Method and data processing device for constructing deep learning model for generating virtual contrast-enhanced CT image

Abstract

A method for constructing a deep learning model for generating a virtual contrast-enhanced computed tomography (CT) image comprises the steps in which a data processing device: receives a pre-contrast CT image and an actual contrast CT image of the pre-contrast CT image; inputs the pre-contrast CT image to a deep learning model to generate a virtual contrast-enhanced CT image; and updates parameters of the deep learning model by using a result obtained by comparing the actual contrast CT image and the virtual contrast-enhanced CT image.

Description

Method for building a deep learning model for generating virtual contrast-enhanced CT images and a data processing device

[0001] The technology described below is a technique for generating virtual contrast-enhanced CT images using a deep learning model.

[0002] Computed Tomography (CT) provides cross-sectional images or reconstructed three-dimensional images of internal structures, and is used to diagnose various diseases. Meanwhile, contrast agents are drugs administered to patients to clearly distinguish between lesions and normal tissue during CT scans. For example, contrast agents are used to diagnose hilar lymphadenopathy or to identify hilar lymph nodes in a chest CT scan.

[0003] However, contrast agents are difficult to use in patients with a history of adverse effects, pregnancy, radioactive iodine-based thyroid treatment, or chronic or acute kidney disease. Therefore, techniques for generating virtual contrast-enhanced CT images using deep learning models have been studied.

[0004] Most conventional techniques for generating contrast-enhanced CT images utilize generative adversarial models. These techniques require pre- and post-contrast CT images as training data. However, these CT images may require registration due to factors such as the subject's breathing or movement. In particular, chest CTs require a registration step due to misalignment between pre- and post-contrast images caused by breathing movement. Therefore, conventional techniques require an additional registration step, and even after registration, certain discrepancies persist, making it difficult to obtain high-performance deep learning models.

[0005] The technology described below is intended to provide a technique for generating a virtual contrast-enhanced CT without a registration step using a dual-energy CT image.

[0006] A method for constructing a deep learning model for generating a virtual contrast-enhanced CT image includes a step in which a data processing device receives a pre-contrast CT (Computed Tomography) image and an actual contrast-enhanced CT image for the pre-contrast CT image, a step in which the data processing device inputs the pre-contrast CT image into a deep learning model to generate a virtual contrast-enhanced CT image, and a step in which the data processing device updates parameters of the deep learning model using a result of comparing the actual contrast-enhanced CT image with the virtual contrast-enhanced CT image.

[0007] A data processing device for generating a virtual contrast-enhanced CT image includes an interface device for receiving a pre-contrast CT image of a specific subject, a storage device for storing a deep learning model for generating a contrast-enhanced CT image, and a computing device for inputting the pre-contrast CT image into the deep learning model to generate a virtual contrast-enhanced CT image.

[0008] Conventional alignment techniques are divided into rigid alignment and non-rigid alignment. Rigid alignment requires frequent user input or similarity measurements during the alignment stage, resulting in high complexity, while non-rigid alignment has the problem of distorting the resulting image.

[0009] The technology described below utilizes dual-energy CT images to generate contrast-enhanced CT images capable of distinguishing detailed features, such as lymph nodes, without the need for complex alignment steps during the training data generation process. Therefore, the technology described below can provide a diagnostic aid necessary for effectively diagnosing lymph node lesions.

[0010] Figure 1 is an example of a system providing contrast-enhanced chest CT images.

[0011] Figure 2 is an example of the learning process of a deep learning model that generates contrast-enhanced chest CT images.

[0012] Figure 3 is an example of the structure of the generator and discriminator of an adversarial generative model.

[0013] Figure 4 is another example of the training process of a deep learning model that generates contrast-enhanced chest CT images.

[0014] Figure 5 is an example of the structure of a residual U-net-based model.

[0015] Figure 6 is an example of the structure of a convolutional block attention module (CBAM).

[0016] Figure 7 is an example of a contrast-enhanced chest CT image generated using a deep learning model.

[0017] Figure 8 is an example of a data processing device that generates a contrast-enhanced chest CT image.

[0018] The technology described below is susceptible to various modifications and embodiments. Specific embodiments are illustrated and described in detail in the drawings. However, this does not limit the technology described below to specific embodiments, and it should be understood that all modifications, equivalents, and alternatives fall within the spirit and scope of the technology described below.

[0019] Terms such as first, second, A, and B may be used to describe various components, but these components are not limited by these terms and are used solely to distinguish one component from another. For example, without departing from the scope of the technology described below, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component. The term "and / or" includes a combination of a plurality of related described items or any of a plurality of related described items.

[0020] As used herein, the singular expressions should be understood to include the plural expressions unless the context clearly dictates otherwise, and the term "comprises" and the like should be understood to mean the presence of a described feature, number, step, operation, component, part, or combination thereof, but not to exclude the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

[0021] Before going into a detailed description of the drawings, it should be made clear that the division of components in this specification is merely a division based on the main function of each component. In other words, two or more components described below may be combined into a single component, or a single component may be further subdivided into two or more components with more detailed functions. In addition to its own main function, each component described below may additionally perform some or all of the functions of other components, and of course, some of the main functions of each component may be exclusively performed by other components.

[0022] Additionally, in performing a method or method of operation, each process constituting the method may occur in a different order than the stated order, unless the context clearly indicates a specific order. That is, each process may occur in the same order as the stated order, may be performed substantially simultaneously, or may be performed in the opposite order.

[0023] The technology described below uses a deep learning model to generate virtual contrast-enhanced CT images. The technology described below is not limited to specific diseases or body parts. However, since registration is essential for chest CT images, the description below focuses on chest CT images.

[0024] The following describes a data processing device that builds a deep learning model for generating contrast-enhanced chest CT images. The data processing device is a computing device capable of processing CT images and training a deep learning model using training images. The data processing device can be implemented as a PC, a networked server, or a chipset embedded with a dedicated program.

[0025] Deep learning models can utilize a variety of models. For example, deep learning models can include U-net-based models and adversarial generative models.

[0026] Below, we describe how a client device generates a contrast-enhanced chest CT image using a trained deep learning model. The client device is a computing device capable of performing CT image processing and generating a virtual contrast-enhanced CT image using a deep learning model. The client device can be implemented as a PC, a networked server, a smart device, or a chipset embedded with a dedicated program.

[0027] Fig. 1 is an example of a system (100) that provides contrast-enhanced chest CT images. In Fig. 1, the data processing device is illustrated as an example of a computer terminal (130) and a server (140).

[0028] The CT equipment (110) generates a chest CT image of a subject (patient). The CT equipment (110) can store the generated chest CT image in an EMR (Electronic Medical Record, 120) or a separate database (DB).

[0029] A user (A) can input a chest CT image using a computer terminal (130) and generate a contrast-enhanced chest CT image. The computer terminal (130) can input a chest CT image from a CT device (110) or an EMR (120) via a wired or wireless network. In some cases, the computer terminal (130) may be a device physically connected to the CT device (110). The computer terminal (130) can preprocess the chest CT image. Here, the preprocessing may be at least one of operations such as changing the image size, changing the resolution, cropping, and removing noise. The computer terminal (130) can input the chest CT image into a pre-trained deep learning model. The computer terminal (130) can generate a contrast-enhanced chest CT image using the deep learning model. User A can view the contrast-enhanced chest CT image on the computer terminal (130).

[0030] The server (140) can receive a chest CT image from a CT device (110) or an EMR (120). The server (140) can preprocess the chest CT image. The server (140) can input the chest CT image into a pre-trained deep learning model. The server (140) can generate a contrast-enhanced chest CT image using the deep learning model. The server (140) can transmit the contrast-enhanced chest CT image to the terminal of user A.

[0031] The computer terminal (130) and / or server (140) may store the contrast-enhanced chest CT image in the EMR (120).

[0032] The researcher explains the process of building the aforementioned deep learning model.

[0033] The researchers built and validated two types of models: a Generative Adversarial Networks (GAN)-based model and a Residual U-net-based model.

[0034] The researchers used dual-energy chest CT images collected by their institution (Samsung Seoul Hospital) from 2,082 individuals. Data from 1,869 individuals served as training and validation data, 153 individuals served as test data for image evaluation, and the remaining 60 individuals served as external validation data.

[0035] First, we explain the process of building a GAN-based model.

[0036] Figure 2 illustrates an example of a learning process (200) of a deep learning model for generating contrast-enhanced chest CT images. In Figure 2, the deep learning model is illustrated as an adversarial generative model such as a Generative Adversarial Network (GAN).

[0037] The image database (DB) stores training images used in the learning process. The image DB stores dual-energy CT images. That is, the deep learning model is trained using dual-energy CT images. The training images include pre-contrast chest CT images and contrast-enhanced chest CT images. The training images consist of pairs of pre-contrast chest CT images and contrast-enhanced chest CT images of the same subject. Figure 2 illustrates the training process of a deep learning model using a single training image. The data processing unit repeatedly performs the training process using multiple training images to build a deep learning model generator.

[0038] The data processing device receives pre-contrast chest CT images and contrast-enhanced chest CT images of the same subject from the image database (210). The input images may have a size of 512×512 pixels. The pre-contrast chest CT image is the image input to the deep learning model, and the contrast-enhanced chest CT image is the correct label value.

[0039] The data processing device can preprocess the input image in a predetermined manner (220). For example, the data processing device can crop a region of interest from the input image. The data processing device can also crop the center of the input CT image to a predetermined size. The cropped image can have a size of 256×256 pixels. Alternatively, the data processing device can extract the region of interest using an image processing technique or a segmentation model.

[0040] The data processing device inputs a pre-contrast chest CT image into the generator. The generator generates a contrast-enhanced image (230). The generated contrast-enhanced chest CT image may have a size of 256×256 pixels.

[0041] A discriminator compares the contrast-enhanced CT image generated by the generator with the labeled contrast-enhanced chest CT image to determine whether the generated contrast-enhanced CT image is a real image or a fake image (240).

[0042] The data processing device updates the parameters of the generator and the discriminator based on the discriminator's judgment result (250). At this time, the loss function is set so that (i) the discriminator can discriminate between the virtual contrast-enhanced image (fake) and the real contrast-enhanced image (real) generated by the discriminator as much as possible, and (ii) the generator can simultaneously determine that the discriminator is a real image for the virtual contrast-enhanced image.

[0043] Figure 3 illustrates an example of the structure of the generator and discriminator of an adversarial generative model. Figure 3(A) illustrates an example of the generator, and Figure 3(B) illustrates an example of the discriminator. Figure 3 illustrates an example of receiving a 256×256 pixel pre-contrast CT image as input and generating a contrast-enhanced CT image of the same size.

[0044] The generator of the adversarial generative model is a residual network combined with a convolutional block attention module (CBAM).

[0045] The generator comprises multiple convolutional layers and multiple convolutional blocks, each of which is composed of a CBAM. Figure 3 illustrates the size (e.g., 3×3) and number (e.g., 32) of filters applied to each convolutional layer. The filter sizes and number of filters in Figure 3 may vary depending on the size of the input image.

[0046] Each convolution block has a structure that combines a convolution layer with a CBAM. This allows the convolution block to output a feature map with enhanced key information from the convolution process. The feature map output from the CBAM can be passed on to the next convolution layer and other layers. The structure and operation of the CBAM are described below.

[0047] The generator has a structure that combines upsampling layers followed by convolution blocks.

[0048] The discriminator of the adversarial generative model receives the correct answer (actual contrast image) for the contrast-enhanced chest CT image and the virtual contrast-enhanced chest CT image output by the generator as input, and outputs a value (0 or 1) that determines whether the image is an actual contrast-enhanced chest CT image or a virtual contrast-enhanced chest CT image. The discriminator consists of multiple convolutional layers.

[0049] Meanwhile, the aforementioned GAN model can also be built as a U-net-based GAN.

[0050] We also describe the construction of residual U-net-based models. All residual U-net-based models were built using a large training dataset.

[0051] Figure 4 illustrates another example of the learning process (300) of a deep learning model for generating contrast-enhanced chest CT images. In Figure 4, the deep learning model is illustrated as an example of a residual U-net-based model.

[0052] The image database (DB) stores training images used in the learning process. The image DB stores dual-energy CT images. That is, the deep learning model is trained using dual-energy CT images. The training images include actual pre-contrast chest CT images and actual contrast-enhanced chest CT images. The training images consist of pairs of pre-contrast and contrast-enhanced chest CT images of the same subject. Figure 4 illustrates the training process of a deep learning model using a single training image. The data processing device builds a deep learning model by repeatedly performing the training process using multiple training images.

[0053] The data processing device receives pre-contrast chest CT images and contrast-enhanced chest CT images of the same subject from the image database (310). The input images may have a size of 512×512 pixels. The pre-contrast chest CT image is the image input to the deep learning model, and the contrast-enhanced chest CT image is the correct label value.

[0054] The data processing device can preprocess the input image in a predetermined manner (320). For example, the data processing device can crop the region of interest from the input image. The data processing device can also crop the center of the input CT image to a predetermined size. The cropped image can have a size of 256×256 pixels. Alternatively, the data processing device can extract the region of interest using an image processing technique or a segmentation model.

[0055] The data processing device inputs a pre-contrast chest CT image into a residual U-net-based model. The residual U-net-based model generates a virtual contrast-enhanced image (330). The generated virtual contrast-enhanced chest CT image may have a size of 256×256 pixels.

[0056] The data processing device updates the parameters of the residual U-net-based model by comparing the virtual contrast-enhanced chest CT image generated by the residual U-net-based model with the actual contrast-enhanced chest CT image, which is the label (correct answer) (340). The residual U-net-based model can update the model parameters based on cross-entropy on a pixel-by-pixel basis.

[0057] Figure 5 is an example of the structure of a residual U-net-based model. A residual U-net-based model can utilize a residual block integrated with CBAM.

[0058] The reduction path of the residual U-net-based model includes four convolutional layers. Each convolutional layer includes two 3×3 convolutional blocks, one batch normalization (BN) block, one ReLU layer, and one CBAM-integrated residual block. That is, CBAM is used as the residual block.

[0059] The expansion path of the residual U-net-based model has a structure corresponding to the four convolutional layers in the reduction path. The expansion path consists of four convolutional layers. Each convolutional layer includes a 2×2 upsampling block, a CBAM-integrated residual block, and a skip connection connected to the reduction path. The residual U-net-based model has a final 1×1 convolutional layer, and the final layer outputs a virtual contrast-enhanced chest CT image of the same size as the input image.

[0060] Figure 6 is an example of the structure of a convolutional block attention module (CBAM). CBAM generates a channel attention map and a spatial attention map for input features, respectively, and multiplies them by the input feature map, thereby generating a feature map with enhanced key features. CBAM generates a 1D (dimension) channel attention map M for input features F. CThe result of multiplying F' (intermediate feature map) by the spatial attention map M S The enhanced feature map F'' is output by sequentially multiplying them. This can be expressed as in mathematical expression 1 below.

[0061]

[0062] stands for element-wise multiplication.

[0063] The channel attention module includes a max pooling layer that performs max pooling (MaxPool) on the input feature map F and an average pooling layer that performs average pooling (AvgPool) on the input feature map F. Each pooling layer outputs a feature with a size of 1×1 channel. The channel attention module includes a common multi-layer perceptron (MLP) to reduce the number of channels output by the pooling layers. The MLP may include one hidden layer. The channel attention module performs element-wise summation ( , element-wise summation) and sigmoid ( , sigmoid) to perform channel attention map M C Outputs M C has a weight per channel. This process can be expressed as shown in mathematical expression 2 below.

[0064]

[0065] σ is the sigmoid function. W0 and W1 are the MLP weights, and W0∈ , W1∈ The output features of MLP are It has a channel of 1. r is the reduction ratio.

[0066] The spatial attention module may include a max pooling layer that performs max pooling on each channel axis and an average pooling layer that receives F' output from the channel attention module as input and performs average pooling. The consecutive pooling layers concatenate the max pooled features (1 × height × width) and the average pooled features (1 × height × width). The spatial attention module includes a 7 × 7 convolutional layer after the pooling layer. The spatial attention module applies a sigmoid function to the final feature to generate a spatial attention map M. S This process can be expressed as in the mathematical formula 3 below. M S includes spatial weights.

[0067]

[0068] CBAM is a spatial attention map M in F' S Multiply by and output the final enhanced feature map F''.

[0069] The researcher evaluated the model's performance using validation data from the aforementioned data. The researcher built several types of models and compared their performance. Using the aforementioned training data, the researcher built the models below.

[0070] Model classification details Model AU-Net based GANModel B Residual U-Net based GANModel CCBAM based Residual U-Net based GANModel DCBAM based Residual U-Net

[0071] U-Net-based GANs are generative models that use a U-Net-structured discriminator. Residual U-Net-based GANs are generative models that use a U-Net-structured discriminator with residual blocks. CBAM-based residual U-Net-based GANs are generative models that use a residual U-Net-structured discriminator that includes CBAM, as shown in the structure of Figure 5. CBAM-based residual U-Nets are models based on residual U-nets, as described in Figures 4 and 5.

[0072] Using validation data, the researchers quantitatively evaluated the quality of virtual contrast-enhanced chest CT images and actual contrast-enhanced chest CT images. The quantitative evaluation indices used were peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and mutual information (MI).

[0073] PSNRSSIMMAEMSERMSEMIModel A24.626(±0.942)0.819(±0.028)0.041(±0.005)0.024(±0.003)0.104(±0.011)1.516(±0.068)Model B24.875(±0.969)0.826(±0.029)0.045(±0.005)0.024(±0.002)0.104(±0.011)1.538(±0.072)Model C24.962(±0.192)0.814(±0.005)0.049(±0.005)0.024(±0.002)0.113(±0.010)1.540(±0.071)Model D25.082(±0.150)0.833(±0.005)0.040(±0.005)0.023(±0.002)0.102(±0.011)1.568(±0.069)

[0074] Overall, the evaluation results showed that models including CBAM performed well. The residual U-Net (Model D) using CBAM showed the highest performance.

[0075] Figure 7 is an example of a contrast-enhanced chest CT image generated using a deep learning model. Figure 7 shows chest CT images for three patients (patient 1, patient 2, and patient 3). In Figure 7, column A represents an actual non-contrast-enhanced chest CT image, column B represents an actual contrast-enhanced chest CT image, column C represents a virtual contrast-enhanced chest CT image generated with Model A, column D represents a virtual contrast-enhanced chest CT image generated with Model B, column E represents a virtual contrast-enhanced chest CT image generated with Model C, and column F represents a virtual contrast-enhanced chest CT image generated with Model D.

[0076] Model A (U-Net-based GAN) showed a checkerboard effect during the upsampling process, resulting in quality degradation. Model B (Residual U-Net-based GAN) reduced the checkerboard effect but showed limitations in contrast enhancement in important regions such as the mediastinum, hilar lymph nodes, and blood vessels. Model C (CBAM-integrated residual U-Net-based GAN) showed improved contrast compared to Model B. Model D (CBAM-integrated residual U-Net) showed contrast enhancement most similar to actual contrast-enhanced chest CT images.

[0077] FIG. 8 illustrates an example of a data processing device (400) that generates a contrast-enhanced chest CT image. The data processing device (400) corresponds to the aforementioned data processing devices (130 and 140 of FIG. 1). The data processing device (400) may be physically implemented in various forms. For example, the data processing device (400) may take the form of a computer device such as a PC, a network server, a data processing-dedicated chipset, etc.

[0078] The data processing device (400) may include a storage device (410), a memory (420), a computing device (430), an interface device (440), a communication device (450), and an output device (460).

[0079] The storage device (410) can store chest CT images generated by CT equipment.

[0080] The storage device (410) can store a virtual contrast-enhanced chest CT image created using a chest CT image.

[0081] The storage device (410) can store a deep learning model that generates a virtual contrast-enhanced chest CT image. The deep learning model may be a GAN-based model, a U-Net-based model, or the like, as described above. However, in this case, the deep learning model is a model that applies CBAM.

[0082] The storage device (410) can store training data (pre-contrast CT images and actual contrast CT images) for training a deep learning model. The training data may be dual-energy CT images as described above.

[0083] The memory (420) can store data and information generated during the process of the data processing device generating a virtual contrast-enhanced chest CT image.

[0084] The interface device (440) is a device that receives certain commands and data from the outside.

[0085] The interface device (440) can receive a pre-contrast chest CT image from a physically connected input device or an external storage device.

[0086] The interface device (440) can also receive a constructed deep learning model.

[0087] The interface device (440) can also receive learning data for building a deep learning model.

[0088] The interface device (440) can also transmit a virtual contrast-enhanced chest CT image generated using a deep learning model to an external object.

[0089] Meanwhile, the interface device (440) may also receive data or information transmitted via the communication device (450) below.

[0090] A communication device (450) refers to a configuration that receives and transmits certain information through a wired or wireless network.

[0091] The communication device (450) can receive a chest CT image from an external object.

[0092] The communication device (450) can receive a deep learning model constructed from an external object.

[0093] The communication device (450) may also receive learning data for building a deep learning model from an external object.

[0094] The communication device (450) can also transmit a virtual contrast-enhanced chest CT image generated using a deep learning model to an external object such as a user terminal.

[0095] The output device (460) is a device that outputs certain information. The output device (460) can output interfaces required for the data processing process, input chest CT images, and generated virtual contrast-enhanced chest CT images.

[0096] The computing device (430) can build the aforementioned deep learning model using training data. The learning process is as described in FIG. 2 or FIG. 4 . In this case, the training data uses dual-energy CT, and the computing device (430) can build the deep learning model without a matching process.

[0097] The computing device (430) can input a pre-contrast chest CT image into the constructed deep learning model to generate a virtual contrast-enhanced chest CT image.

[0098] The computing device (430) may be a device such as a processor, AP, or chip embedded with a program that processes data and performs certain operations.

[0099] Although Figure 6 illustrates a chest CT image, the data processing device (400) can generate virtual contrast-enhanced CT images for CT images of various body parts. In this case, a deep learning model can be constructed using learning data for a specific disease or body part during the learning process.

[0100] Additionally, the method for generating contrast-enhanced CT images or the method for constructing a deep learning model for generating contrast-enhanced CT images as described above may be implemented as a program (or application) including an executable algorithm that can be executed on a computer. The program may be stored and provided on a temporary or non-transitory computer-readable medium.

[0101] A non-transitory readable medium refers to a medium that stores data semi-permanently and can be read by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, the various applications or programs described above may be stored and provided on a non-transitory readable medium, such as a CD, DVD, hard disk, Blu-ray disk, USB, memory card, ROM (read-only memory), PROM (programmable read only memory), EPROM (Erasable PROM, EPROM), EEPROM (Electrically EPROM), or flash memory.

[0102] Temporarily readable media refers to various types of RAM, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchronous DRAM (Synclink DRAM, SLDRAM), and Direct Rambus RAM (DRRAM).

[0103] The present embodiment and the drawings attached to the present specification only clearly illustrate a part of the technical idea included in the above-described technology, and it will be obvious that all modified examples and specific embodiments that can be easily inferred by a person skilled in the art within the scope of the technical idea included in the specification and drawings of the above-described technology are included in the scope of the rights of the above-described technology.

Claims

1. A step in which a data processing device receives a pre-contrast CT (Computed Tomography) image and an actual contrast CT image for the pre-contrast CT image; A step in which the data processing device inputs the pre-contrast CT image into a deep learning model to generate a virtual contrast-enhanced CT image; and The data processing device includes a step of updating the parameters of the deep learning model using the results of comparing the actual contrast CT image and the virtual contrast-enhanced CT image, A method for constructing a deep learning model for generating a virtual contrast-enhanced CT image, wherein the deep learning model includes a convolutional block attention module that generates enhanced features using a channel attention map for input values ​​and a spatial attention map for input values.

2. In paragraph 1, A method for constructing a deep learning model for generating a virtual contrast-enhanced CT image, which is a dual-energy CT image, from the above pre-contrast CT image and the above actual contrast-enhanced CT image.

3. In paragraph 1, The above deep learning model is an adversarial generative model, The generator of the above adversarial generative model includes a number of convolution blocks, A method for constructing a deep learning model for generating a virtual contrast-enhanced CT image, wherein the convolution block includes at least one convolution layer and the convolution block attention module.

4. In paragraph 1, The above deep learning model is a residual U-Net-based model, The above residual U-Net based model is a method for constructing a deep learning model that generates a virtual contrast-enhanced CT image including the convolution block attention module as a residual block.

5. In paragraph 1, A method for constructing a deep learning model for generating a virtual contrast-enhanced CT image, wherein the convolution block attention module includes a channel attention module that generates the channel attention map by performing channel attention on an input feature map and a spatial attention module that generates the spatial attention map by performing spatial attention on an intermediate feature map generated by multiplying the input feature map by the channel attention map, and generates an enhanced feature map by multiplying the intermediate feature map by the spatial attention map.

6. An interface device that receives pre-contrast CT (Computed Tomography) images of a specific subject; A storage device for storing a deep learning model that generates contrast-enhanced CT images; and Including a computing device that inputs the above pre-contrast CT image into the deep learning model to generate a virtual contrast-enhanced CT image, A data processing device for generating a virtual contrast-enhanced CT image, wherein the deep learning model includes a convolutional block attention module that generates enhanced features using a channel attention map for input values ​​and a spatial attention map for input values.

7. In paragraph 6, The above deep learning model is an adversarial generative model, The generator of the above adversarial generative model includes a number of convolution blocks, A data processing device for generating a virtual contrast-enhanced CT image, wherein the convolution block comprises at least one convolution layer and the convolution block attention module.

8. In paragraph 6, The above deep learning model is a residual U-Net-based model, The residual U-Net-based model is a data processing device that generates a virtual contrast-enhanced CT image including the convolution block attention module as a residual block.

9. In paragraph 6, A data processing device for generating a virtual contrast-enhanced CT image, wherein the convolution block attention module includes a channel attention module that generates the channel attention map by performing channel attention on an input feature map and a spatial attention module that generates the spatial attention map by performing spatial attention on an intermediate feature map generated by multiplying the input feature map by the channel attention map, and generates an enhanced feature map by multiplying the intermediate feature map by the spatial attention map.

Images