JP2023133132A - Transformer feature sharing-based pet image reconstruction method, apparatus and device and medium - Google Patents

Abstract

To provide transformer feature sharing-based PET image reconstruction method, apparatus and device, and medium.SOLUTION: The invention comprises the steps: obtaining a back projection image containing PET original data information; and obtaining a PET image with a pre-trained PET image reconstruction network model based on transformer feature sharing. The PET image reconstruction network model is composed of two groups of encoders-decoders, one group establishes mapping from a PET back projection image to a PET reconstruction image, and the other group establishes mapping from the PET back projection image to a prior information image. Meanwhile, the two groups of encoders-decoders are optimized, so that noise in the target PET image is reduced by utilizing a prior information image, and image detail information is reserved at the same time. And a transform unit is used between the two groups of encoders, so that parameter sharing of the autonomous learning encoders in a reconstruction network training process is realized, reconstruction errors are further reduced, and the quality of the reconstructed image is improved.SELECTED DRAWING: Figure 1

Description

The present invention relates to the field of medical imaging technology, and in particular to PET image reconstruction methods, apparatus, devices and media based on transformer feature sharing.

Positron emission tomography (PET) is a medical imaging technology that can simultaneously provide biological function/metabolic information as well as morphological/anatomical structural information of the human body, and is useful for tumor diagnosis, neuropsychiatric disorders, and cardiac disease. It is widely used for clinical applications such as diagnosis, treatment, and evaluation of vascular diseases. The process of producing a PET image involves injecting a radioactive tracer into the patient before the scan, and as the tracer takes part in physiological metabolism in the body, it decays to produce positrons, which are then released into nearby tissues. generating a high-energy photon pair that annihilates with an electron and moves in the opposite direction; detecting a coincidence of the photon pair with two detectors at opposite positions and recording a coincidence response line; A certain number of matched response lines are collected and arranged as three-dimensional PET raw data according to position and angle information, and after correcting the raw data, a reconstruction algorithm is used to express the metabolic intensity of each tissue in the body. obtaining an original PET image.

With the upgrade of computer hardware and the rapid development of deep learning technology, the construction of a PET image reconstruction network based on deep learning has become a hot topic in recent years. Over time, advances have been made in techniques such as establishing end-to-end mapping between low-quality and high-quality PET images using encoders and decoders, and establishing mapping between PET raw data and PET images. ing. However, existing reconstruction networks based on a single encoding-decoding structure rely on prior knowledge about PET images present in medical images of other modalities (e.g. computed tomography CT, magnetic resonance imaging MRI) in the reconstruction process. Since it is not possible to introduce PET images, there is still room for improvement in the quality of PET images.

Multitask deep neural networks consisting of multiple sets of encoders and decoders have already been successfully applied to tasks such as segmentation and detection of natural and medical images. The network structure can realize the introduction of prior information of subtasks into the main task, and the feature sharing methods commonly used in the network structure are encoder sharing, or encoder parameter sharing using convolution-based spatial and channel attention mechanisms. It is learning. Compared to encoder-shared forced parameter sharing, a learnable parameter sharing mechanism can lead network training to a better solution. Nevertheless, existing convolution-based parameter sharing learning lacks globality and can only realize local parameter sharing by calculating attention coefficients between local features. Meanwhile, a network structure called a transformer, which replaces the traditional convolutional structure, is making groundbreaking progress in fields such as text translation and speech recognition. The transformer consists of a global multi-head self-attention mechanism and a multilayer perceptron.

An object of the present invention is to provide a PET image reconstruction method, apparatus, device and medium based on transformer feature sharing to address the shortcomings of existing PET image reconstruction techniques based on deep learning.

Based on the structure and function of transformers, the present invention constructs a novel PET image reconstruction network model based on transformer feature sharing. The model consists of two encoder-decoder sets, one set of encoder-decoders establishes the mapping between the PET backprojection image and the PET reconstructed image, and the other set of encoder-decoders establishes the mapping between the PET backprojection image and the PET reconstructed image. To reduce the reconstruction error of PET image reconstruction network using prior knowledge in prior information images by establishing a mapping between images and prior information images and optimizing two sets of encoder-decoders simultaneously. Realize. A transformer unit is used instead of a convolution-based attention mechanism between two sets of encoders to realize self-learning of encoder parameter sharing in the training process of the reconstruction network, specifically, interleaving based on the transformer unit. By designing the connections and realizing the calculation of global channel and spatial attention coefficients, the parameter sharing of the encoder is self-learning in the training process of the reconstruction network, further reducing the reconstruction error of the network, and improving the reconstruction PET Improve image quality.

The present invention is realized by the following technical solutions.
A PET image reconstruction method based on transformer feature sharing, the method comprising:
obtaining a backprojection image containing PET raw data information;
inputting the back-projection image containing the PET raw data information into a pre-trained transformer feature sharing based PET image reconstruction network model to obtain a PET image.

The PET image reconstruction network model based on transformer feature sharing includes a first encoder-decoder, a second encoder-decoder, and at least one transformer unit, the first encoder-decoder mapping and predicting a backprojected image. A second encoder-decoder is used to map the back-projection image to generate a predictive prior information image, which is used to generate a predicted prior information image from other images collected in the same batch as the PET images. A type of image of modality, the transformer unit connects the convolution unit in the first encoder-decoder, the second encoder-decoder encoding structure, and the convolution unit in the first encoder-decoder, second encoder-decoder encoding structure. and feed the processed output to the next convolution unit in the first encoder-decoder, second encoder-decoder encoding structure, each transformer unit being connected in turn to an attention mechanism module and a multi-layer Contains a perceptron module,
The PET image reconstruction network model based on pre-trained transformer feature sharing comprises a predicted PET image generated by a first encoder-decoder, a predicted a priori information generated by a second encoder-decoder, based on a training data set. Minimizing the image-to-truth loss and maximizing the structural similarity between the predicted PET image generated by the first encoder-decoder and the predicted a priori information image generated by the second encoder-decoder. This can be obtained by training for the purpose of

moreover,
Each set of training data in the training data set further includes a back-projected image including PET raw data information, a reconstructed PET image, and a priori information image.

Furthermore, the prior information image is a CT image or an MRI image acquired in the same batch as the PET image.

Furthermore, the back projection image containing the PET raw data information is
It is obtained by regularizing, attenuating, randomizing, and scattering correcting the PET raw data and then back-projecting it onto the image area to obtain a back-projected image containing the PET raw data information.

Furthermore, the PET image is
It is obtained by repeatedly reconstructing physically corrected PET raw data to obtain a PET image.

Further, the attention mechanism module obtains a query vector group, a keyword vector group, a feature vector group based on the output splicing features of the convolution unit in the first encoder-decoder, second encoder-decoder encoding structure, and obtains the query vector group. Perform an inner product operation on each vector in the keyword vector group with all vectors in the keyword vector group in order, obtain the attention parameters of the corresponding vector for itself and all other vectors by the inner product operation, and use the obtained attention parameters to calculate the features. Calculate the weighted sum of all vectors in the vector group, complete the update of the corresponding feature vector, then update all feature vectors in order, splice all updated feature vectors, and calculate the output feature. Then, the result of adding the output features and the output splicing features of the convolution unit is used to be the final output of the attention mechanism module.

A PET image reconstruction device, the device comprising:
an image acquisition module for obtaining backprojection images containing PET raw data information;
a reconstruction module for inputting the back-projection image containing the PET raw data information into a pre-trained transformer feature sharing based PET image reconstruction network model to obtain a PET image.

Furthermore, the device includes:
Based on the training data set, minimize the loss between the predicted PET image generated by the first encoder-decoder, the predicted a priori information image generated by the second encoder-decoder, and the true value, and generate the predicted PET image by the first encoder-decoder. a PET image based on pre-trained transformer feature sharing by training with the aim of maximizing the structural similarity between the predicted PET image to be generated and the predicted a priori information image generated by the second encoder-decoder; It further includes a training module for obtaining a reconstructed network model.

An electronic device including a memory, a processor, and a computer program stored in the memory and operable on the processor, wherein when the processor executes the computer program, the PET image reconstruction method is performed.

A storage medium containing computer-executable instructions that, when executed by a computer processor, perform the PET image reconstruction method.

The present invention has the following beneficial effects compared to existing PET image reconstruction methods. The present invention realizes mapping from PET backprojection images to PET reconstructed images and medical images of other modalities (e.g., CT, MRI) by training two encoder-decoder sets simultaneously, The prior information in is used to improve the quality of PET image reconstruction and improve the generalization ability of the reconstruction network. The present invention proposes to use a transformer unit instead of a convolution layer and connect two sets of encoders in an interleaved manner to realize an attention mechanism to calculate globality between different channels and different spatial location features. , outperforms existing attention mechanisms realized by convolutional networks based on local features. Furthermore, by optimizing one loss function, the present invention realizes self-learning on how to share features between two sets of encoders, and compared to existing network structures that share one encoder, By selectively increasing useful information and discarding useless information, we approach an optimal network solution.

1 is a schematic diagram of the flow of a PET image reconstruction method based on transformer feature sharing provided by Example 1 of the present invention; FIG. FIG. 2 is a structural diagram of a PET image reconstruction network model based on transformer feature sharing provided by Example 1 of the present invention; FIG. 3 is a schematic diagram of the operation of the attention mechanism module provided by the embodiment 1 of the present invention; FIG. 2 is a comparison diagram of the reconstruction results of the PET image reconstruction method based on transformer feature sharing and the PET image reconstruction method based on a single encoder-decoder provided by Example 1 of the present invention. FIG. 2 is a structural schematic diagram of a PET reconstruction device provided by Example 2 of the present invention; FIG. 3 is a structural schematic diagram of an electronic device provided by Example 3 of the present invention;

Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. Where the following description refers to drawings, the same numbers in different drawings indicate the same or similar elements, unless stated otherwise. The embodiments described in the illustrative examples below do not represent all embodiments consistent with the present invention. On the contrary, they are merely examples of apparatus and methods consistent with certain aspects of the invention, as detailed in the appended claims.

The terminology used in the present invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

As used in this invention and the appended claims, the singular forms "a", "a", and "the" are intended to include the plural as well, unless the context clearly dictates otherwise. . It should also be understood that the term "and/or" as used herein refers to and encompasses any or all possible combinations of one or more of the associated listed items.

Note that in the present invention, various information may be explained using terms such as "first," "second," and "third," but these information are not limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, first information may also be referred to as second information, and similarly, second information may be referred to as first information without departing from the scope of the invention. Depending on the context, the word "if" as used herein may be interpreted as "when" or "with" or "as determined."

Embodiment 1: FIG. 1 is a flow diagram of a PET image reconstruction method based on transformer feature sharing provided by Embodiment 1 of the present invention, and as shown in FIG. includes the following steps:
Step 101: Obtain a backprojection image containing PET raw data information.
Specifically, the PET raw data is regularized, attenuated, randomized, and scatter-corrected, and then back-projected onto the image area to obtain a back-projected image containing the raw data information.

Step 102, inputting the back-projection image containing the PET raw data information into a pre-trained PET image reconstruction network model based on transformer feature sharing to obtain a PET image.

The PET image reconstruction network model based on transformer feature sharing includes a first encoder-decoder, a second encoder-decoder, and a transformer unit, and the first encoder-decoder maps a backprojected image to generate a predicted PET image. a second encoder-decoder is used to map the back-projection image to generate a predictive prior information image, said prior information image being acquired in the same batch as the PET images of other modalities; The transformer unit connects the convolution unit in the first encoder-decoder, second encoder-decoder encoding structure, and connects the convolution unit in the first encoder-decoder, second encoder-decoder encoding structure. The transformer unit is used to receive the output and feed the processed output to the next convolution unit in the first encoder-decoder, second encoder-decoder encoding structure, and each transformer unit is connected in turn to an attention mechanism module and a multilayer perceptron. Contains modules.

The encoder-decoder has a general neural network structure and is composed of a plurality of convolution units and deconvolution units.

The present invention establishes an encoder-decoder to generate a predicted PET image and a predicted prior information image by mapping, respectively, and uses the prior knowledge in the prior information image to reduce the noise in the target PET image while At the same time, by using the fact that the transformer unit is composed of a multi-head self-attention mechanism and a multi-layer perceptron, the global attention coefficient can be calculated, and between the two sets of encoders, the transformer unit By designing interleaved connections based on , and realizing the calculation of global channel and spatial attention coefficients, the encoder parameter sharing is self-learning in the training process of the reconstruction network, which further reduces the reconstruction error of the network. , improving the quality of reconstructed PET images.

The prior information image refers to one type of image of another modality that is collected in the same batch as the PET image, such as a CT image or an MRI image. Since the time interval between collecting the prior information image and collecting the PET image is short and structural changes in each organ and tissue in the body can be ignored, the prior information image and the PET image have the same anatomical structure. In addition, since the prior information image has a higher resolution than the PET image, the local smoothness of the prior information image is used to reduce internal noise of the organ or tissue corresponding to the PET image, preserving detailed structural information. can do.

The transformer unit may be one or more and may be used to connect any convolution unit in the first encoder-decoder, second encoder-decoder encoding structure, preferably as shown in FIG. FIG. 2 is a structural diagram of a PET image reconstruction network model based on transformer feature sharing provided by embodiment 1 of the invention, with the number of transformer units and convolution units in the first encoder-decoder and second encoder-decoder encoding structures; The number is the same and the sharing of encoder parameters for each set of convolution units is self-learned to maximize the quality of reconstructed PET images.

The output of the convolution unit in the encoder-decoder encoding structure is composed of a plurality of channels, and in a preferred embodiment, the attention mechanism module performs the operation processing as shown in FIG. 3 for each channel, specifically is as follows.
The output splicing features of the convolution unit in the first encoder-decoder and the second encoder-decoder encoding structure are determined by projection matrices W _i ^qry , W _i ^key , W _i ^val (where i is the i-th channel ), respectively, and project each vector in the query vector group as a query vector group, keyword vector group, and feature vector group, perform an inner product operation on each vector in the query vector group with all vectors in the keyword vector group in order, and calculate the corresponding vector by the inner product operation. Obtain the attention parameters for itself and all other vectors, use the obtained attention parameters to calculate the weighted sum of all vectors in the feature vector group, complete the update of the corresponding feature vector, and then The feature vectors of the other channels are updated one after another, and the steps are repeated to complete the update of the feature vector groups of other channels. The updated feature vector group of each channel is spliced, the output feature 1 is calculated by the projection matrix W ^o , and the result of adding the output feature 1 and the output splicing feature of the convolution unit is the final output of the attention mechanism module.

An update method for a feature vector of a vector of a single channel may be expressed as follows.
where f _i is the splicing feature output from a single channel of the convolution unit in the first encoder-decoder and the second encoder-decoder. f _i W _i ^qry , (W _i ^key ) ^τ f _i , f _i W _i ^val are three projection operations. softmax( ) is a nonlinear activation function and d is the dimension of the single channel feature vector group. The attention mechanism module proposed above connects two sets of encoders in an interleaved manner instead of a convolutional layer to realize an attention mechanism that calculates the globality between different channels and different spatial location features, and Outperforms existing attention mechanisms realized by feature-based convolutional networks.

The PET image reconstruction network model based on pre-trained transformer feature sharing comprises a predicted PET image generated by a first encoder-decoder, a predicted a priori information generated by a second encoder-decoder, based on a training data set. Minimizing the image-to-truth loss and maximizing the structural similarity between the predicted PET image generated by the first encoder-decoder and the predicted a priori information image generated by the second encoder-decoder. This can be obtained by training for the purpose of

Specifically, each set of training data in the training data set includes a back-projection image including PET raw data information, a reconstructed PET image, and a priori information image. The back-projection image containing PET raw data information becomes the input of the PET image reconstruction network model based on transformer feature sharing, and the reconstructed PET image is used as the learning label of the first encoder-decoder (by the first encoder-decoder). The prior information image becomes the learning label of the second encoder-decoder (the true value corresponding to the predicted prior information image generated by the second encoder-decoder).

The data serving as the learning label needs to express the distribution situation of the PET radioactive tracer in each organ in the body with the highest possible quality, and in this example, the PET image is
It is obtained by repeatedly reconstructing physically corrected PET raw data to obtain a PET image.

Furthermore, during training, a gradient optimization algorithm is used to minimize the loss between the predicted PET image generated by the first encoder-decoder, the predicted a priori information image generated by the second encoder-decoder, and the true value; The first encoder-decoder, which is interleaved by the transformer unit, maximizes the structural similarity between the predicted PET image generated by the encoder-decoder and the predicted a priori information image generated by the second encoder-decoder. 2 Update encoder-decoder parameters. The loss between the predicted PET image generated by the first encoder-decoder, the predicted a priori information image generated by the second encoder-decoder, and the true value is defined as follows.
however,
is the output of the first encoder-decoder, I ^PET is the learning label of the first encoder-decoder, that is, the PET image obtained by reconstruction,
is the output of the second encoder-decoder, and ^ICT is the learning label of the second encoder-decoder, which is a CT image in this example. SSIM( ) is a structural similarity index measurement function, which is used to calculate the structural similarity between the outputs of two encoder-decoder sets. β is a weighting factor and is used to adjust the importance of structural similarity to the loss function. Mapping from PET backprojection images to PET reconstructed images and medical images of other modalities (e.g., CT, MRI) is realized by training two sets of encoder-decoders with one loss function simultaneously, A priori information in the images is used to improve the quality of PET image reconstruction and improve the generalization ability of the reconstruction network. In addition, network training realizes self-learning on how to share features between two sets of encoders, and compared to existing network structures that share one encoder, useful information is selectively increased and useless information is increased. By discarding such information, the encoder-decoder output can be brought closer to the optimal solution.

FIG. 4 is a comparison of reconstruction results for a patient's PET whole body image by the PET image reconstruction method based on transformer feature sharing provided by the first embodiment of the present invention and the PET image reconstruction method based on a single encoder-decoder. Here, the one based on a single encoder-decoder is obtained by using PET back projection images as input and training conventional iteratively reconstructed images as learning labels, and the one based on a single encoder-decoder The reconstruction result is shown in FIG. 4(a), and the reconstruction result is very close to the conventional iterative reconstruction result. The reconstruction result using the multi-encoding-decoding network based on transformer feature sharing proposed by the present invention is shown in Fig. 4(b), and the noise in the generated PET image is shown in Fig. 4(a). It is clearly less and the contrast of the lung tumor (eg, indicated by the arrow) is not reduced and the quality of the PET reconstructed image is significantly improved.

Embodiment 2: Corresponding to the embodiment of the PET image reconstruction method based on transformer feature sharing, this embodiment further provides a PET image reconstruction device based on transformer feature sharing, and FIG. 5 is a structural schematic diagram of a PET reconstruction device provided by Example 2, and as shown in FIG. 5, the device includes an image acquisition module 110 and a reconstruction module 120;
The image acquisition module 110 is used to obtain backprojected images containing PET raw data information;
The reconstruction module 120 inputs the back-projection image containing the PET raw data information into a pre-trained PET image reconstruction network model based on transformer feature sharing to obtain a PET image.

The structure of the PET image reconstruction network model based on the transformer feature sharing is shown in FIG. 2 and will not be described here.

Furthermore, the apparatus minimizes the loss between the predicted PET image generated by the first encoder-decoder, the predicted a priori information image generated by the second encoder-decoder, and the true value based on the training data set; a transformer pre-trained by training with the aim of maximizing the structural similarity between the predicted PET image produced by one encoder-decoder and the predicted a priori image produced by a second encoder-decoder; It further includes a training module 130 for obtaining a PET image reconstruction network model based on feature sharing.

The PET image reconstruction device provided by this embodiment can significantly improve the quality of PET reconstructed images.

Embodiment 3: Corresponding to the embodiment of the PET image reconstruction method based on transformer feature sharing, this embodiment further provides an electronic device based on transformer feature sharing, and FIG. 6 shows an embodiment of the present invention. 6 is a structural schematic diagram of an electronic device provided by an embodiment of the present invention, as shown in FIG. a computer program; when the processor executes the computer program, the PET image reconstruction method based on transformer feature sharing is implemented.

The electronic device of the present invention may be placed in any device having data processing capability, and the any device having data processing capability may be, for example, a device or apparatus such as a computer.

As an apparatus in the logical sense, it is formed by reading corresponding computer program instructions in a non-volatile memory into memory and executing them by a processor of any device having data processing capacity in which it is located. In terms of hardware, as shown in FIG. 6, which is a hardware structural diagram of the electronic device of the present invention, in addition to the processor, memory, network interface, and nonvolatile memory shown in FIG. Any device with data processing capabilities may typically include other hardware, which will not be described here, depending on the actual functionality of any device with data processing capabilities.

Regarding the process of realizing the functions and effects of each unit of the electronic device, the process of realizing the corresponding step in the method can be specifically referred to, and the description thereof will be omitted here.

Since the electronic device embodiment substantially corresponds to the method embodiment, reference may be made to the description of the method embodiment for related parts. The electronic device embodiments described above are illustrative only, and units described as separate parts may or may not be physically separate, and parts described as units may or may not be physically separate. may or may not be a physical unit, i.e., co-located or distributed over multiple network units. In order to achieve the objective of the solution of the present invention, some or all modules may be selected among them according to actual needs. Those skilled in the art can understand and implement without creative efforts.

Embodiments of the present invention further provide a computer-readable storage medium storing a program, which when executed by a processor performs the PET image reconstruction method based on transformer feature sharing of the embodiments. be done.

The computer readable storage medium may be an internal storage unit of any device having data processing capabilities described in any of the embodiments above, such as a hard disk or memory. The computer readable storage medium may be any device having data processing capability, such as a plug-in hard disk installed in the device, a smart memory card (Smart Media Card, SMC), an SD card, etc. cards, flash memory cards, etc. Further, the computer-readable storage medium may include both internal storage units and external storage devices of any device that has data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data necessary for any device having the data processing capability, and is used to temporarily store data that has already been output or is to be output. It may also be used to store information in

It should be noted that what has been described above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the protection scope of the invention. You will understand something. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention. However, the scope of the invention is determined by the claims appended hereto.

Claims (10)

A PET image reconstruction method based on transformer feature sharing, the method comprising:
obtaining a backprojection image containing PET raw data information;
inputting the back-projection image containing the PET raw data information into a pre-trained transformer feature sharing based PET image reconstruction network model to obtain a PET image;
The PET image reconstruction network model based on transformer feature sharing includes a first encoder-decoder, a second encoder-decoder, and at least one transformer unit, the first encoder-decoder mapping a backprojected image. A second encoder-decoder is used to map the backprojection image to generate a predicted prior information image, the prior information image being collected in the same batch as the PET image, and The transformer unit connects the convolution unit in the first encoder-decoder, second encoder-decoder encoding structure, and performs the convolution in the first encoder-decoder, second encoder-decoder encoding structure. The transformer unit is used to receive the output of the unit and feed the processed output to the next convolution unit in the first encoder-decoder, second encoder-decoder encoding structure, and each transformer unit is connected in turn to the attention mechanism module and Contains a multilayer perceptron module,
The PET image reconstruction network model based on pre-trained transformer feature sharing comprises a predicted PET image generated by a first encoder-decoder, a predicted a priori information generated by a second encoder-decoder, based on a training data set. Minimizing the image-to-truth loss and maximizing the structural similarity between the predicted PET image generated by the first encoder-decoder and the predicted a priori information image generated by the second encoder-decoder. A PET image reconstruction method based on transformer feature sharing, characterized in that it is obtained by training for the purpose of. 1 . Each set of training data in the training data set further comprises a back-projection image including PET raw data information, a reconstructed PET image, and a priori information image. The method described in. 2. The method of claim 1, wherein the prior information image is a CT image or an MRI image acquired in the same batch as the PET image. The back-projection image containing the PET raw data information is
2. The method according to claim 1, wherein the PET raw data is regularized, attenuated, randomized, and scatter-corrected, and then back-projected onto an image area to obtain a back-projected image containing the PET raw data information. the method of. The PET image is
2. The method of claim 1, wherein the PET images are obtained by iteratively reconstructing physically corrected raw PET data. The attention mechanism module obtains a query vector group, a keyword vector group, and a feature vector group based on the output splicing features of the convolution unit in the first encoder-decoder and the second encoder-decoder encoding structure, and Perform an inner product operation on each vector with all vectors in the keyword vector group in order, obtain the attention parameters of the corresponding vector for itself and all other vectors by the inner product operation, and use the obtained attention parameters to group the feature vectors. Calculate the weighted sum of all the vectors in the process, complete the updating of the corresponding feature vectors, update all the feature vectors in order, splice all the updated feature vectors, and calculate the output features. 2. The method of claim 1, wherein the result of adding the output features and the output splicing features of the convolution unit is used to be the final output of the attention mechanism module. A PET image reconstruction device that implements the PET image reconstruction method according to claim 1, comprising:
an image acquisition module for obtaining backprojection images containing PET raw data information;
a reconstruction module for inputting the back-projection image containing the PET raw data information into a pre-trained PET image reconstruction network model based on transformer feature sharing to obtain a PET image. Image reconstruction device. Based on the training data set, minimize the loss between the predicted PET image generated by the first encoder-decoder, the predicted a priori information image generated by the second encoder-decoder, and the true value, and generate the predicted PET image by the first encoder-decoder. a PET image based on pre-trained transformer feature sharing by training with the aim of maximizing the structural similarity between the predicted PET image to be generated and the predicted a priori information image generated by the second encoder-decoder; 8. The apparatus of claim 7, further comprising a training module for obtaining a reconstructed network model. 7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and operable on the processor, when the processor executes the computer program, the PET according to any one of claims 1 to 6. An electronic device that performs an image reconstruction method. A storage medium comprising computer-executable instructions, the computer-executable instructions, when executed by a computer processor, performing a PET image reconstruction method according to any one of claims 1 to 6. Characteristic storage media.