and fusion of low-dose PET and ACCT images at different scales, taking into account the features of the two modalities.

[0062] First, the NAC (Non-attenuation-corrected) PET image is cropped to a size of 128 × 128. Then, the first feature map corresponding to the ACCT image is Perform two downsamplings, use a convolution kernel size of 3×3×3 and a step size of 2, so that the feature map and feature images The same size, j = 2, 3, ..., N; for j = 1, the first feature map with the same size as the cropped low-dose unattenuated PET image is obtained

[0063] Based on feature map Perform affine transformation to obtain affine transformation parameters (γ, β), j = 1, 2, ..., N:

[0064] Where: H represents affine transformation.

[0065] In a specific embodiment, the ACCT intermediate feature map is subjected to a 3×3×3 filter and a sigmoid function to obtain affine transformation parameters (γ, β) for the feature image of the PET image. Zooming and panning operations are performed on the channel, which can be expressed as follows:

[0066] In the above-mentioned spatial position alignment module, the feature matching and fusion of low-dose PET and ACCT images are promoted by performing adaptive affine transformation on the channel dimension of the intermediate feature map.

[0067] The lower left part of Figure 1 shows a matching and fusion method. Specifically, based on the feature map and Use Hadamard product operation to achieve matching and obtain feature map Based on the first feature map Get the second feature map The first feature map of the same size For j = 1, the first feature map of the same size as the cropped low-dose unattenuated PET image is obtained. The feature map and The fusion is realized by using Hadamard addition operation, j = 1, 2, ..., N. In this process, the feature map and When matching, two downsamplings are performed; after matching, the data is retrieved from the cropping unit. That is, again the feature map After downsampling, the result is fused with the matching result to finally obtain the output result of the spatial position alignment module, which is recorded as the third feature map set.

[0068] For the third feature map Perform M times of scale-invariant feature extraction to obtain the feature map of attenuation correction M is a set value. For example, three residual modules connected in series are used to implement three-time scale-invariant feature extraction.

[0069] Will The decoder is fed into the decoder, which predicts normal-dose AC-PET images through corresponding upsampled transposed convolution operations and short connections. The decoder also has five residual modules, each implemented using a 3D transposed deconvolution layer (using a filter size of 3×3×3 with a stride of 2) and a convolution layer (using a filter size of 3×3×3 with a stride of 1). Furthermore, a batch normalization layer and an activation function layer are added between these two transposed deconvolution layers. Skip connections connect the fused PET and ACCT image features with the upsampled features of the corresponding scale, preserving more contextual information.

[0070] Refer to the blue column in Figure 1. From bottom to top, the decoder first transforms the feature map Upsample and enlarge it to the same size as Same size and and Attenuation correction feature map is obtained by splicing through jump connections The image size change diagram is shown in the corresponding column of Figure 2. After that, a visual standard measurement PET image is obtained through channel restoration.

[0071] In summary, the dual low-dose attenuation correction system for PET / CT system uses a dual low-dose attenuation calibration model to take low-dose ACCT images and low-dose non-attenuation-corrected PET images as input, and takes standard-dose attenuation-corrected PET images as output.

[0072] Attenuation correction is a commonly used technique in nuclear medicine imaging (such as positron emission tomography-computed tomography (PET-CT)) to correct for variations in image brightness caused by radioisotope decay. It compensates for radiation absorption in various regions of the image, making them more accurate and reliable. In this study, a dual low-dose attenuation calibration model was trained to calibrate low-dose, unattenuation-corrected PET images using low-dose ACCT images, resulting in standard-dose, attenuation-corrected PET images. The training dataset (PET and CT data) was obtained from a uEXPLORER whole-body PET / CT scanner operating at a tube voltage of 120 kVp and a tube current of 20 mA. The dataset included unattenuation-corrected whole-body PET images at various dose levels. To simulate the noise distribution in PET images, we roughly categorized the noise distribution into four categories, designated as 1.0%, 2.5%, 5.0%, and 25% dose. All simulated doses were extracted from the standard raw data by randomly selecting a certain proportion of count events. The reconstructed PET images were 192 × 192 × 673 pixels in size and 2.9 mm thick. For low-dose ACCT images, the scanning tube voltage was 120 kVp and the tube current was 20 mA. The thickness of the low-dose ACCT image was set to 3 mm. The resulting ACCT image size was 512 × 512 × 673.

[0073] Assume that a training dataset D = {(x1, y1), (x2, y2), ..., (x n ,y n )}, where {x1, x2, ..., x n} are samples in the low-dose ACCT image dataset, {y1, y2, ..., y n} are samples from a low-dose unattenuated PET image dataset, {z1, z2, ..., z n} is a sample in the PET image dataset after standard dose attenuation correction, and n is the total number of training samples.

[0074] Next, a loss function for training the low-dose attenuation calibration model is constructed. One implementation method is to use the MSE function as the loss function. The training loss can be expressed as:

[0075] Where: G represents the mapping relationship, θ i represents the network parameters, x i is a sample in the low-dose ACCT image dataset, y i is a sample in a low-dose unattenuated PET image dataset with various dose levels, z i is the sample in the PET image dataset after standard dose attenuation correction, and n is the total number of training samples.

[0076] The dual low-dose attenuation calibration model is trained using the above training dataset, that is, the low-dose unattenuated corrected whole-body PET image y i (See Figure 3) and low-dose ACCT images x i (See Figure 4) as the network input, the PET image z after attenuation correction with standard dose i As the reference standard (ground truth) (see Figure 5), and the Adam optimizer is used for optimization.

[0077] After the trained dual low-dose attenuation calibration model is input with low-dose ACCT images and low-dose unattenuated PET images, the processing process of the two is as follows:

[0078] Perform multi-scale feature extraction on low-dose ACCT images to obtain the first feature map set Perform multi-scale feature extraction on low-dose unattenuated PET images to obtain the second feature map set N is the set value;

[0079] The first feature map Second feature map Perform adaptive spatial position alignment, match and fuse the two aligned feature maps to obtain the third feature map set

[0080] The third feature map Perform M times of scale-invariant feature extraction to obtain the feature map of attenuation correction M is the set value;

[0081] The feature map Enlarge it to the same size as Same size and and Splicing to get the attenuation correction feature map

[0082] Based on feature map Obtain visualizable standard-dose PET images.

[0083] As can be seen from the implementation process described above, the dual dose reduction strategy employed in this case reduces the radiation dose in whole-body PET / CT scans. This is of great significance for ultra-low-dose scanning protocols for whole-body scans, ensuring the integrity of the scanning protocol. When performing whole-body PET attenuation correction, the ACCT dose is reduced rather than eliminated, making it easier to gain clinical acceptance. Current whole-body CT scans take approximately 20 seconds and introduce respiratory artifacts. However, the dual low-dose whole-body PET attenuation correction strategy can reduce CT scan time and minimize respiratory artifacts present during CT scans, thereby ensuring the quality of PET images after attenuation correction. This case allows for the acquisition of PET images at the lowest clinically acceptable dose level while simultaneously reducing both PET and CT doses.

[0084] It should be noted that this technical solution can be applied to tracer PET / SPECT imaging, which helps to provide diagnostic reference and formulate clinical diagnostic guidelines.

[0085] In the above process, the terms "first," "second," and "third" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, the terms "first," "second," and "third" may explicitly or implicitly include one or more of the features.

[0086] Through the above description of the embodiments, those skilled in the art will clearly understand that the disclosed method or system can be implemented by means of software plus necessary general-purpose hardware. Of course, it can also be implemented by means of dedicated hardware, including application-specific integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. In general, any function performed by a computer program can be easily implemented by the corresponding hardware. Moreover, the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for the present disclosure, software program implementation is often the preferred embodiment.

[0087] It should be noted that references to "one embodiment," "another embodiment," or "an embodiment" in this specification refer to specific features, structures, or characteristics described in conjunction with that embodiment as included in at least one embodiment generally described in this application. The appearance of the same expression in multiple places in the specification does not necessarily refer to the same embodiment. Furthermore, when a specific feature, structure, or characteristic is described in conjunction with any embodiment, it is intended that such feature, structure, or characteristic, when implemented in conjunction with other embodiments, also falls within the scope of the present invention.

[0088] Although the embodiments of the present invention have been described above with reference to the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments and application fields. The above-mentioned specific embodiments are merely illustrative and instructive, and are not restrictive. A person skilled in the art, guided by this specification and without departing from the scope of protection of the claims of the present invention, may also devise various forms, all of which fall within the scope of protection of the present invention.

Claims

1. A dual low-dose attenuation correction method for a PET / CT system, characterized in that: The method comprises the following steps: Perform multi-scale feature extraction on low-dose ACCT images to obtain the first feature map set Perform multi-scale feature extraction on low-dose unattenuated PET images to obtain a second feature map set N is the set value; The first feature map Second feature map Perform adaptive spatial position alignment, match and fuse the two aligned feature maps to obtain the third feature map set The third feature map Perform M times of scale-invariant feature extraction to obtain the feature map of attenuation correction M is the set value; The feature map Enlarge it to the same size as The same size and and Splice to get the attenuation correction feature map Based on feature map Obtain visual standard-dose PET images.

2. The method according to claim 1, characterized in that The steps of adaptive spatial position alignment include: Based on the first feature map Get the second feature map The first feature map of the same size For j = 1, the first feature map with the same size as the cropped low-dose unattenuated PET image is obtained. Based on feature map Perform affine transformation to obtain affine transformation parameters (γ, β), j = 1, 2, ..., N; Using the affine transformation parameters (γ, β), the feature map Perform scaling and translation operations on the channel to obtain the second feature map 3. The method according to claim 2, characterized in that Matching and fusion, the implementation steps include: Based on feature map and Use Hadamard product operation to achieve matching and obtain feature map Based on the first feature map Get the second feature map The first feature map of the same size For j = 1, the first feature map with the same size as the cropped low-dose unattenuated PET image is obtained. The feature map and Fusion is achieved using Hadamard addition operation, j=1, 2,…,N.

4. A dual low-dose attenuation correction system for a PET / CT system, characterized in that: The system uses a dual low-dose attenuation calibration model to take a low-dose ACCT image and a low-dose PET image without attenuation correction as input, and a standard-dose PET image after attenuation correction as output; The dual low-dose attenuation calibration model includes a first encoder, a second encoder, a spatial position alignment module and a decoder, wherein: The first encoder is configured to perform multi-scale feature extraction on the low-dose ACCT image to obtain a first feature map set N is the set value; The second encoder is configured to perform multi-scale feature extraction on the low-dose unattenuated PET image to obtain a second feature map set The spatial position alignment module is configured to realize the first feature map Second feature map Perform adaptive spatial position alignment, match and fuse the two aligned feature maps to obtain the third feature map set The decoder is configured to implement the feature map Upsample to make it The same size and and Splice to get the attenuation correction feature map Among them: feature map Based on the third feature map The scale-invariant feature is extracted M times, where M is a set value.

5. The system according to claim 4, characterized in that: The first encoder, the second encoder and the decoder are all composed of residual modules; Each residual module constituting the encoder is composed of two three-dimensional convolutional layers, and a batch normalization layer and an activation function layer are set between the two convolutional layers; Each residual module constituting the decoder is composed of a 3D transposed deconvolution layer and a convolution layer, with a batch normalization layer and an activation function layer between the transposed deconvolution layer and the convolution layer.

6. The system according to claim 4, characterized in that: The loss function for the dual low-dose attenuation calibration model training is as follows: Where: G represents the mapping relationship, θ i represents the network parameters, x i is a sample in the low-dose ACCT image dataset, y i is a sample in a low-dose unattenuated PET image dataset with various dose levels, z i is the sample in the PET image dataset after standard dose attenuation correction, and n is the total number of training samples.

7. The system according to claim 4, characterized in that: The spatial position alignment module includes a cropping unit, an affine unit, and a scaling and translation unit; among them: A cropping unit is configured to: Get the second feature map The first feature map of the same size For j = 1, the first feature map with the same size as the cropped low-dose unattenuated PET image is obtained. Affine units are configured based on feature maps Perform affine transformation to obtain affine transformation parameters (γ, β), j = 1, 2, ..., N; The scaling and translation unit is configured to use the affine transformation parameters (γ, β) to transform the feature map Perform scaling and translation operations on the channel to obtain the second feature map 8. The system according to claim 7, characterized in that: The affine unit uses filters and sigmoid functions to obtain affine transformation parameters.

9. The system according to claim 7, characterized in that The spatial position alignment module also includes a matching unit and a fusion unit; wherein: The matching unit is configured to be based on the feature map and Use Hadamard product operation to achieve matching and obtain feature map The fusion unit is configured to combine the feature maps and Fusion is achieved using Hadamard addition, j = 1, 2, ..., N, where: feature map By getting it again from the clipping unit.

10. The system according to claim 4, characterized in that: Use M residual modules to transform the third feature map Perform M times of scale-invariant feature extraction to obtain the feature map of attenuation correction M is the set value.

Images