PET / CT imaging lung image restoration model training method, restoration method and device

Abstract

The present invention discloses a lung image restoration model training method, restoration method, and device for PET / CT imaging. The model training method includes: dividing a sample lung image into n first image data and reducing the resolution to obtain n second image data; selecting four adjacent second image data, and obtaining forward feature parameters and reverse feature parameters based on the pacth array of a global special token and the pacth array of a universal human body token; performing corrections through a long short-term memory network to obtain forward target feature parameters and reverse target feature parameters; obtaining n third image data and a compensation value; and obtaining a target loss function based on the compensation value, the first image data, and the second image data to obtain a lung image restoration model. The pacth array of the universal human body token can be used to share this feature vector when calculating all images, thereby achieving a balance between the complexity and accuracy of machine learning, ensuring the accuracy of the image restoration results while reducing the time and cost spent.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image restoration, and in particular to a PET / CT lung image restoration model training method, a restoration method, and a device. Background Art

[0002] Computed Tomography (CT) medical imaging technology arranges raw data collected from detectors into a two-dimensional matrix with the detector channels as the horizontal axis and the scanning field of view as the vertical axis. The raw CT projection data used for image reconstruction, also known as a sinogram, is essentially a superposition of the curves formed by each point on the image. Furthermore, before a PET / CT examination, a drug, an imaging agent, is injected. This agent, similar to sugars and involved in human metabolism, primarily enhances image clarity. Therefore, lesions with high metabolic activity require energy and absorb large amounts of the imaging agent. Consequently, the imaging agent accumulates in these lesions, which is visualized by PET / CT imaging, allowing accurate lesion detection. PET / CT results for lung examinations include two images: the CT image represents the entire lung, while the PET image marks the lesions. The computational methods used for CT and PET are the same.

[0003] Currently, CT and PET are commonly used methods for examining lung lesions, providing high accuracy. However, much old lung imaging data exists. Due to issues like pixel loss caused by improper storage or low image resolution resulting from limitations in existing imaging technology, this data is less reliable and difficult to accurately analyze lesions. This requires restoration of these old lung imaging data to restore high-definition lung images.

[0004] In existing technologies, conventional neural network models are used to directly analyze old lung image data. The data volume is too small to achieve medical-grade accuracy, and labeling also consumes a lot of resources (generally requiring the labeler to have a certain level of medical knowledge), resulting in a waste of human resources and often unable to achieve accurate repair effects.

[0005] Therefore, those skilled in the art are in urgent need of developing a new technical solution to solve the above problems. Summary of the Invention

[0006] To overcome the problems existing in the related art, the present invention discloses a lung image restoration model training method, a restoration method and an apparatus for PET / CT imaging.

[0007] According to a first aspect of the disclosed embodiments of the present invention, a method for training a lung image restoration model for PET / CT imaging is provided, the method comprising:

[0008] Acquire a sample lung image obtained by PET / CT imaging, and divide the sample lung image into n pieces of first image data according to layers of cross-sectional scanning;

[0009] Performing resolution reduction processing on the first image data to obtain n pieces of second image data;

[0010] Selecting four adjacent second image data, and obtaining forward feature parameters and reverse feature parameters of the four adjacent second image data according to the patch array of the global special token and the patch array of the universal human body token;

[0011] Correcting the forward feature parameters and the reverse feature parameters through a long short-term memory network to obtain corrected forward target feature parameters and reverse target feature parameters;

[0012] performing initial amplification processing on each piece of the second image data by cubic interpolation to obtain n pieces of third image data;

[0013] Obtaining a compensation value for each third image data according to a self-attention mechanism, a cross-attention mechanism of a forward target feature parameter, and a cross-attention mechanism of a reverse target feature parameter;

[0014] obtaining a target loss function according to the compensation value, the first image data, and the second image data;

[0015] The target loss function is back-propagated to obtain a lung image restoration model.

[0016] Optionally, performing resolution reduction processing on the first image data to obtain n pieces of second image data includes:

[0017] The n first image data are counted as P1, P2, ..., P n , wherein the resolution of each first image data is w×h;

[0018] Let u = w / 2, v = h / 2 respectively, perform resolution reduction processing on the first image data, and obtain the second image data x1, x2, ..., x n , where the resolution of the second image data is u×v.

[0019] Optionally, the selecting of four adjacent second image data and obtaining the forward feature parameters and the reverse feature parameters of the four adjacent second image data according to the patch array of the global special token and the patch array of the universal human body token includes:

[0020] Get four adjacent second image data;

[0021] Each second image data is divided into m*m patches through the C-ViT model, and each patch contains p*p feature vector values;

[0022] Generate a patch array PTK of a global special token and a patch array BTK of a universal human body token, so that each second image data contains (2+m*m) patches and the four second image data contain (5+4*m*m) patches;

[0023] Obtain the forward feature parameters BR of the four adjacent second image data i =[BTK i PTK i-2 PTK i-1 PTK i PTK i+1 ] and reverse characteristic parameter CR i =[BTK i PTK i-1 PTK i PTK i+1 PTK i+2 ].

[0024] Optionally, performing initial amplification processing on each piece of the second image data by cubic interpolation to obtain n pieces of third image data includes:

[0025] Each of the second image data x is interpolated by cubic interpolation. i Perform initial magnification processing to obtain n third image data K i , where each third image data K i Contains 2m*2m patches.

[0026] Optionally, obtaining the compensation value of the third image data according to the self-attention mechanism of each third image data, the cross-attention mechanism of the forward target feature parameter, and the cross-attention mechanism of the reverse target feature parameter includes:

[0027] According to each third image data K i The self-attention mechanism is calculated in the 7*7 area around the patch;

[0028] Forward target feature parameters BY i and reverse target characteristic parameter BZ i Perform cross-attention mechanism calculations;

[0029] For the second image data x i-2 、xi-1 、x i+1 、x i+2 The 3*3 area is used for cross attention mechanism calculation;

[0030] Stitch all corrected patches to obtain the compensation value dx of the third image data i .

[0031] Optionally, obtaining a target loss function according to the compensation value, the first image data, and the second image data includes:

[0032] According to the compensation value dx i and the second image data x i Acquire the fourth image data Ax i =x i -dx i ;

[0033] According to the fourth image data Ax i With the first image data P i The target loss function is obtained by summing the absolute values of the error values between each pixel. , where j and k represent the pixel at the jth row and kth column.

[0034] According to a second aspect of the disclosed embodiments of the present invention, a method for restoring lung images using PET / CT imaging is provided. The method is based on the lung image restoration model described in the first aspect of the disclosed embodiments of the present invention, and the method comprises:

[0035] Acquiring fifth image data of the lung to be repaired imaged by PET / CT imaging;

[0036] The fifth image data is used as input of the lung image restoration model, and the restored sixth image data is obtained according to the output result of the lung image restoration model.

[0037] According to a third aspect of the disclosed embodiments of the present invention, there is provided a PET / CT imaging lung image restoration model training device, the device comprising:

[0038] a sample image segmentation module, which obtains a sample lung image obtained by PET / CT imaging and divides the sample lung image into n pieces of first image data according to layers of cross-sectional scanning;

[0039] a resolution reduction processing module, connected to the sample image segmentation module, performing resolution reduction processing on the first image data to obtain n pieces of second image data;

[0040] a feature parameter extraction module, connected to the resolution reduction processing module, selecting four adjacent second image data, and obtaining forward feature parameters and reverse feature parameters of the four adjacent second image data based on the patch array of the global special token and the patch array of the universal human body token;

[0041] A feature parameter correction module is connected to the feature parameter extraction module, and corrects the forward feature parameters and the reverse feature parameters through a long short-term memory network to obtain corrected forward target feature parameters and reverse target feature parameters;

[0042] an initial amplification processing module, connected to the characteristic parameter correction module, performing initial amplification processing on each piece of the second image data by cubic interpolation to obtain n pieces of third image data;

[0043] a compensation value acquisition module, connected to the initial amplification processing module, for acquiring a compensation value of the third image data according to a self-attention mechanism of each third image data, a cross-attention mechanism of a forward target feature parameter, and a cross-attention mechanism of a reverse target feature parameter;

[0044] a loss function calculation module, connected to the compensation value acquisition module, and acquiring a target loss function according to the compensation value, the first image data, and the second image data;

[0045] The repair model acquisition module is connected to the loss function calculation module, and performs backpropagation on the target loss function to obtain a lung image repair model.

[0046] According to a fourth aspect of the disclosed embodiments of the present invention, there is provided a lung image restoration device for PET / CT imaging, the device being based on the lung image restoration model according to the first aspect of the disclosed embodiments of the present invention, the device comprising:

[0047] a module for acquiring an image to be repaired, for acquiring fifth image data of an image of the lung to be repaired from PET / CT imaging;

[0048] The clear image acquisition module is connected to the image acquisition module to be repaired, uses the fifth image data as the input of the lung image repair model, and obtains the repaired sixth image data according to the output result of the lung image repair model.

[0049] In summary, the present invention discloses a lung image restoration model training method, restoration method, and device for PET / CT imaging. The model training method includes: dividing a sample lung image into n first image data and performing resolution reduction to obtain n second image data; selecting four adjacent second image data, and obtaining forward feature parameters and reverse feature parameters based on a patch array of a global special token and a patch array of a universal human body token; performing corrections through a long short-term memory network to obtain forward target feature parameters and reverse target feature parameters; obtaining n third image data and compensation values; obtaining a target loss function based on the compensation values, the first image data, and the second image data; and obtaining a lung image restoration model. By generating a patch array of universal human body tokens, this feature vector is shared in all image calculations, thereby finding a balance between the complexity and accuracy of machine learning. The lung image restoration model obtained in this way not only ensures the accuracy of the image restoration results, but also reduces the time and cost spent on training a large amount of sample data.

[0050] Other features and advantages disclosed in the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are used to provide a further understanding of the present disclosure and constitute a part of the specification. Together with the following detailed description, they are used to explain the present disclosure but do not constitute a limitation of the present disclosure. In the accompanying drawings:

[0052] Figure 1 is a flowchart illustrating a method for training a lung image restoration model for PET / CT imaging according to an exemplary embodiment;

[0053] Figure 2 is based on Figure 1 A schematic flow chart of a feature parameter extraction method is shown;

[0054] Figure 3 is based on Figure 1 A schematic flow chart of a feature parameter extraction method is shown;

[0055] Figure 4 is based on Figure 3 A flow chart of a characteristic parameter correction method is shown;

[0056] Figure 5 is a flowchart illustrating a method for restoring lung images using PET / CT imaging according to an exemplary embodiment;

[0057] Figure 6 This is a structural block diagram of a lung image restoration model training device for PET / CT imaging according to an exemplary embodiment;

[0058] Figure 7 The figure is a structural block diagram of a lung image restoration device for PET / CT imaging according to an exemplary embodiment. DETAILED DESCRIPTION

[0059] The following is a detailed description of the specific embodiments disclosed in the present invention in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0060] PET, short for Positron Emission Tomography (PET), commonly referred to as "PET," is a relatively advanced nuclear medicine imaging technique. CT, short for Computed Tomography (CT), is a widely used and rapidly developing X-ray tomography technique. The integration of these two techniques into a single device, merging and displaying images of different natures, is known as PET / CT ("PET-CT"). Specifically, PET can visualize metabolic processes within the body. PET imaging is based on the detection of gamma-ray pairs indirectly emitted by positron-emitting radionuclides (also known as radiopharmaceuticals, radionuclides, or radiotracers). The tracer is injected intravenously with a bioactive molecule, typically a sugar used for cellular energy. The PET system's sensitive detectors capture gamma-ray emissions within the body, and software is used to triangulate the emission sources, creating a three-dimensional computed tomography image of the tracer concentration within the body.

[0061] PET / CT imaging technology, due to its unique imaging technology, is widely used in tumor screening, especially in lung cancer diagnosis, and is currently one of the most accurate imaging examination technologies. In the disclosed embodiments of the present invention, a lung image restoration model is used to restore unclear lung image data due to improper storage or damage, thereby obtaining clear lung image data. This facilitates medical personnel to diagnose diseases based on lung image data.

[0062] Figure 1 FIG. 1 is a flow chart of a method for training a lung image restoration model for PET / CT imaging according to an exemplary embodiment. Figure 1 As shown, the method includes:

[0063] In step 101, a sample lung image obtained by PET / CT imaging is acquired, and the sample lung image is divided into n pieces of first image data according to layers of cross-sectional scanning.

[0064] For example, it can be understood that sample lung images are selected to train the lung image restoration model. Before the training begins, based on the characteristics of the lung structure, the sample lung images of PET / CT imaging are divided into n first image data according to the tomographic images. Characteristics of the lung structure: The lung structure is physically highly correlated with other levels and has pathways. Bones, soft tissues, body fluids, fat, etc. are physically interconnected. When the information of this layer is less, information from other levels can be extracted and fused. For example, there can be no muscle in the center of the bone, and the middle part between two layers of bones is mostly bone tissue. Therefore, by dividing the sample lung images into layers, it is possible to obtain first image data of n layers that are correlated with each other and connected by pathways.

[0065] In step 102, the first image data is subjected to resolution reduction processing to obtain n pieces of second image data.

[0066] For example, it can be understood that steps 101 and 102 are the process of creating data and labels. The sample lung image and the first image data after segmentation are both clear lung images. In the disclosed embodiment of the present invention, the clear lung image is downgraded to simulate old lung image data (i.e., blurred images that need to be repaired) for training. At the same time, the clear first image data is stored as a label for evaluating the target variable during the training process.

[0067] Specifically, the first image data is subjected to resolution reduction processing to obtain n pieces of second image data, including: counting the n pieces of first image data as P1, P2, ..., P n , where the resolution of each first image data is w×h; let u=w / 2, v=h / 2 respectively, perform resolution reduction processing on the first image data, and obtain the second image data x1, x2, ..., x n , where the resolution of the second image data is u×v.

[0068] For example, a high-definition lung image (i.e., n first image data) is selected for resolution reduction compression processing. If the compression ratio is a non-integer, it can be calculated by linear interpolation. Preferably, the pixel resolution of the image at each level is divided by 2, that is, even-numbered row pixels and even-numbered column pixels are discarded. If the resolution of each first image data is w×h, the resolution of the second image data after resolution reduction processing is u×v, where u=w / 2, v=h / 2, and the calculation result is rounded up. In addition, it can be understood that when scanning an image, an axis is usually used as a reference, such as the sagittal axis, the coronal axis, the vertical axis, etc. Any of the axes is used as a reference for scanning and the layers are divided along the axis direction, and the subsequent resolution reduction processing is the same.

[0069] In step 103, four adjacent second image data are selected, and the forward feature parameters and reverse feature parameters of the four adjacent second image data are obtained according to the patch array of the global special token and the patch array of the universal human body token.

[0070] For example, the transformer model is selected in the embodiment disclosed in the present invention. Considering that when extracting the feature parameters of the image for training, the more relevant feature parameters are extracted, the more accurate the training results are. However, the larger the amount of input data, the more difficult it is for the model to converge. Therefore, 4 images are selected from the second image data for accurately calculating their correlation. Since the tissue structure of the human body has a unified law and a unified basic framework, it can be fed back to the local tissue through the overall feature parameters, such as the lung structure. When extracting the feature parameters of the 4 adjacent second image data, while dividing the image into several patch arrays, a patch array of global special tokens is generated based on the global specificity of the image, and then considering the similarity characteristics of the human body structure, a more general patch array of universal human body tokens is generated.

[0071] Specifically, Figure 2 is based on Figure 1 A flow chart of a feature parameter extraction method is shown in FIG. Figure 2 As shown, step 103 includes:

[0072] In step 1031, four adjacent second image data are obtained.

[0073] For example, Figure 3 As shown, four adjacent second image data are selected to generate forward feature parameters and reverse feature parameters.

[0074] In step 1032, each second image data is divided into m*m patches using the C-ViT model, and each patch contains p*p feature vector values.

[0075] For example, the C-ViT model is an improved ViT, which divides each second image data into m*m patches. This step is also the process of dividing the original second image data into several small squares. Subsequently, each patch needs to be projected into a vector of fixed length and sent to the encoder, that is, each patch contains p*p feature vector values.

[0076] In step 1033, a patch array PTK of a global special token and a patch array BTK of a universal human body token are generated so that each second image data contains (2+m*m) patches and the four second image data contain (5+4*m*m) patches.

[0077] In step 1034, the forward feature parameters BR of the four adjacent second image data are obtained. i =[BTK i PTK i-2 PTK i-1 PTK i PTK i+1 ] and reverse characteristic parameter CR i =[BTK i PTK i-1 PTK i PTK i+1 PTK i+2 ].

[0078] For example, since it involves image classification, a special token is added to the input sequence, and the output corresponding to the token is the final category prediction. In the existing solution, it is necessary to generate a patch array of a global special token for this image (denoted as PTK). After adding, each second image data contains (1+m*m) patches. In the embodiment disclosed in the present invention, considering the great similarity of human body structure, a more general human body token patch array (denoted as BTK) is generated. This feature vector array is shared when calculating all images, and finally each second image data contains (2+m*m) patches. It can be understood that when the 4 second image data are synthesized, due to different PTKs and the same BTK, the 4 second image data contain (5+4*m*m) patches, and the forward feature parameter BR i =[BTK i PTK i-2 PTK i-1 PTK i PTK i+1 ], and the reverse characteristic parameter CR i =[BTK i PTK i-1 PTK i PTK i+1 PTK i+2 ].

[0079] In step 104, the forward feature parameters and the reverse feature parameters are corrected by using a long short-term memory network to obtain corrected forward target feature parameters and reverse target feature parameters.

[0080] For example, in the simple mode, the forward characteristic parameter BR is obtained i and reverse characteristic parameter CR iThen the subsequent decoder process can be entered. In the preferred embodiment disclosed by the present invention, in order to further improve the calculation accuracy and make the final trained model have the ability to consider longer-term parameters, a small LSTM (long short-term memory network) is used to fuse earlier parameters, observe longer tissue connections, and calculate the forward feature parameter BR i and reverse characteristic parameter CR i Make corrections to obtain the corrected forward target feature parameters BY i and reverse target characteristic parameter BZ i . Specifically, Figure 4 As shown in Figure 3, the correction process of the long short-term memory network for the forward feature parameters and the reverse feature parameters is shown.

[0081] In step 105 , each piece of second image data is initially amplified by cubic interpolation to obtain n pieces of third image data.

[0082] For example, in the decoder process, each second image data after resolution reduction is initially amplified to restore the resolution of the second image data to w×h, but the image clarity is not restored to the level before resolution reduction. Therefore, the cubic interpolation method is used to enlarge each second image data x i Perform initial magnification processing to obtain n third image data K i , where each third image data K i It contains 2m*2m patches, which can better reflect the authenticity, interactivity and smoothness of human body structure.

[0083] In step 106, a compensation value of the third image data is obtained according to the self-attention mechanism of each third image data, the cross-attention mechanism of the forward target feature parameter, and the cross-attention mechanism of the reverse target feature parameter.

[0084] Specifically, according to the self-attention mechanism of each third image data, the cross-attention mechanism of the forward target feature parameter and the cross-attention mechanism of the reverse target feature parameter, the compensation value of the third image data is obtained, including: according to each third image data K i The self-attention mechanism is calculated for the 7*7 area around the patch; the positive target feature parameter BY i and reverse target characteristic parameter BZ i Perform cross attention mechanism calculation; for the second image data x i-2 、x i-1 、x i+1 、x i+2 The 3*3 area is used for cross attention mechanism calculation; all the corrected patches are spliced together to obtain the compensation value dx of the third image data i .

[0085] For example, each third image data K i The patch needs to consider the self-attetion of the 7*7 area around it, and BY i +BZ i (a total of ten patches) cross attention, and x i-2 、x i-1 、x i+1 、x i+2 Cross attention in the 3*3 area, concatenating all patches is recorded as dx i Among them, 7*7 and 3*3 are set to simplify the training process and time. The range can also be increased to improve accuracy, but it will increase training time and resource overhead.

[0086] In step 107 , a target loss function is obtained according to the compensation value, the first image data, and the second image data.

[0087] Specifically, obtaining the target loss function according to the compensation value, the first image data and the second image data includes: according to the compensation value dx i and the second image data x i Acquire the fourth image data Ax i =x i -dx i According to the fourth image data Ax i With the first image data P i The target loss function L is obtained by summing the absolute values of the errors between the pixels. It can be understood that the fourth image data Ax i is the corrected image data obtained after the above model training, and the difference between the image data and the real first image data is P i There is still a certain error, so it is necessary to calculate the target loss function , where j and k represent the pixel at the jth row and kth column.

[0088] In step 108, the target loss function is back-propagated to obtain a lung image restoration model.

[0089] For example, backpropagation is performed using a loss function, with the goal of reducing the loss. After multiple iterations of training, model parameters are obtained, ultimately resulting in the corresponding target model (i.e., the lung image restoration model) and parameters. Using these target model and parameters, the old image to be restored can be restored.

[0090] Figure 5 FIG. 1 is a flow chart of a method for restoring lung images using PET / CT imaging according to an exemplary embodiment. Figure 5As shown, based on the lung image restoration model, the method includes:

[0091] In step 501 , fifth image data of a lung image to be repaired obtained by PET / CT imaging is acquired.

[0092] In step 502, the fifth image data is used as input to a lung image restoration model, and restored sixth image data is obtained according to an output result of the lung image restoration model.

[0093] Figure 6 FIG. 1 is a structural block diagram of a PET / CT lung image restoration model training device according to an exemplary embodiment. Figure 6 As shown, the apparatus 600 includes:

[0094] The sample image segmentation module 610 obtains a sample lung image obtained by PET / CT imaging and divides the sample lung image into n pieces of first image data according to the layers of cross-sectional scanning;

[0095] A resolution reduction processing module 620 is connected to the sample image segmentation module 610 and performs resolution reduction processing on the first image data to obtain n pieces of second image data;

[0096] A feature parameter extraction module 630 is connected to the resolution reduction processing module 620 and selects four adjacent second image data. Based on the patch array of the global special token and the patch array of the universal human body token, it obtains the forward feature parameters and the reverse feature parameters of the four adjacent second image data.

[0097] The feature parameter correction module 640 is connected to the feature parameter extraction module 630 and corrects the forward feature parameter and the reverse feature parameter through a long short-term memory network to obtain the corrected forward target feature parameter and the reverse target feature parameter;

[0098] An initial magnification processing module 650 is connected to the characteristic parameter correction module 640 and performs initial magnification processing on each second image data by cubic interpolation to obtain n third image data;

[0099] a compensation value acquisition module 660 connected to the initial magnification processing module 650, for acquiring a compensation value for each third image data based on a self-attention mechanism, a cross-attention mechanism of a forward target feature parameter, and a cross-attention mechanism of a reverse target feature parameter;

[0100] a loss function calculation module 670 connected to the compensation value acquisition module 660, and acquiring a target loss function according to the compensation value, the first image data, and the second image data;

[0101] The restoration model acquisition module 680 is connected to the loss function calculation module 670, and performs backpropagation on the target loss function to obtain a lung image restoration model.

[0102] Figure 7 FIG. 1 is a structural block diagram of a lung image restoration device for PET / CT imaging according to an exemplary embodiment. Figure 7 As shown, the apparatus 700 includes:

[0103] The image acquisition module 710 for repairing obtains fifth image data of the lung image to be repaired from PET / CT imaging;

[0104] The clear image acquisition module 720 is connected to the image acquisition module 710 to be repaired, and uses the fifth image data as the input of the lung image repair model, and obtains the repaired sixth image data according to the output result of the lung image repair model.

[0105] In summary, the present invention discloses a lung image restoration model training method, restoration method, and device for PET / CT imaging. The model training method includes: dividing a sample lung image into n first image data and performing resolution reduction to obtain n second image data; selecting four adjacent second image data, and obtaining forward feature parameters and reverse feature parameters based on a patch array of a global special token and a patch array of a universal human body token; performing corrections through a long short-term memory network to obtain forward target feature parameters and reverse target feature parameters; obtaining n third image data and compensation values; obtaining a target loss function based on the compensation values, the first image data, and the second image data; and obtaining a lung image restoration model. By generating a patch array of universal human body tokens, this feature vector is shared in all image calculations, thereby finding a balance between the complexity and accuracy of machine learning. The lung image restoration model obtained in this way not only ensures the accuracy of the image restoration results, but also reduces the time and cost spent on training a large amount of sample data.

[0106] The preferred embodiments of the present disclosure are described in detail above in conjunction with the accompanying drawings. However, the present disclosure is not limited to the specific details of the above embodiments. Within the technical concept of the present disclosure, various simple modifications can be made to the technical solutions of the present disclosure, and these simple modifications all fall within the scope of protection of the present disclosure.

[0107] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any appropriate manner without contradiction. In order to avoid unnecessary repetition, the present disclosure will not further describe various possible combinations.

[0108] In addition, the various embodiments of the present disclosure may be arbitrarily combined, and as long as they do not violate the concept of the present disclosure, they should also be regarded as the contents disclosed by the present disclosure.

Claims

1. A method for training a lung image restoration model for PET / CT imaging, characterized in that: The method comprises: Acquire a sample lung image obtained by PET / CT imaging, and divide the sample lung image into n pieces of first image data according to layers of cross-sectional scanning; Performing resolution reduction processing on the first image data to obtain n pieces of second image data; Selecting four adjacent second image data, and obtaining forward feature parameters and reverse feature parameters of the four adjacent second image data according to the patch array of the global special token and the patch array of the universal human body token; Correcting the forward feature parameters and the reverse feature parameters through a long short-term memory network to obtain corrected forward target feature parameters and reverse target feature parameters; performing initial amplification processing on each piece of the second image data by cubic interpolation to obtain n pieces of third image data; Obtaining a compensation value for each third image data according to a self-attention mechanism, a cross-attention mechanism of a forward target feature parameter, and a cross-attention mechanism of a reverse target feature parameter; obtaining a target loss function according to the compensation value, the first image data, and the second image data; Back-propagating the target loss function to obtain a lung image restoration model; The selecting of four adjacent second image data and obtaining forward feature parameters and reverse feature parameters of the four adjacent second image data according to the patch array of the global special token and the patch array of the universal human body token includes: obtaining the four adjacent second image data; dividing each second image data into m*m patches using the C-ViT model, each patch containing p*p feature vector values; Generate a patch array PTK of a global special token and a patch array BTK of a universal human body token, so that each second image data contains (2+m*m) patches and the four second image data contain (5+4*m*m) patches; obtain the forward feature parameters BR of the four adjacent second image data i =[BTK i PTK i-2 PTK i-1 PTK i PTK i+1 ] and reverse characteristic parameter CR i =[BTK i PTK i-1 PTK i PTK i+1 PTK i+2 ].

2. The lung image restoration model training method for PET / CT imaging according to claim 1, characterized in that: The performing resolution reduction processing on the first image data to obtain n pieces of second image data includes: The n first image data are counted as P1, P2, ..., P n , wherein the resolution of each first image data is w×h; Let u = w / 2, v = h / 2 respectively, perform resolution reduction processing on the first image data, and obtain the second image data x1, x2, ..., x n , where the resolution of the second image data is u×v.

3. The lung image restoration model training method for PET / CT imaging according to claim 1, characterized in that: The performing initial amplification processing on each piece of the second image data by cubic interpolation to obtain n pieces of third image data includes: Each of the second image data x is interpolated by cubic interpolation. i Perform initial magnification processing to obtain n third image data K i , where each third image data K i Contains 2m*2m patches.

4. The lung image restoration model training method for PET / CT imaging according to claim 1, characterized in that: The obtaining of the compensation value of the third image data according to the self-attention mechanism of each third image data, the cross-attention mechanism of the forward target feature parameter, and the cross-attention mechanism of the reverse target feature parameter includes: According to each third image data K i The self-attention mechanism is calculated in the 7*7 area around the patch; Forward target feature parameters BY i and reverse target characteristic parameter BZ i Perform cross-attention mechanism calculations; For the second image data x i-2 、x i-1 、x i+1 、x i+2 The 3*3 area is used for cross attention mechanism calculation; Stitch all corrected patches to obtain the compensation value dx of the third image data i .

5. The lung image restoration model training method for PET / CT imaging according to claim 4, characterized in that: The obtaining of a target loss function according to the compensation value, the first image data, and the second image data includes: According to the compensation value dx i and the second image data x i Acquire the fourth image data Ax i =x i -dx i ; According to the fourth image data Ax i With the first image data P i The target loss function is obtained by summing the absolute values of the error values between each pixel. , where j and k represent the pixel at the jth row and kth column.

6. A method for restoring lung images from PET / CT imaging, characterized in that: The method is based on the lung image restoration model according to any one of claims 1 to 5, and the method comprises: Acquiring fifth image data of the lung to be repaired imaged by PET / CT imaging; The fifth image data is used as input of the lung image restoration model, and the restored sixth image data is obtained according to the output result of the lung image restoration model.

7. A PET / CT lung image restoration model training device, characterized in that: The device comprises: a sample image segmentation module, which obtains a sample lung image obtained by PET / CT imaging and divides the sample lung image into n pieces of first image data according to layers of cross-sectional scanning; a resolution reduction processing module, connected to the sample image segmentation module, performing resolution reduction processing on the first image data to obtain n pieces of second image data; a feature parameter extraction module, connected to the resolution reduction processing module, selecting four adjacent second image data, and obtaining forward feature parameters and reverse feature parameters of the four adjacent second image data based on the patch array of the global special token and the patch array of the universal human body token; A feature parameter correction module is connected to the feature parameter extraction module, and corrects the forward feature parameters and the reverse feature parameters through a long short-term memory network to obtain corrected forward target feature parameters and reverse target feature parameters; an initial amplification processing module, connected to the characteristic parameter correction module, performing initial amplification processing on each piece of the second image data by cubic interpolation to obtain n pieces of third image data; a compensation value acquisition module, connected to the initial amplification processing module, for acquiring a compensation value of the third image data according to a self-attention mechanism of each third image data, a cross-attention mechanism of a forward target feature parameter, and a cross-attention mechanism of a reverse target feature parameter; a loss function calculation module, connected to the compensation value acquisition module, and acquiring a target loss function according to the compensation value, the first image data, and the second image data; a restoration model acquisition module, connected to the loss function calculation module, performing backpropagation on the target loss function to acquire a lung image restoration model; The selecting of four adjacent second image data and obtaining forward feature parameters and reverse feature parameters of the four adjacent second image data according to the patch array of the global special token and the patch array of the universal human body token includes: obtaining the four adjacent second image data; dividing each second image data into m*m patches using the C-ViT model, each patch containing p*p feature vector values; Generate a patch array PTK of a global special token and a patch array BTK of a universal human body token, so that each second image data contains (2+m*m) patches and the four second image data contain (5+4*m*m) patches; obtain the forward feature parameters BR of the four adjacent second image data i =[BTK i PTK i-2 PTK i-1 PTK i PTK i+1 ] and reverse characteristic parameter CR i =[BTK i PTK i-1 PTK i PTK i+1 PTK i+2 ].

8. A lung image restoration device for PET / CT imaging, characterized in that: The device is based on the lung image restoration model according to any one of claims 1 to 5, and comprises: a module for acquiring an image to be repaired, for acquiring fifth image data of an image of the lung to be repaired from PET / CT imaging; The clear image acquisition module is connected to the image acquisition module to be repaired, uses the fifth image data as the input of the lung image repair model, and obtains the repaired sixth image data according to the output result of the lung image repair model.

Images