WO2021097703A1

WO2021097703A1 - Image reconstruction method, apparatus and device, and storage medium

Info

Publication number: WO2021097703A1
Application number: PCT/CN2019/119681
Authority: WO
Inventors: 王珊珊; 郑海荣; 祁可翰; 刘新
Original assignee: 深圳先进技术研究院
Priority date: 2019-11-20
Filing date: 2019-11-20
Publication date: 2021-05-27

Abstract

An image reconstruction method, apparatus and device, and a storage medium. The method comprises: acquiring sampled data of a target object (S201); and inputting the sampled data into a trained deep learning network for processing to obtain a reconstructed image corresponding to the sampled data, wherein the sampled data is under-sampled frequency domain data in a non-Cartesian coordinate system acquired on the basis of a preset sampling mode (S202). Compared with the technical solution of performing image reconstruction on under-sampled frequency domain data in a non-Cartesian coordinate system on the basis of non-uniform fast Fourier transformation, in the method, a deep learning network is pre-trained, and reconstruction can be directly performed to obtain a corresponding image according to under-sampled frequency domain data input in a non-Cartesian coordinate system, without the need to manually select/adjust parameters, such as a scale factor, such that the reconstruction speed of the under-sampled frequency domain data in the non-Cartesian coordinate system is improved.

Description

Image reconstruction method, device, equipment and storage medium

Technical field

This application relates to the field of image processing technology, for example, to an image reconstruction method, device, device, and storage medium.

Background technique

In the current medical imaging imaging, such as computed tomography (CT), magnetic resonance imaging (magnetic resonance imaging, MRI), positron emission tomography (positron emission tomography, PET), it is necessary to analyze the target object. The sampled data is used for image reconstruction to form a high-definition image of the scanned area.

Taking magnetic resonance imaging MRI as an example, in order to enhance the clinical practicability of magnetic resonance imaging technology and shorten the scanning time, magnetic resonance equipment often uses a sampling frequency far lower than the Nyquist sampling frequency for data sampling to obtain target objects in non-Cartesian The under-sampling frequency domain data in the coordinate system is then reconstructed from the under-sampling frequency domain data in the non-Cartesian coordinate system to form a high-definition image of the target object.

However, the current image reconstruction method based on the under-sampled data in a non-Cartesian coordinate system has a longer image reconstruction time and a slower imaging speed.

Summary of the invention

technical problem

One of the objectives of the embodiments of the present application is to provide an image reconstruction method, device, equipment, and storage medium to solve the problem of image reconstruction time when performing image reconstruction based on the under-sampling number in a non-Cartesian coordinate system in the prior art. Long technical problem.

The solution to the problem

Technical solutions

In order to solve the above technical problems, the technical solutions adopted in the embodiments of this application are:

In the first aspect, an embodiment of the present application provides an image reconstruction method, including:

Acquire sampling data of the target object; wherein, the sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system acquired based on a preset sampling mode;

Input the sampled data into the trained deep learning network for processing, and obtain the reconstructed image corresponding to the sampled data; among them, the deep learning network is based on the under-sampled sample frequency domain data and the imaging target in the non-Cartesian coordinate system of multiple imaging targets Fully sampled sample images, obtained by training the initial deep learning network.

In a possible implementation of the first aspect, inputting sampled data into a trained deep learning network for processing to obtain a reconstructed image corresponding to the sampled data includes:

Input the sampled data into the trained deep learning network for coordinate conversion to obtain the under-sampled non-uniform frequency domain data in the Cartesian coordinate system corresponding to the sampled data;

Perform density compensation processing on the under-sampled non-uniform frequency domain data in the Cartesian coordinate system to obtain the under-sampled uniform frequency domain data in the Cartesian coordinate system;

Perform reconstruction processing on the under-sampled uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data.

In a possible implementation of the first aspect, performing reconstruction processing on the under-sampled uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data includes:

Fill in the under-sampled uniform frequency domain data to obtain fully-sampled uniform frequency domain data;

Perform inverse Fourier transform on the uniform frequency domain data to obtain the reconstructed image corresponding to the under-sampled uniform frequency domain data.

In a possible implementation of the first aspect, before inputting the sampled data into the trained deep learning network for image reconstruction, the method further includes:

Acquire multiple training samples, each training sample includes a fully sampled sample image of the imaging target and sample sample data obtained by processing the imaging target based on a preset sampling mode; the fully sampled sample image is used to predict the image output by the initial deep learning network Compare; the interval between two adjacent sampling points in the preset sampling mode is not uniform;

According to multiple training samples, the initial deep learning network is iteratively trained, and the training is stopped when the preset conditions are met, and the trained deep learning network is obtained.

In a possible implementation of the first aspect, there are multiple imaging targets used for training;

Obtaining multiple training samples includes:

Obtain the full sample sample image corresponding to each imaging target, and the full sample sample image is collected by the medical imaging equipment;

Obtain sample sampling data of each imaging target based on the preset sampling mode, the sample sampling data is the under-sampling frequency domain data in a non-Cartesian coordinate system;

The fully sampled sample image and sample sample data corresponding to each imaging target are used as a training sample.

In a possible implementation of the first aspect, the initial deep learning network includes: a coordinate conversion convolution module, a density compensation convolution module, and an image reconstruction convolution module that are sequentially cascaded;

According to a plurality of the training samples, the initial deep learning network is iteratively trained, and the training is stopped when the preset conditions are met. The deep learning network obtained after training includes:

Initialize the model parameters of the initial deep learning network;

Use the coordinate conversion convolution module to perform convolution operation on the sample sample data in the training sample to obtain the under-sampled non-uniform sample data in the Cartesian coordinate system;

Perform density compensation on the under-sampled non-uniform sample data in the Cartesian coordinate system through the density compensation convolution module to generate corresponding under-sampled uniform sample data;

Perform convolution operation on the under-sampled uniform sample data through the image reconstruction convolution module to generate the predicted image corresponding to the under-sampled uniform sample data;

If the preset conditions are not currently met, update the model parameters of the current deep learning network, and return to perform the convolution operation of the sample sampling data in the training sample through the coordinate conversion convolution module to obtain the under-sampling non-uniformity in the Cartesian coordinate system The step of sample data; if the preset conditions are met, the current model parameters are saved to obtain the deep learning network.

In a possible implementation of the first aspect, the preset sampling mode includes radial sampling or spiral scanning sampling.

In the second aspect, an embodiment of the present application provides an image reconstruction device, including:

The sampling module is used to obtain sampling data of the target object; wherein, the sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system obtained based on a preset sampling mode;

The reconstruction module is used to input the sampled data into the trained deep learning network for processing to obtain the reconstructed image corresponding to the sampled data; among them, the deep learning network is based on the frequency domain of under-sampled samples in a non-Cartesian coordinate system of multiple imaging targets The data and the fully sampled sample images of the imaging target are obtained by training the initial deep learning network.

In a third aspect, an embodiment of the present application provides an image reconstruction device, including a memory, a processor, and computer-readable instructions stored in the memory and running on the processor. The processor executes the computer-readable instructions to implement the above-mentioned first On the one hand, the steps of any method.

In a fourth aspect, the embodiments of the present application provide a computer-readable storage medium, the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions implement the steps of any one of the methods in the first aspect when the computer-readable instructions are executed by a processor. .

In a fifth aspect, embodiments of the present application provide a computer program product, which when the computer program product runs on an image reconstruction device, causes the image reconstruction device to execute the method in any one of the above-mentioned first aspects.

Compared with the prior art, the embodiment of the present application has the beneficial effect that the sampled data of the target object is processed through the trained deep learning network, thereby obtaining the reconstructed image corresponding to the sampled data, and realizing the image of the sampled data of the target object Reconstruction; where the sampled data is the under-sampled frequency domain data in a non-Cartesian coordinate system, and the deep learning network is based on the under-sampled sample frequency domain data of multiple imaging targets and the fully sampled sample image of the imaging target, and the initial deep learning Obtained from network training. Compared with the prior art technical solution for image reconstruction based on non-uniform fast Fourier transform for under-sampled frequency domain data in a non-Cartesian coordinate system, the deep learning network in the embodiment of the application is pre-trained and can be directly based on The input sampling data obtains the corresponding reconstruction graph without manual selection/adjustment of parameters such as scale factors, which improves the reconstruction speed of the under-sampled frequency domain data in the non-Cartesian coordinate system; on the other hand, the deep learning network in this application Accelerated calculations can be performed directly based on the graphics processor GPU. Under the premise of ensuring the accuracy of image reconstruction, the image reconstruction time of under-sampling frequency domain data in the non-Cartesian coordinate system is further shortened.

It can be understood that, for the beneficial effects of the second aspect to the fifth aspect described above, reference may be made to the related description in the first aspect described above, and details are not repeated here.

The beneficial effects of the invention

Brief description of the drawings

Description of the drawings

In order to more clearly describe the technical solutions in the embodiments of the present application, the following will briefly introduce the accompanying drawings that need to be used in the embodiments or exemplary technical descriptions. Obviously, the accompanying drawings in the following description are only of the present application. For some embodiments, those of ordinary skill in the art can obtain other drawings based on these drawings without creative work.

In order to more clearly describe the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only of the present application. For some embodiments, those of ordinary skill in the art can obtain other drawings based on these drawings without creative work.

FIG. 1 is a schematic diagram of the architecture of an image reconstruction system provided by an embodiment of the present application;

FIG. 2 is a schematic flowchart of an image reconstruction method provided by an embodiment of the present application;

FIG. 3 is a schematic diagram of the architecture of a deep learning network provided by an embodiment of the present application;

FIG. 4 is a schematic diagram of a process of obtaining a reconstructed image corresponding to sampling data according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a process of acquiring a trained deep learning network provided by an embodiment of the present application;

FIG. 6 is a schematic diagram of a process of obtaining training samples provided by an embodiment of the present application;

FIG. 7 is a schematic diagram of a process flow for iterative training of a deep learning network provided by an embodiment of the present application;

FIG. 8 is a schematic diagram of the composition of an image reconstruction device provided by an embodiment of the present application;

FIG. 9 is a schematic diagram of the composition of an image reconstruction device provided by another embodiment of the present application;

FIG. 10 is a schematic structural diagram of an image reconstruction device provided by an embodiment of the present application.

Invention embodiment

Embodiments of the present invention

In order to have a more detailed understanding of the characteristics and technical content of the embodiments of the present application, the implementation of the embodiments of the present application will be described in detail below with reference to the accompanying drawings. The attached drawings are for reference and explanation purposes only, and are not used to limit the embodiments of the present application. In the following technical description, for the convenience of explanation, a number of details are used to provide a sufficient understanding of the disclosed embodiments. However, without these details, one or more embodiments can still be implemented. In other cases, in order to simplify the drawings, well-known structures and devices may be simplified for display.

It should be understood that when used in the specification and appended claims of this application, the term "comprising" indicates the existence of the described features, wholes, steps, operations, elements and/or components, but does not exclude one or more other The existence or addition of features, wholes, steps, operations, elements, components, and/or collections thereof.

In addition, in the description of the specification of this application and the appended claims, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

FIG. 1 is a schematic diagram of the architecture of an image reconstruction system provided by an embodiment of the application. As shown in FIG. 1, the image reconstruction system includes an image acquisition device 10 and an image reconstruction device 20. As shown in FIG.

The image acquisition device 10 refers to a device for providing medical imagery to users. For example, magnetic resonance equipment.

The image acquisition device 10 can obtain different sampling data according to different sampling modes, including but not limited to non-uniform under-sampling frequency domain data, uniform under-sampling frequency domain data, uniform full-sampling frequency domain data, and the like. For example, when the sampling mode is spiral scanning sampling, the sampling data is non-uniform under-sampling frequency domain data in a non-Cartesian coordinate system.

The image reconstruction device 20 is configured to receive sampling data in a non-Cartesian coordinate system sent by the image acquisition device 10 to perform image reconstruction. The image reconstruction device 20 communicates with the image acquisition device 10 through a network. The aforementioned networks include, but are not limited to, wide area networks and local area networks.

The image reconstruction device 20 may be a cloud server. The cloud server may be a server that implements a single function or a server that implements multiple functions. Specifically, it may be an independent physical server or a cluster of physical servers.

Illustratively, it is assumed that the image reconstruction method provided in the embodiment of the present application is executed by the image reconstruction device 20, and the image acquisition device 10 is a magnetic resonance device. The magnetic resonance equipment stores the sampling points in a preset arrangement, generates sampling data, and sends the sampling data to the image reconstruction device 20, and the image reconstruction device 20 performs subsequent image reconstruction.

The arrangement method is determined by the sampling mode. In practical applications, if the magnetic resonance equipment adopts linear scanning sampling, the sampling points are evenly distributed on the grid points. The image reconstruction device 20 can obtain the signal intensity of each voxel at a certain position in the two-dimensional plane through the inverse Fourier transform, and convert it into the corresponding gray value to obtain the magnetic resonance image.

However, the imaging speed of linear sampling is slow, which limits the application and development of magnetic resonance equipment in the field of clinical medical imaging. In order to improve imaging efficiency, spiral trajectories can be used for scanning and sampling. After sampling with spiral trajectory scanning, the sampling points are located on a series of spiral trajectories on the polar coordinate plane (non-Cartesian coordinate system). They are not regular rectangular grid points. If you directly reconstruct the image of the sampling points obtained by spiral trajectory scanning, It is necessary to perform grid processing in advance and use window functions (scale factors) for smoothing processing, which requires a very large amount of calculation.

In the image reconstruction method provided by the embodiment of the application, the image reconstruction device 20 obtains the under-sampling data in the non-Cartesian coordinate system collected by the magnetic resonance device 10, and inputs the under-sampling data in the non-Cartesian coordinate system to the depth after training The learning network performs reconstruction processing and directly obtains the reconstructed image of the target object.

In order to make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will exemplarily describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are A part of the embodiments of the present invention, but not all the embodiments. Based on the embodiments listed below, all other embodiments obtained by a person of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

Fig. 2 is a schematic flowchart of an image reconstruction method provided by an embodiment of the application. The image reconstruction method is executed by the image reconstruction device shown in Fig. 1. As shown in Fig. 2, the image reconstruction method includes:

S201. Acquire sampling data of the target object; where the sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system acquired based on a preset sampling mode;

In this embodiment, the target object is a specimen to be detected by a medical imaging device. Exemplarily, the target object includes a simulation body, a living body (animal or human body), an isolated organ or tissue, and the like.

In this embodiment, the data collected by the preset sampling mode may be data in a non-Cartesian coordinate system; after the sampling point in the preset sampling mode is projected to the Cartesian coordinate system, the interval between two adjacent sampling points It can be uneven. For example, the preset sampling mode may include any one of the following: radial sampling, spiral trajectory sampling, rotating scan sampling, echo planar imaging (EPI) based sampling, and interleave sampling. Among them, EPI relies on the continuous reverse direction switching of the gradient coil after a pulse excitation to collect a series of gradient echo signals.

Exemplarily, it is assumed that the preset sampling mode is that the magnetic resonance device scans the target object based on the spiral trajectory. Spiral trajectory scanning generally starts from the center of the storage space of the sampling point, and then expands outward in a spiral shape. That is to say, the sampling points after one excitation are not arranged sequentially on a two-dimensional grid. By adjusting the gradient waveform, the data is filled along the spiral trajectory. After the spiral trajectory is used for scanning and imaging, each sampling point is located in the polar coordinate plane. A series of spiral trajectories on (non-Cartesian coordinate system) are not regular rectangular grid points; after the sampling points are projected to the Cartesian coordinate system, the intervals between adjacent sampling points are not the same. The magnetic resonance device sends a series of initial sampling points located on the polar coordinate plane to the image reconstruction device, so that the image reconstruction device performs image reconstruction on the data to obtain a reconstructed image of the target object.

S202. Input the sampled data into the trained deep learning network for processing, and obtain a reconstructed image corresponding to the sampled data; wherein, the deep learning network is based on the under-sampled sample frequency domain data in the non-Cartesian coordinate system of multiple imaging targets and imaging The fully sampled sample image of the target is obtained by training the initial deep learning network.

The deep learning network in this embodiment is a deep learning network constructed based on a deep learning framework. The input of the deep learning network is frequency-domain data of under-sampled samples in a non-Cartesian coordinate system, and the output is a high-precision image close to full sampling, that is, a reconstructed image.

In this embodiment, multiple preset sampling modes may correspond to one deep learning network, or each preset sampling model may correspond to one deep learning network. The division should be based on the sampling principle in the preset sampling mode, which is not limited here.

The deep learning network is based on the under-sampled sample frequency domain data in the non-Cartesian coordinate system of multiple imaging targets and the fully sampled sample image of the imaging target, which is obtained by training the initial deep learning network. During the training process, the frequency domain data of the under-sampled samples in the non-Cartesian coordinate system of the imaging target is obtained by using a preset sampling mode. The full-sampled sample image is used to compare with the predicted image output by the initial deep learning network to be based on The comparison result adjusts the model parameters of the initial deep learning network.

In this embodiment, the deep learning network includes a cascaded coordinate conversion convolution module, a density compensation convolution module, and an image reconstruction convolution module that are sequentially cascaded. Among them, the coordinate conversion convolution module is used to receive the input under-sampled frequency domain data in a non-Cartesian coordinate system, and perform convolution operations on the under-sampled frequency-domain data in a non-Cartesian coordinate system to obtain the under-sampled frequency domain data in the non-Cartesian coordinate system. Undersample uniform frequency domain data. The density compensation convolution module is used to perform density compensation processing on the under-sampled non-uniform frequency domain data in the Cartesian coordinate system to obtain the under-sampled uniform frequency domain data in the Cartesian coordinate system. The image reconstruction convolution module is used to reconstruct the under-sampled uniform frequency domain data to obtain the reconstructed image corresponding to the under-sampled uniform frequency domain data.

Please refer to FIG. 3 together. FIG. 3 mainly describes how to obtain the reconstructed image corresponding to the sampling data in S202. Assume that the preset sampling mode is spiral trajectory sampling.

As shown in Figure 3, the sampled data is input into the trained deep learning network for processing, and the reconstructed image corresponding to the sampled data is obtained, including S2011 to S2013, as follows:

S2021: Input the sampled data into the trained deep learning network to perform coordinate conversion, and obtain the under-sampled non-uniform frequency domain data in the Cartesian coordinate system corresponding to the sampled data.

The coordinate conversion convolution module receives the input sampling data, that is, the under-sampled frequency domain data in the non-Cartesian coordinate system, and performs convolution operation on the under-sampled frequency domain data in the non-Cartesian coordinate system to obtain the data in the Cartesian coordinate system. Undersample uniform frequency domain data.

Among them, the convolution operation may be a real number convolution operation or a complex number convolution operation. In an embodiment, the coordinate transformation convolution module performs complex convolution operations, and the coordinate transformation convolution module includes a first complex convolution block cascaded in multiple layers.

The complex convolution operation of the first complex convolution block can be expressed as:

W*C=(A+iB)*(a+ib)=(Aa-Bb)+i(Ab+Ba) (1)

Among them, W represents the input sampled data, A represents the real part of the sampled data, B represents the imaginary part of the sampled data, C represents the complex convolution kernel, a represents the real part of the complex convolution kernel, and b represents the imaginary part of the complex convolution kernel. Part, convolve the real and imaginary parts of the sampled data respectively.

S2022 Perform density compensation processing on the under-sampled non-uniform frequency domain data in the Cartesian coordinate system to obtain the under-sampled uniform frequency domain data in the Cartesian coordinate system.

The density compensation convolution module includes a second complex convolution block cascaded in multiple layers. The multi-layer second complex convolution block sequentially implements a complex convolution operation to perform density compensation processing on the under-sampled non-uniform frequency domain data. The complex convolution operation is the same as the above formula (1), and will not be repeated here.

S2023: Perform reconstruction processing on the under-sampled uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data.

Image reconstruction convolution module, fills the under-sampled uniform frequency domain data to obtain fully sampled uniform frequency domain data; then performs inverse Fourier transform on the uniform frequency domain data to obtain the reconstructed image corresponding to the under-sampled uniform frequency domain data .

The image reconstruction convolution module includes a multi-layer cascaded third complex convolution block, and the multi-layer third complex convolution block sequentially implements the filling process of the under-sampled uniform frequency domain data through the complex convolution operation. The complex convolution operation is the same as the above formula (1), and will not be repeated here. From the perspective of signal and image processing, complex numbers introduce phase information compared to real numbers. The phase information of the image provides a detailed description of the shape, edge and direction of the image, which can be used to restore the amplitude information of the image, based on complex convolution operations. The deep learning network has better image reconstruction effects.

In this embodiment, the accuracy of the reconstructed image is affected by the size of the convolution kernel in the first complex convolution block, the second complex convolution block, and the third complex convolution block. The convolution kernel can be set according to the accuracy requirements of the reconstructed image. The size is not limited here.

Exemplarily, please refer to FIG. 4 together. FIG. 4 is a schematic structural diagram of a deep learning network provided by an embodiment of the application. As shown in FIG. 4, the deep learning network includes cascaded coordinate conversion convolution modules that are sequentially cascaded, Density compensation convolution module and image reconstruction convolution module. The input of the coordinate conversion convolution module is sampled data, and the output is under-sampling non-uniform frequency domain data in Cartesian coordinates; the input of the density compensation convolution module is under-sampling frequency Domain data, the output is the under-sampled uniform frequency domain data in the Cartesian coordinate system, the input of the image reconstruction convolution module is the under-sampled uniform frequency domain data, and the output is the reconstructed image.

Wherein, the coordinate conversion convolution module includes a first complex convolution block cascaded in multiple layers. The density compensation convolution module includes a multi-layer cascaded second complex convolution block; the image reconstruction convolution module includes a multi-layer cascaded third complex convolution block.

In practical applications, the coordinate conversion convolution module receives the under-sampled frequency domain data in a non-Cartesian coordinate system, and the cascaded first complex convolution block sequentially performs complex convolutions on the under-sampled frequency domain data in the non-Cartesian coordinate system. Product operation, that is, the output of the previous first complex convolution block is the input of the next first complex convolution block, until multiple first complex convolution modules have completed the complex convolution operation to generate the undercartesian coordinate system Sampling frequency domain data; the first second complex convolution block in the second complex convolution block of multi-layer cascade receives the output of the last first complex convolution block (that is, the under-sampled frequency domain data in the Cartesian coordinate system) , And perform complex convolution operations on the output, and then the cascaded second complex convolution block sequentially performs complex convolution operations on the output of the previous second complex convolution block, until the output under-sampling in the Cartesian coordinate system is uniform Frequency domain data;

The cascaded third complex convolution block sequentially performs complex convolution operations on the under-sampled uniform frequency domain data, that is, the output of the previous third complex convolution block is the input of the next third complex convolution block, up to multiple The third complex convolution block all completes the complex convolution operation and outputs the reconstructed image.

Compared with the prior art, the advantageous effect of the image reconstruction method provided in this embodiment is that the sampled data of the target object is processed through the trained deep learning network, so as to obtain the reconstructed image corresponding to the sampled data. , The image reconstruction of the sampled data of the target object is realized; among them, the sampled data is the under-sampled frequency domain data in a non-Cartesian coordinate system, and the deep learning network is based on the under-sampled sample frequency domain data of multiple imaging targets and the imaging target's image reconstruction. Fully sampled sample images, obtained by training the initial deep learning network. Compared with the prior art technical solution for image reconstruction based on non-uniform fast Fourier transform for under-sampled frequency domain data in a non-Cartesian coordinate system, the deep learning network in the embodiment of the application is pre-trained and can be directly based on The input sampling data obtains the corresponding reconstruction graph without manual selection/adjustment of parameters such as scale factors, which improves the reconstruction speed of the under-sampled frequency domain data in the non-Cartesian coordinate system; on the other hand, the deep learning network in this application Accelerated calculations can be directly based on the graphics processor GPU. Under the premise of ensuring the accuracy of image reconstruction, the image reconstruction time of under-sampling frequency domain data in non-Cartesian coordinates is further shortened.

Please refer to FIG. 5 together. FIG. 5 is a schematic diagram of a process for obtaining a trained deep learning network according to an embodiment of the application. FIG. 5 mainly describes how to obtain the trained deep learning network in step S202. As shown in Figure 5, the method for obtaining the trained deep learning network includes S501 to S502, which are specifically as follows:

S501. Obtain a plurality of training samples, each training sample includes a fully sampled sample image of the imaging target and sample sample data obtained by processing the imaging target based on a preset sampling mode; the fully sampled sample image is used for output from the initial deep learning network The predicted images are compared; the interval between two adjacent sampling points in the preset sampling mode is not uniform.

The multiple training samples correspond to multiple imaging targets, and the training samples correspond to the imaging targets on a one-to-one basis. The imaging targets may include the target object in step 202, which is a specimen used for detection by a medical imaging device. The imaging target may refer to objects at different positions on a target specimen, or may refer to different target specimens.

The fully sampled sample image of the imaging target can be acquired from a medical imaging device based on a low-power under-sampling factor.

The sample sampling data in each training sample can be obtained by processing the imaging target based on a preset sampling mode.

Please refer to FIG. 6 together. FIG. 6 mainly describes how to obtain training samples in the foregoing S501. There are multiple imaging targets, and multiple training samples correspond to multiple imaging targets.

As shown in Figure 6, the steps of obtaining multiple training samples include S5011 to S5013, which are specifically as follows:

S5011. Obtain a full-sampling sample image corresponding to each imaging target, and the full-sampling sample image is collected by the medical imaging equipment.

In this embodiment, the hospital imaging equipment may be a magnetic resonance equipment, and the magnetic resonance equipment may acquire a scan image of the imaging target from the magnetic resonance equipment based on a low-power under-collection factor, and then preprocess the acquired scan image and process it The latter image is used as a fully sampled sample image of the imaging target. Among them, the preprocessing method can include image selection processing, normalization processing, and so on. Image selection processing is used to extract some images with low quality or with more noise data to improve the efficiency of training. The normalization process is to facilitate the input of the fully sampled image in the complex convolutional neural network to adapt to subsequent training.

S5012: Acquire sample sampling data of each imaging target based on a preset sampling mode, where the sample sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system.

For a certain imaging target, the initial under-sampling frequency domain data is acquired from the same MRI device based on the preset sampling mode, that is, the frequency domain data before imaging. The imaging target is in a non-Cartesian coordinate system. Frequency domain data of under-sampled samples.

S5013. Use the fully sampled sample image and sample sample data corresponding to each imaging target as a training sample.

Combine its fully sampled image and undersampled frequency domain data into a training sample.

The acquisition of multiple training samples provided by this embodiment preprocesses the fully sampled initial scan image of the imaging target obtained by the medical imaging equipment to obtain the fully sampled image of the imaging target, which partially eliminates the adverse effects caused by singular sample data and improves Training efficiency.

S502: Perform iterative training on the initial deep learning network according to multiple training samples, and stop training when a preset condition is met, to obtain a trained deep learning network.

In this embodiment, the structure of the initial deep learning network is the same as the structure of the deep learning network in step 202. The initial deep learning network includes a cascaded coordinate conversion convolution module, a density compensation convolution module, and an image reconstruction convolution module.

In this embodiment, the iterative training of the deep learning network based on multiple training samples includes, for each training sample, using a fully sampled image in the training sample as a label, and sample sampling data in the training sample as input. Obtain the predicted image output by the deep learning network, compare the predicted image output by the deep learning network with the fully sampled image, and adjust the model parameters of the deep learning network according to the comparison results; after that, start the next round of iterative training; if the current number of training When the preset number of iterations is met, or the error between the accuracy of the predicted image obtained in this iterative training and the accuracy of the fully sampled sample image in the training sample is less than or equal to the preset error threshold, the training is stopped, and the trained deep learning network is obtained . Among them, the error is calculated by using a preset loss function.

In this embodiment, the training loss function may be a minimum absolute value deviation loss function, a minimum square error loss function, etc., which are not specifically limited here. The loss function is used to calculate the error value between the predicted image output by the deep learning network and the fully sampled image. For example, image resolution difference, image sharpness difference, or image similarity difference, etc.

In this embodiment, the deep learning network is trained based on multiple training samples, and the model parameters of the complex neural network model are optimized. Based on this training, the deep learning network is input to the non-Cartesian coordinate system of any target object. By under-sampling the frequency domain data, a high-resolution image of the target object, that is, a reconstructed image, can be obtained. The high-resolution image is an image close to a fully-sampled image, which can meet the actual application requirements in medical imaging.

In another embodiment, the deep learning network can be trained based on an end-to-end training mechanism. Please also refer to FIG. 7. FIG. 7 mainly describes the iterative training of the deep learning network in step S502 as an example.

According to a plurality of the training samples, the initial deep learning network is iteratively trained, and when the preset conditions are met, the training is stopped to obtain the trained deep learning network including:

S701: Initialize model parameters of the initial deep learning network.

The initial values of the model parameters are preset values.

S702: Perform a convolution operation on the sample sample data in the training sample by the coordinate conversion convolution module to obtain the under-sampled non-uniform sample data in the Cartesian coordinate system.

Obtain multiple training samples according to multiple imaging targets, and use multiple training samples to perform end-to-end training in sequence.

In this embodiment, the sample sampling data of the training samples are input to the coordinate conversion convolution module in the initial deep learning network, and forward propagation based on the current model parameters of the initial deep learning network. Specifically, the first complex convolution block cascaded in the coordinate conversion convolution module sequentially performs complex convolution operations, until multiple first complex convolution blocks have completed the complex convolution operation, and the under-sampling frequency in the Cartesian coordinate system is generated. Domain data.

S703: Perform density compensation on the under-sampled non-uniform sample data in the Cartesian coordinate system through the density compensation convolution module, and generate corresponding under-sampled uniform sample data.

This step is the same as the processing of step S2022, and will not be repeated here.

S704: Perform a convolution operation on the under-sampled uniform sample data through the image reconstruction convolution module to generate a predicted image corresponding to the under-sampled uniform sample data.

This step is the same as the processing of step S2023, and will not be repeated here.

S705. If the preset conditions are not currently met, update the model parameters of the current deep learning network, and return to perform the convolution operation on the sample sample data in the training sample through the coordinate conversion convolution module to obtain the undersampling in the Cartesian coordinate system The step of non-uniform sample data; if the preset conditions are met, the current model parameters are saved to obtain a deep learning network.

After S704 is executed, it is determined whether the preset condition is currently met. The preset condition may be: the current number of training reaches the preset number of iterations, or the error between the accuracy of the predicted image obtained in this iteration training and the accuracy of the fully sampled sample image in the training sample is less than or equal to the preset error threshold .

If the preset conditions are not currently met, update the model parameters of the current deep learning network, and return to S702, and continue to execute S702 to S704.

Exemplarily, assuming that the error between the accuracy of the predicted image obtained in this iterative training and the accuracy of the fully sampled sample image in the training sample is greater than the preset error threshold, then based on the error, backpropagation is performed in the deep learning network. Update the model parameters of the current deep learning network. Then return to execute S702~S704, execute the next iteration, that is, take the sample sampling data of the training sample as the input, and propagate the deep learning network based on the updated model parameters forward, and the output of the current deep learning network (predicted image) is again with the current training The fully sampled sample images in the sample are compared until the error between the two is less than the preset threshold, the current model parameters are saved, and the deep learning network is obtained. Among them, back propagation is the process of adjusting model parameters based on the results and errors of forward propagation.

If the preset conditions are met, the current model parameters are saved, and the deep learning network is obtained.

The iterative training method of the deep learning network provided in this embodiment adopts an end-to-end training method, and directly inputs the sample data obtained by the collection into the deep learning network to obtain a predicted image. The predicted image is combined with the fully sampled samples in the training sample. The image comparison will get an error, the error is back propagated, and the model parameters of the deep learning network are updated until the accuracy error between the output of the deep learning network and the fully sampled sample image is less than the preset threshold. This method saves the data labeling work required before the execution of each independent learning task, and can directly perform training calculations based on the graphics processor GPU, which improves the training efficiency.

It should be understood that the size of the sequence number of each step in the foregoing embodiment does not mean the order of execution. The execution sequence of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

Based on the image reconstruction method provided in the foregoing embodiment, the embodiment of the present invention further provides an embodiment of an apparatus for implementing the foregoing method embodiment.

FIG. 8 is a schematic diagram of the composition of an image reconstruction device provided by an embodiment of the application. As shown in FIG. 8, the image reconstruction device 80 includes: a sampling module 801 and a reconstruction module 802.

The sampling module 801 is configured to obtain sampling data of a target object; wherein the sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system obtained based on a preset sampling mode;

The reconstruction module 802 is used to input the sampled data into the trained deep learning network for processing, and obtain the reconstructed image corresponding to the sampled data; wherein, the deep learning network is based on the under-sampled sample frequency in the non-Cartesian coordinate system of multiple imaging targets The domain data and the fully sampled sample image of the imaging target are obtained by training the initial deep learning network.

The reconstruction module 802 is specifically used for:

The reconstruction module 802 is also specifically used for:

The image reconstruction device provided in this embodiment processes the sampled data of the target object through the trained deep learning network, thereby obtaining the reconstructed image corresponding to the sampled data, and realizes the image reconstruction of the sampled data of the target object; where the sampled data is For the under-sampled frequency domain data in a non-Cartesian coordinate system, the deep learning network is obtained by training the initial deep learning network based on the under-sampled sample frequency domain data of multiple imaging targets and the fully sampled sample image of the imaging target. Compared with the prior art technical solution for image reconstruction based on non-uniform fast Fourier transform for under-sampled frequency domain data in a non-Cartesian coordinate system, the deep learning network in the embodiment of the application is pre-trained and can be directly based on The input sampling data obtains the corresponding reconstruction graph without manual selection/adjustment of parameters such as scale factors, which improves the reconstruction speed of the under-sampled frequency domain data in the non-Cartesian coordinate system; on the other hand, the deep learning network in this application Accelerated calculations can be directly based on the graphics processor GPU. Under the premise of ensuring the accuracy of image reconstruction, the image reconstruction time of under-sampling frequency domain data in non-Cartesian coordinates is further shortened.

FIG. 9 is a schematic diagram of the composition of an image reconstruction device provided by another embodiment of the application. As shown in FIG. 9, the image reconstruction device 80 further includes a training module 803.

Training module 803, used for:

Optionally, the training module 803 is specifically used for:

Optionally, the deep learning network includes: a coordinate transformation convolution module, an inverse Fourier transformation module, and a second complex convolution layer that are sequentially cascaded; the training module 803 is also specifically used for:

Initialize the model parameters of the initial deep learning network;

Optionally, the coordinate transformation convolution module includes a plurality of cascaded first complex convolution modules, and the second complex convolution layer includes a plurality of cascaded second complex convolution modules; the first complex convolution module and the second complex convolution module; The size of the convolution kernel of the complex convolution module is the same.

Optionally, the preset sampling mode includes radial sampling or spiral scanning sampling.

The image reconstruction device provided in this embodiment trains the initial deep learning network based on multiple training samples, and optimizes the model parameters of the initial deep learning network. Based on this training, the deep learning network is used to input any target object. The under-sampling frequency domain data in the Cartesian coordinate system can obtain a high-resolution image of the target object, that is, a reconstructed image. The high-resolution image is an image close to the full-sampled image, which can meet the actual application requirements in medical imaging. .

The image reconstruction apparatus provided by the embodiments shown in FIG. 8 and FIG. 9 can be used to implement the technical solutions in the foregoing method embodiments, and their implementation principles and technical effects are similar, and will not be repeated here in this embodiment.

Fig. 10 is a schematic diagram of an image reconstruction device provided by an embodiment of the present application. As shown in FIG. 9, the image reconstruction terminal device 100 of this embodiment includes: at least one processor 1001, a memory 1002, and computer-readable instructions stored in the memory 1002 and executable on the processor 1001. The image reconstruction device further includes a communication component 1003, wherein the processor 1001, the memory 1002, and the communication component 1003 are connected by a bus 1004.

The processor 1001 implements the steps in the foregoing image reconstruction method embodiments when executing computer-readable instructions, such as step S201 to step S202 in the embodiment shown in FIG. 2. Alternatively, when the processor 1001 executes the computer-readable instructions, the functions of the modules/units in the foregoing device embodiments, for example, the functions of the modules 801 to 802 shown in FIG. 8 are realized.

Exemplarily, the computer-readable instructions may be divided into one or more modules/units, and the one or more modules/units are stored in the memory 1002 and executed by the processor 1001 to complete the present application. The one or more modules/units may be a series of computer-readable instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer-readable instructions in the image reconstruction device 100.

Those skilled in the art can understand that FIG. 10 is only an example of the image reconstruction device and does not constitute a limitation on the image reconstruction device. It may include more or less components than shown in the figure, or a combination of certain components, or different components. Components, such as input and output devices, network access devices, buses, etc.

The processor 1001 can be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), ready-made Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.

The memory 1002 may be an internal storage unit of the image reconstruction device, or an external storage device of the image reconstruction device, such as a plug-in hard disk, a smart media card (SMC), or a secure digital (SD) card. Flash Card, etc. The memory 1002 is used to store the computer readable instructions and other programs and data required by the image reconstruction device. The memory 1002 can also be used to temporarily store data that has been output or will be output.

The bus can be an Industry Standard Architecture (ISA) bus, Peripheral Component (PCI) bus, or Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus and so on. For ease of representation, the buses in the drawings of this application are not limited to only one bus or one type of bus.

The embodiments of the present application also provide a computer-readable storage medium, and the computer-readable storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the steps in the foregoing method embodiments can be realized.

The embodiments of the present application provide a computer program product. When the computer program product runs on an image reconstruction device, the image reconstruction device can implement the steps in the foregoing method embodiments when executed.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the implementation of all or part of the processes in the above-mentioned embodiments and methods in this application can be accomplished by instructing relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the computer-readable instructions are executed by the processor, they can implement the steps of the foregoing method embodiments. Wherein, the computer-readable instruction includes computer-readable instruction code, and the computer-readable instruction code may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may at least include: any entity or device capable of carrying computer-readable instruction codes to the image reconstruction equipment, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunications signal and software distribution medium. For example, U disk, mobile hard disk, floppy disk or CD-ROM, etc. In some jurisdictions, according to legislation and patent practices, computer-readable media cannot be electrical carrier signals and telecommunication signals.

In the above-mentioned embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail or recorded in an embodiment, reference may be made to related descriptions of other embodiments.

A person of ordinary skill in the art may realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraint conditions of the technical solution. Professionals and technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered beyond the scope of this application.

In the embodiments provided in this application, it should be understood that the disclosed apparatus/network equipment and method may be implemented in other ways. For example, the device/network device embodiments described above are only illustrative. For example, the division of the modules or units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units. Or components can be combined or integrated into another system, or some features can be omitted or not implemented. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, a person of ordinary skill in the art should understand that it can still implement the foregoing The technical solutions recorded in the examples are modified, or some of the technical features are equivalently replaced; these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the application, and should be included in Within the scope of protection of this application.

Claims

An image reconstruction method, characterized in that it comprises:

Acquiring sampling data of the target object; wherein the sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system acquired based on a preset sampling mode;

Input the sampled data into the trained deep learning network for processing, and obtain the reconstructed image corresponding to the sampled data; wherein, the deep learning network is based on the under-sampled sample frequency in the non-Cartesian coordinate system of multiple imaging targets. The domain data and the fully sampled sample image of the imaging target are obtained by training the initial deep learning network.
8. The image reconstruction method according to claim 1, wherein said inputting said sampled data into a trained deep learning network for processing to obtain a reconstructed image corresponding to said sampled data comprises:

Input the sampled data into the trained deep learning network to perform coordinate conversion to obtain the under-sampled non-uniform frequency domain data in the Cartesian coordinate system corresponding to the sampled data;

Performing density compensation processing on the under-sampled non-uniform frequency domain data in the Cartesian coordinate system to obtain the under-sampled uniform frequency domain data in the Cartesian coordinate system;

Perform reconstruction processing on the under-sampled uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data.
3. The image reconstruction method according to claim 2, wherein the performing reconstruction processing on the under-sampled uniform frequency domain data to obtain the reconstructed image corresponding to the under-sampled uniform frequency domain data comprises:

Performing filling processing on the under-sampled uniform frequency domain data to obtain fully-sampled uniform frequency domain data;

Perform inverse Fourier transform on the uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data.
The image reconstruction method according to any one of claims 1 to 3, wherein before the input of the sampled data into the trained deep learning network for image reconstruction, the method further comprises:

Acquire a plurality of training samples, each training sample includes a fully sampled sample image of the imaging target and sample sample data obtained by processing the imaging target based on the preset sampling mode; the fully sampled sample image is used to communicate with the The predicted images output by the initial deep learning network are compared; the interval between two adjacent sampling points in the preset sampling mode is not uniform;

According to a plurality of the training samples, iterative training is performed on the initial deep learning network, and the training is stopped when a preset condition is met, to obtain a trained deep learning network.
The image reconstruction method according to claim 4, wherein there are multiple imaging targets used for training;

The obtaining multiple training samples includes:

Acquiring a fully sampled sample image corresponding to each imaging target, the fully sampled sample image being collected by a medical imaging device;

Acquiring sample sampling data of each imaging target based on the preset sampling mode, where the sample sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system;

The fully sampled sample image and sample sample data corresponding to each imaging target are used as a training sample.
5. The image reconstruction method according to claim 4, wherein the initial deep learning network comprises: a coordinate conversion convolution module, a density compensation convolution module, and an image reconstruction convolution module that are sequentially cascaded;

The iterative training of the initial deep learning network based on a plurality of the training samples, and stopping the training when a preset condition is met, obtaining the trained deep learning network includes:

Initialize the model parameters of the initial deep learning network;

Performing a convolution operation on the sample sample data in the training sample by the coordinate conversion convolution module to obtain under-sampled non-uniform sample data in a Cartesian coordinate system;

Performing density compensation on the under-sampled non-uniform sample data in the Cartesian coordinate system by the density compensation convolution module, and generate corresponding under-sampled uniform sample data;

Performing a convolution operation on the under-sampled uniform sample data by the image reconstruction convolution module to generate a predicted image corresponding to the under-sampled uniform sample data;

If the preset conditions are not currently met, update the model parameters of the current deep learning network, and return to perform the convolution operation of the sample sample data in the training sample through the coordinate conversion convolution module to obtain the Cartesian coordinate system The step of under-sampling non-uniform sample data; if the preset conditions are met, save the current model parameters to obtain the deep learning network.
8. The image reconstruction method according to claim 1, wherein the preset sampling mode includes radial sampling or helical scanning sampling.
An image reconstruction device, characterized in that it comprises:

An acquisition module for acquiring sampling data of the target object; wherein the sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system acquired based on a preset sampling mode;

The reconstruction module is used to input the sampled data into the trained deep learning network for processing to obtain the reconstructed image corresponding to the sampled data; wherein, the deep learning network is based on multiple imaging targets in a non-Cartesian coordinate system The under-sampled sample frequency domain data of and the fully sampled sample image of the imaging target are obtained by training the initial deep learning network.
An image reconstruction device, comprising a memory, a processor, and computer-readable instructions stored in the memory and running on the processor, wherein the processor executes the computer-readable instructions to implement The following steps:

Acquiring sampling data of the target object; wherein the sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system acquired based on a preset sampling mode;

Input the sampled data into the trained deep learning network for processing, and obtain the reconstructed image corresponding to the sampled data; wherein, the deep learning network is based on the under-sampled sample frequency in the non-Cartesian coordinate system of multiple imaging targets. The domain data and the fully sampled sample image of the imaging target are obtained by training the initial deep learning network.
9. The image reconstruction device according to claim 9, wherein said inputting said sampled data into a trained deep learning network for processing to obtain a reconstructed image corresponding to said sampled data comprises:

Input the sampled data into the trained deep learning network to perform coordinate conversion to obtain the under-sampled non-uniform frequency domain data in the Cartesian coordinate system corresponding to the sampled data;

Performing density compensation processing on the under-sampled non-uniform frequency domain data in the Cartesian coordinate system to obtain the under-sampled uniform frequency domain data in the Cartesian coordinate system;

Perform reconstruction processing on the under-sampled uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data.
11. The image reconstruction device according to claim 10, wherein the performing reconstruction processing on the under-sampled uniform frequency domain data to obtain the reconstructed image corresponding to the under-sampled uniform frequency domain data comprises:

Performing filling processing on the under-sampled uniform frequency domain data to obtain fully-sampled uniform frequency domain data;

Perform inverse Fourier transform on the uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data.
The image reconstruction device according to any one of claims 9-11, wherein the processor executes the computer-readable instructions before inputting the sampled data into the trained deep learning network for image reconstruction It also implements the following steps:

Acquire a plurality of training samples, each training sample includes a fully sampled sample image of the imaging target and sample sample data obtained by processing the imaging target based on the preset sampling mode; the fully sampled sample image is used to communicate with the The predicted images output by the initial deep learning network are compared; the interval between two adjacent sampling points in the preset sampling mode is not uniform;

According to a plurality of the training samples, iterative training is performed on the initial deep learning network, and the training is stopped when a preset condition is met, to obtain a trained deep learning network.
The image reconstruction device according to claim 12, wherein there are multiple imaging targets used for training;

The obtaining multiple training samples includes:

Acquiring a fully sampled sample image corresponding to each imaging target, the fully sampled sample image being collected by a medical imaging device;

Acquiring sample sampling data of each imaging target based on the preset sampling mode, where the sample sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system;

The fully sampled sample image and sample sample data corresponding to each imaging target are used as a training sample.
The image reconstruction device according to claim 12, wherein the initial deep learning network comprises: a coordinate conversion convolution module, a density compensation convolution module, and an image reconstruction convolution module that are sequentially cascaded;

The iterative training of the initial deep learning network based on a plurality of the training samples, and stopping the training when a preset condition is met, obtaining the trained deep learning network includes:

Initialize the model parameters of the initial deep learning network;

Performing a convolution operation on the sample sample data in the training sample by the coordinate conversion convolution module to obtain under-sampled non-uniform sample data in a Cartesian coordinate system;

Performing density compensation on the under-sampled non-uniform sample data in the Cartesian coordinate system by the density compensation convolution module, and generate corresponding under-sampled uniform sample data;

Performing a convolution operation on the under-sampled uniform sample data by the image reconstruction convolution module to generate a predicted image corresponding to the under-sampled uniform sample data;

If the preset conditions are not currently met, update the model parameters of the current deep learning network, and return to perform the convolution operation of the sample sample data in the training sample through the coordinate conversion convolution module to obtain the Cartesian coordinate system The step of under-sampling non-uniform sample data; if the preset conditions are met, save the current model parameters to obtain the deep learning network.
A computer-readable storage medium, the computer-readable storage medium storing computer-readable instructions, wherein the computer-readable instructions are executed by a processor to implement the following steps:

Acquiring sampling data of the target object; wherein the sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system acquired based on a preset sampling mode;

Input the sampled data into the trained deep learning network for processing, and obtain the reconstructed image corresponding to the sampled data; wherein, the deep learning network is based on the under-sampled sample frequency in the non-Cartesian coordinate system of multiple imaging targets. The domain data and the fully sampled sample image of the imaging target are obtained by training the initial deep learning network.
15. The computer-readable storage medium of claim 15, wherein the inputting the sampled data into a trained deep learning network for processing to obtain a reconstructed image corresponding to the sampled data comprises:

Input the sampled data into the trained deep learning network to perform coordinate conversion to obtain the under-sampled non-uniform frequency domain data in the Cartesian coordinate system corresponding to the sampled data;

Performing density compensation processing on the under-sampled non-uniform frequency domain data in the Cartesian coordinate system to obtain the under-sampled uniform frequency domain data in the Cartesian coordinate system;

Perform reconstruction processing on the under-sampled uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data.
15. The computer-readable storage medium of claim 16, wherein the performing reconstruction processing on the under-sampled uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data comprises:

Performing filling processing on the under-sampled uniform frequency domain data to obtain fully-sampled uniform frequency domain data;

Perform inverse Fourier transform on the uniform frequency domain data to obtain a reconstructed image corresponding to the under-sampled uniform frequency domain data.
The computer-readable storage medium according to any one of claims 15-17, wherein, before the sampling data is input into the trained deep learning network for image reconstruction, the computer-readable instructions are processed by the processor The following steps are also implemented during execution:

Acquire a plurality of training samples, each training sample includes a fully sampled sample image of the imaging target and sample sample data obtained by processing the imaging target based on the preset sampling mode; the fully sampled sample image is used to communicate with the The predicted images output by the initial deep learning network are compared; the interval between two adjacent sampling points in the preset sampling mode is not uniform;

According to a plurality of the training samples, iterative training is performed on the initial deep learning network, and the training is stopped when a preset condition is met, to obtain a trained deep learning network.
18. The computer-readable storage medium of claim 18, wherein there are multiple imaging targets used for training;

The obtaining multiple training samples includes:

Acquiring a fully sampled sample image corresponding to each imaging target, the fully sampled sample image being collected by a medical imaging device;

Acquiring sample sampling data of each imaging target based on the preset sampling mode, where the sample sampling data is under-sampling frequency domain data in a non-Cartesian coordinate system;

The fully sampled sample image and sample sample data corresponding to each imaging target are used as a training sample.
18. The computer-readable storage medium of claim 18, wherein the initial deep learning network comprises: a coordinate conversion convolution module, a density compensation convolution module, and an image reconstruction convolution module that are sequentially cascaded;

The iterative training of the initial deep learning network based on a plurality of the training samples, and stopping the training when a preset condition is met, obtaining the trained deep learning network includes:

Initialize the model parameters of the initial deep learning network;

Performing a convolution operation on the sample sample data in the training sample by the coordinate conversion convolution module to obtain under-sampled non-uniform sample data in a Cartesian coordinate system;

Performing density compensation on the under-sampled non-uniform sample data in the Cartesian coordinate system by the density compensation convolution module, and generate corresponding under-sampled uniform sample data;

Performing a convolution operation on the under-sampled uniform sample data by the image reconstruction convolution module to generate a predicted image corresponding to the under-sampled uniform sample data;

If the preset conditions are not currently met, update the model parameters of the current deep learning network, and return to perform the convolution operation of the sample sample data in the training sample through the coordinate conversion convolution module to obtain the Cartesian coordinate system The step of under-sampling non-uniform sample data; if the preset conditions are met, save the current model parameters to obtain the deep learning network.