System, method and medical imaging apparatus for generating a three-dimensional medical image based on a CT scout view

Abstract

The application particularly relates to a system, a method and a medical imaging device for generating a three-dimensional medical image based on a CT positioning sheet, the system comprising an acquisition unit, a deep learning unit and a display unit; the deep learning unit comprising a preprocessing module, a network generation module and a network discrimination module, the preprocessing module being used for data processing on input positioning sheet image information; the network generation module comprising an encoder, a feature converter and a decoder, the encoder being used for extracting shallow and deep features of the positioning sheet image after data processing, the feature converter being used for dimension migration and feature conversion of the 2D deep features extracted by the encoder to a three-dimensional high-latitude feature space to obtain 3D deep features, and the decoder being used for expanding and mining the 3D deep features after the feature converter to generate a three-dimensional medical image. The three-dimensional medical image generated by the application has determined CT value information, contains rich medical anatomy information features, and can effectively distinguish medical organ tissues therein.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging technology, specifically relating to systems, methods, and medical imaging equipment for generating three-dimensional medical images based on CT localization films. Background Technology

[0002] CT localization film scanning is an indispensable part of the CT examination process. Its main purpose is to determine the approximate location of the problem to be detected and parameters such as the Region of Interest (ROI) before the formal CT tomographic scan. The main process involves the CT tube rotating to a certain angle to emit X-rays of a specific energy range. These rays interact with matter within their effective range, causing energy attenuation and changes. The detector collects the corresponding changed signals for rapid reconstruction. Therefore, CT localization films are two-dimensional images that primarily characterize the attenuation and changes in the accumulated matter along the radiation path.

[0003] Although CT localization films are not three-dimensional images, their physical principle is based on the reflection of X-rays after attenuation throughout space. They record the cumulative attenuation integral of X-rays in space, a physical quantity reflecting the thickness of the material encountered by the X-rays. Therefore, in a sense, they can be used for three-dimensional image estimation. However, how to generate effective three-dimensional images using two-dimensional localization film information is a pressing problem that needs to be solved. Existing research has yielded relevant results. For example, Phlipp Henzler proposed a deep learning model for estimating skeletal models using X-ray images, but this study focuses on animal skull data, and the model outputs geometric phantom information, lacking effective anatomical information of tissues and organs, resulting in unclear medical features. Additionally, Weinan Song et al. proposed an image generation model for X-ray images of teeth and jawbones, but this is limited to human teeth and jawbones, and the input data is a tiled 2D image. It also requires input from optical images used in dental examinations, making the model very cumbersome and limiting its application scenarios. Similarly, patent document CN113272869A also discloses 3D shape reconstruction from positioning films in medical imaging, which is only based on the transformation of 2D images into 3D shapes. The goal of this implementation is more inclined to the distribution of three-dimensional geometric shapes, and cannot include CT value information or any radiographic information of tissues and organs. Moreover, its generation network only includes a positioning film encoder for input positioning film and a decoder for outputting 3D shapes from the output of the decoder, which makes the texture features of the output 3D shape limited. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, the purpose of this invention is to provide a method, system, and medical imaging equipment for generating three-dimensional medical images based on CT localization images. This invention combines Generative Adversarial Networks (GANs) and employs deep learning technology to learn features from localization images, generating relatively realistic three-dimensional medical images. These images contain definite CT value information and rich radiographic anatomical image information. The generated images clearly show the characteristics of medical organs and tissues, facilitating conversion into medical DICOM files and effectively supporting medical diagnosis. Furthermore, it requires no additional image information input, has a simple structure, intelligently generates three-dimensional images, and saves time and resource costs.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] A system for generating three-dimensional medical images based on CT scans includes:

[0007] The acquisition unit is used to acquire image information from the positioning patch.

[0008] The deep learning unit is configured with a generative adversarial network model to generate three-dimensional medical images based on the input localization image information.

[0009] A display unit is used to display the generated three-dimensional medical images;

[0010] The deep learning unit includes a preprocessing module, a network generation module, and a network discrimination module. The preprocessing module is used to process the input localization patch image information.

[0011] The network generation module includes an encoder, a feature converter, and a decoder. The encoder is used to extract shallow and deep features from the localization image after data processing. The feature converter is used to perform dimensionality transfer and feature transformation on the 2D deep features extracted by the encoder to a three-dimensional high-dimensional feature space to obtain 3D deep features. The decoder is used to expand and mine the 3D deep features after the feature converter to generate three-dimensional medical images.

[0012] The network discrimination module is used to judge whether the 3D medical images generated by the network generation module are real or fake, in order to train the generative adversarial network model until the model converges.

[0013] As a preferred embodiment, the preprocessing module includes geometric range cropping, numerical truncation, and normalized linear transformation performed sequentially. Geometric range cropping is used to select the target range of the image based on the starting range of the scanning coordinates and the spatial units of pixels in the X and Z directions in the positioning patch image information. Numerical truncation is used to truncate the image based on the pixel values ​​of the target range. Normalized linear transformation is used to perform a normalized linear transformation on the pixel values ​​of the truncated image to obtain the positioning patch image to be input into the network generation module.

[0014] As a preferred embodiment, the encoder includes a 2D convolution module, a 2D downsampling module, and a first activation function connected in sequence.

[0015] As a preferred embodiment, the 2D downsampling submodule includes a residual module and a 2D pooling layer connected in sequence. The residual module consists of several residual units connected in sequence. Each residual unit consists of several processing sub-units connected in series. Each processing sub-unit includes a convolutional layer, an IN layer, and a LeakyReLU function connected in sequence. The input features of the 2D downsampling submodule are superimposed in the last processing sub-unit and then processed by the LeakyReLU function in the last processing sub-unit to obtain the input of the feature converter.

[0016] As a preferred embodiment, the feature converter includes a convolutional layer, a projection unit module, and a second activation function connected in sequence. The projection unit module consists of a fully connected layer and a reshaping layer connected in sequence.

[0017] As a preferred embodiment, the decoder includes a 3D convolution, a 3D upsampling module and a third activation function connected in sequence, wherein the 3D upsampling module includes a residual module and a 3D deconvolution connected in sequence.

[0018] As a preferred embodiment, the network discrimination module includes a generated image, a real image, a discriminator, and a discrimination result. The discriminator is used to judge the authenticity of the generated image and the real image to obtain the discrimination result.

[0019] The discriminator includes a 3D downsampling module, which comprises a residual module and a 3D pooling layer connected in sequence.

[0020] This invention also provides a method for generating three-dimensional medical images based on CT scan images, applicable to the system described in any of the above embodiments, the method comprising the following steps:

[0021] S1. Acquire CT localization images;

[0022] S2. Input the CT localization image into the generative adversarial network model to generate a three-dimensional medical image;

[0023] S3. Display the generated three-dimensional medical image.

[0024] As a preferred embodiment, the training process of the generative adversarial network model includes:

[0025] S01. Obtain the two-dimensional positioning film and its corresponding three-dimensional CT image;

[0026] S02. Process the two-dimensional positioning film and its corresponding three-dimensional CT image to obtain the training dataset;

[0027] S03. Establish a generative adversarial network model for converting two-dimensional localization films into three-dimensional CT images;

[0028] S04. Train the generative adversarial network model using the training dataset and loss function until the model converges.

[0029] The present invention also provides medical imaging equipment configured with the system or application described in any of the above embodiments as described in any of the above embodiments.

[0030] Compared with the prior art, the beneficial effects of this invention are:

[0031] The present invention relates to a method, system, and medical imaging device for generating three-dimensional medical images based on CT localization films. By utilizing generative adversarial networks (GANs) in deep learning, three-dimensional image features are generated from two-dimensional localization film information, resulting in a three-dimensional image with CT value information. This image contains rich medical reflective imaging and anatomical information features, has good image resolution, and can effectively distinguish medical organs and tissues. In addition, the method is simple, does not rely on other modal information, and the results are stable and accurate, saving time and resource costs. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the architecture of a deep learning unit according to an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of the preprocessing module architecture according to an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the architecture of the network generation module according to an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the encoder architecture according to an embodiment of the present invention;

[0036] Figure 5 This is a schematic diagram of the architecture of the 2D downsampling module according to an embodiment of the present invention;

[0037] Figure 6 This is a schematic diagram of the residual module and residual unit according to an embodiment of the present invention;

[0038] Figure 7 This is a schematic diagram of the feature converter architecture according to an embodiment of the present invention;

[0039] Figure 8 This is a schematic diagram of the projection unit module according to an embodiment of the present invention;

[0040] Figure 9 This is a schematic diagram of the decoder architecture according to an embodiment of the present invention;

[0041] Figure 10This is a schematic diagram of the architecture of the 3D upsampling module according to an embodiment of the present invention;

[0042] Figure 11 This is a schematic diagram of the architecture of the network discrimination module according to an embodiment of the present invention;

[0043] Figure 12 This is a schematic diagram of the architecture of the 3D downsampling module according to an embodiment of the present invention;

[0044] Figure 13 This is a flowchart of the training process of the generative adversarial network model according to an embodiment of the present invention;

[0045] Figure 14 This is a diagram showing the processing details of the generative adversarial network model in an embodiment of the present invention;

[0046] Figure 15 This is a comparison image of the effect between the image generated by the embodiment of the present invention and the real image; wherein, a is a cross-sectional view of a certain layer of the real image, b is a cross-sectional view of the same position of the 3D image generated by the positioning piece using the above method of the embodiment of the present invention, c is a coronal bitmap of the real image, and d is a coronal bitmap of the same position of the 3D image generated by the positioning piece using the above method of the embodiment of the present invention. Detailed Implementation

[0047] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0048] This invention constructs a generative adversarial network model for generating 3D medical images from CT scans. It is an efficient image feature generation and transformation model based on improved generative adversarial networks, and has the following characteristics:

[0049] (1) The model is an improvement on the generative adversarial network model in the field of deep learning and has some characteristics of the deep learning generative adversarial network model.

[0050] (2) The model has the function of generating, reconstructing or estimating three-dimensional medical images from two-dimensional images;

[0051] (3) The three-dimensional medical images (i.e. 3D medical images) output by the model not only have the surface properties of a solid in general geometry, but the image pixel values ​​also have medical radioactivity information;

[0052] (4) The three-dimensional medical images generated by the model have features such as the lines and contours of organs and tissues.

[0053] Specifically, the following describes in detail the components and processing methods of the system for generating three-dimensional medical images based on CT positioning films according to embodiments of the present invention.

[0054] The system for generating three-dimensional medical images based on CT localization films according to embodiments of the present invention includes an acquisition unit, a deep learning unit, and a display unit.

[0055] The acquisition unit is used to acquire positioning film image information; the acquired CT positioning film image information can be a single positioning film or a combination of multiple positioning films obtained from CT scanner scans at any angle, such as 0° or 90°. The acquisition unit can utilize the X-ray imager of the CT scanner.

[0056] The deep learning unit is configured with a generative adversarial network (GAN) model to generate 3D medical images based on input localization slice image information. Functionally, the deep learning unit comprises three main parts: an image preprocessing component, a generative network component, and a network discriminative component. More specifically, such as... Figure 1 As shown, the deep learning unit includes a preprocessing module 100, a network generation module 200, and a network discrimination module 300.

[0057] The preprocessing module 100 is used to process the input localization slice image information or 3D real medical images, and includes the following three functions: geometric range cropping, numerical truncation, and normalized linear transformation. More specifically, such as... Figure 2 As shown, the preprocessing module 100 includes geometric range clipping 101, numerical truncation 102, and normalized linear transformation 103 performed sequentially.

[0058] Geometric range cropping is used to select a target area of ​​the image based on the starting range of the scan coordinates in the positioning patch image information and the spatial units of pixels in the X and Z directions, as follows:

[0059] The main operation of Geometric Range Clipping 101 is to read the starting range value of the scan coordinates from the CT localization film DICOM file (xscan). start zscan start xscan end zscan end The spatial units of pixels in the X and Z directions, spaceX and spaceZ, are transformed according to the following formula:

[0060]

[0061] The relevant parameters idx, △idx, idz, and △idz are obtained, and finally, the image within the range of [idx, idx+△idx] and [idz, idz+△idz] is selected.

[0062] Numerical truncation 102 is used to truncate the image based on the pixel values ​​of the target area, mainly based on the truncation range [v]. min ,v max The pixel value v of the positioning patch image is judged and truncated according to the following formula:

[0063]

[0064] Normalized linear transformation 103 is used to perform a normalized linear transformation on the truncated image pixel values ​​to obtain the localization patch image to be input into the network generation module. Specifically, normalized linear transformation 103 mainly performs a normalized linear transformation on the image pixel values ​​v of the localization patch and the 3D image. For explanation, Gaussian normalized linear transformation is used, as shown in the following formula:

[0065]

[0066] The network generation module of this invention includes an encoder, a feature converter, and a decoder. The encoder extracts shallow and deep features from the localization image after data processing. The feature converter performs dimensionality transfer and feature transformation on the 2D deep features extracted by the encoder to a three-dimensional high-dimensional feature space to obtain 3D deep features. The decoder expands and mines the 3D deep features processed by the feature converter to generate a three-dimensional medical image. The encoder includes a 2D convolutional module, a 2D downsampling module, and a first activation function connected in sequence. The 2D downsampling submodule includes a residual module and a 2D pooling layer connected in sequence. The residual module consists of several residual units connected in sequence, each consisting of several small processing units. Each small processing unit includes a convolutional layer, an IN layer, and a LeakyReLU function connected in sequence. The input features of the last small processing unit and the 2D downsampling submodule are superimposed, and then processed by the LeakyReLU function in the last small processing unit to obtain the input of the feature converter. The feature converter includes a convolutional module, a projection unit module, and a second activation function connected in sequence. The projection unit module consists of a fully connected layer and a reshaping layer connected in sequence. The decoder includes a 3D convolutional module, a 3D upsampling module, and a third activation function connected in sequence. The 3D upsampling module includes a residual module and a 3D deconvolution module connected in sequence.

[0067] More specifically, such as Figure 3 As shown, the network generation module 200 consists of the following module units: input image 201, encoder 202, feature converter 203, decoder 204 and generated image 205.

[0068] The main function of encoder 202 is to extract shallow and deep features from the localization patch image through multiple combinations and stacking operations such as convolution, downsampling, pooling, and activation functions in deep learning. The encoder's typical feature extraction is achieved through a 2D image downsampling module, whose main characteristics are the use of residual modules and 2D image pooling. For example... Figure 4 As shown, the encoder 202 mainly consists of a 2D convolution module 202_1, a 2D downsampling module 202_2, and an activation function 202_3, which are basic feature processing units. Its function is to perform feature analysis processing on the input image. For example... Figure 5 As shown, the processing of the 2D downsampling module 202_2 consists of a residual module K0 and a 2D pooling layer K2. The residual module K0 mainly extracts and expresses more potential 2D image features, while the 2D pooling layer K2 performs pooling operations on the 2D features to summarize and organize the features.

[0069] The following provides a detailed description of the residual module K0. The residual module consists of multiple residual units stacked sequentially. The main characteristic of each residual unit is that it undergoes multiple convolution + IN layers + LeakyReLU operations, and finally sums the input and output before passing the sum through LeakyReLU again for output. Figure 6 As shown, the residual module K0 is a composite processing unit consisting of repeated residual units K1. Residual unit K1 is composed of three basic processing steps: convolutional layer, IN layer, and LeakyReLU function, stacked multiple times in sequence. Finally, it is superimposed on the input features of the entire module and then processed by LeakyReLU to obtain the final output. The specific number of overlapping residual units K0, convolutional layers, IN layers, and LeakyReLU functions can be chosen by the specific data sample space and is not subject to any specific limitation.

[0070] The feature converter 203 in the network generation module 200 functions to transfer and transform the diverse and complex 2D deep features obtained from the encoder to a 3D high-dimensional feature space. The main feature of the feature converter is that it achieves the aforementioned function through a projection unit module, which consists of two main functional modules: a fully connected layer and a reshaping module. Figure 7 As shown, the feature converter 203 mainly consists of convolution 203_1, projection unit module 203_2, and activation function 203_3. The primary function of the feature converter 203 is to synthesize the diverse and complex 2D deep features obtained from the encoder 202, performing dimensionality transfer and feature transformation onto a 3D high-dimensional feature space. This function is mainly accomplished by the projection unit module 203_2, which is primarily composed of fully connected layers and reshaping layers, such as... Figure 8 As shown.

[0071] The decoder 204 in the network generation module 200 functions to expand and mine deeper 3D features after the feature converter, abstracting and extracting them into more detailed and figurative features. The main features enabling this function are the residual module and 3D deconvolution operation. Specifically, as... Figure 9 As shown, decoder 204 is the process for expanding and mining deeper 3D features. Its main components consist of 3D convolution 204_1, 3D upsampling module 204_2, and activation function 204_3. The processing is primarily completed by 3D upsampling module 204_2, as shown below. Figure 10 As shown, the 3D upsampling module 204_2 is mainly composed of the residual module K0 and 3D deconvolution.

[0072] The network discrimination module 300 in this embodiment of the invention is used to judge the authenticity of the three-dimensional medical images generated by the network generation module and real images, so as to train the generative adversarial network model until the model converges. The real images required for network training are three-dimensional medical images obtained from CT scanner scans.

[0073] Specifically, the network discrimination module 300 is a discrimination network whose main function is to distinguish between the generated image obtained by the network generation module 200 and the real image to determine their authenticity. Figure 11 As shown, the network discrimination module 300 mainly consists of a generated image 205, a real image 301, a discriminator 302, and a discrimination result 303. The most important discrimination function is performed by the discriminator 302. The discriminator's main function is to effectively distinguish between the generated 3D medical image and the real 3D medical image during network training, thereby assisting the entire network in learning its feature representation capabilities more comprehensively. Its main components are a residual module and a 3D pooling layer. Figure 12 As shown, the main process of the discriminator 302 is implemented by the 3D downsampling module 302_1, which mainly consists of the residual module K0 and the 3D pooling layer.

[0074] The display unit in this embodiment of the invention is used to display the generated three-dimensional medical image.

[0075] For the generative adversarial network model in this embodiment of the invention, the network training incorporates a three-part loss function: the main loss, the network generation loss, and the network discriminant loss, as detailed below:

[0076] L = L main +ε*L gen +η*L adv

[0077] Among them, L main It mainly consists of multi-scale structural similarity coefficients, and the network generation loss L genIt is mainly used to evaluate the L2 loss between the image generated by the generative network and the real image, and the network discrimination loss L. adv It primarily measures the distinction between generated and real images.

[0078] ε and η are both positive real numbers less than 1, and their specific values ​​can be chosen freely depending on the specific training task. Additionally, the main loss L... main Network generation loss L gen and network discrimination loss L adv They are defined as follows:

[0079]

[0080] The members l, c, and s are defined as follows:

[0081]

[0082] The above parameters are explained as follows: L main It mainly consists of multi-scale structural similarity coefficients (SSIM), where I and G are both 3D images. I is a real 3D medical image, and G is a generated image output by the network generation module 200 of this embodiment of the invention. It is three-dimensional by default and is also represented by I(M,N,H) and G(M,N,H). The number of pixels in each dimension is M, N, and H respectively. For ease of explanation, i, j, and k are used to represent the pixel index value in each dimension.

[0083] in, C1, C2, and C3 are fixed at values ​​of 6.5, 58, and 29, respectively. I′ and G′ are the 2D images indexed by I and G to each dimension i, respectively. μ is the operation function for calculating the mean of the images, and σ is the operation function for calculating the standard deviation. The specific meanings are explained according to the subscripts, such as μ I′ The result of taking the mean of image I′, σ G′ This is the result of calculating the standard deviation of the image G′, and it is particularly important to note that σ is... I′G′ The goal is to find the joint covariance of I′ and G′.

[0084] Network generation loss L gen The definition is as follows:

[0085]

[0086] Network generation loss L gen It is mainly used to evaluate the L2 loss between the image generated by the generative network and the real image, that is, to calculate the mean of the sum of squares of the differences between pixels in images I and G.

[0087] Network discrimination loss L advEssentially, it is the Wasserstein loss with a penalty term, defined as follows:

[0088]

[0089] Where E is the L1 loss function, and D is the result of the network discriminant module 300. It is a variable representing a generated image sample obtained by the generator network in each training iteration. These are the real image sample variables used in each training iteration. D(x) and D(x) are the generated image sample variables after passing through the discrimination module, respectively. The result of the discrimination between the real image sample variable x, where It is the weight coefficient of a certain generated image sample variable. It is the gradient operator, which can be interpreted as the change in weights during each training iteration, where λ can be chosen as a positive real value.

[0090] Based on the system for generating three-dimensional medical images based on CT localization films described above in the embodiments of the present invention, the embodiments of the present invention also provide a method for generating three-dimensional medical images based on CT localization films, including the following steps:

[0091] S1. Acquire CT localization images;

[0092] S2. Input the CT localization image into the generative adversarial network model to generate a three-dimensional medical image;

[0093] S3. Display the generated three-dimensional medical image.

[0094] Among them, such as Figure 13 and 14 As shown, the training process of a generative adversarial network model includes:

[0095] S01. Obtain the two-dimensional positioning film and its corresponding three-dimensional CT image; the image data involved in this step are all raw data, stored in DICOM data format;

[0096] S02. Data processing is performed on the two-dimensional positioning films and their corresponding three-dimensional CT images to obtain a training dataset; specifically, the obtained data is resampled, which mainly includes two processing steps:

[0097] First, the data is interpolated to the same matrix dimension. Since the sizes of the localization images and corresponding 3D CT data obtained from different patients are not the same, all data are interpolated to ensure the neural network can train and converge properly. Localization images are interpolated using bilinear interpolation to a predetermined size; 3D CT data are interpolated using cubic interpolation to a predetermined size.

[0098] Secondly, the CT values ​​of the data are normalized. Since the distribution of CT values ​​in medical images is wider than that of pixel values ​​in natural images, the localization image data and 3D data are normalized in order for the network to find the optimal solution more quickly.

[0099] S03. Establish a generative adversarial network model for converting two-dimensional localization films into three-dimensional CT images, including two models: a generator and a discriminator.

[0100] S04. Train the generative adversarial network model using the training dataset and the set loss function until the model converges.

[0101] Specifically, set the loss function L. gen L adv and L main The generator and discriminator are trained alternately until convergence, so as to obtain the weights of the trained generator network, that is, to obtain and save the weights of the entire network.

[0102] This invention also provides medical imaging equipment configured with the system described in this invention or applying the method described in this invention. Examples of medical imaging equipment include CT scanners.

[0103] like Figure 15 As shown, the three-dimensional medical image generated by the method described in the embodiments of the present invention is basically consistent with the real image, thereby verifying that the method described in the embodiments of the present invention has the ability to generate, reconstruct, or estimate three-dimensional medical images from two-dimensional positioning images.

[0104] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A system for generating a three-dimensional medical image based on a CT scout view, characterized by, include: The acquisition unit is used to acquire image information from the positioning patch. The deep learning unit is configured with a generative adversarial network model to generate three-dimensional medical images based on the input localization image information. A display unit is used to display the generated three-dimensional medical images; The deep learning unit includes a preprocessing module, a network generation module, and a network discrimination module. The preprocessing module is used to process the input localization patch image information. The network generation module includes an encoder, a feature converter, and a decoder. The encoder is used to extract shallow and deep features from the localization image after data processing. The feature converter is used to perform dimensionality transfer and feature transformation on the 2D deep features extracted by the encoder to a three-dimensional high-dimensional feature space to obtain 3D deep features. The decoder is used to expand and mine the 3D deep features after the feature converter to generate three-dimensional medical images. The network discrimination module is used to judge whether the 3D medical images generated by the network generation module are real or fake, in order to train the generative adversarial network model until the model converges; The preprocessing module includes geometric range cropping, numerical truncation, and normalized linear transformation performed sequentially. Geometric range cropping is used to select the target range of the image based on the starting range of the scanning coordinates and the spatial units of pixels in the X and Z directions in the positioning patch image information. Numerical truncation is used to truncate the image based on the pixel values ​​of the target range. Normalized linear transformation is used to perform normalized linear transformation on the pixel values ​​of the truncated image to obtain the positioning patch image to be input into the network generation module. The encoder includes a 2D convolution module, a 2D downsampling module, and a first activation function connected in sequence. The 2D downsampling module includes a residual module and a 2D pooling layer connected in sequence. The residual module consists of several residual units connected in sequence. Each residual unit consists of several processing units connected in series. Each processing unit includes a convolutional layer, an IN layer and a LeakyReLU function connected in sequence. The input features of the 2D downsampling module are superimposed in the last processing unit and then processed by the LeakyReLU function in the last processing unit to obtain the input of the feature converter. The feature converter includes a convolutional and projection unit modules connected in sequence, and a second activation function. The projection unit module consists of a fully connected layer and a reshaping layer connected in sequence. The decoder includes a 3D convolution, a 3D upsampling module and a third activation function connected in sequence. The 3D upsampling module includes a residual module and a 3D deconvolution connected in sequence. The network discrimination module includes a generated image, a real image, a discriminator, and a discrimination result. The discriminator is used to judge whether the generated image and the real image are real or fake in order to obtain the discrimination result. The discriminator includes a 3D downsampling module, which comprises a residual module and a 3D pooling layer connected in sequence.

2. Method for generating a three-dimensional medical image on the basis of a CT scout view, for use in a system as claimed in claim 1, characterized in that The method includes the following steps: S1. Acquire CT localization images; S2. Input the CT localization image into the generative adversarial network model to generate a three-dimensional medical image; S3. Display the generated three-dimensional medical image.

3. The method of generating a three-dimensional medical image based on a CT positioning sheet according to claim 2, wherein, The training process of the generative adversarial network model includes: S01, acquire two-dimensional positioning film and its corresponding three-dimensional CT image; S02, data processing is carried out to two-dimensional positioning film and its corresponding three-dimensional CT image, and training data set is obtained; S03, a generative adversarial network model for converting two-dimensional positioning film into three-dimensional CT image is established; S04, the generative adversarial network model is trained by using training data set and loss function until the model converges.

4. Medical imaging apparatus, characterized in that Configuring the system as claimed in claim 1 or applying the method as claimed in claim 2 or 3.

Images