A medical image reconstruction method based on an adversarial sample and a related device

Abstract

The application discloses a medical image reconstruction method based on an adversarial sample and a related device. The method comprises the following steps: obtaining a medical image to be reconstructed, and inputting the medical image into a trained recovery network model, wherein the recovery network model is trained by using training medical images and corresponding deinterference images of the training medical images, and the deinterference images are generated based on adversarial samples generated from the training medical images; and determining a reconstructed image corresponding to the medical image by using the recovery network model. The application adopts the method of generating deinterference images and training images based on adversarial samples corresponding to training medical images to train a recovery network model to reconstruct medical images, so that the image quality of the reconstructed medical image is improved. This is because the use of deinterference images as clean images can solve the problem of the decline of the recovery performance of an adversarial training method on clean samples, thereby improving the model performance of the recovery network model.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a medical image reconstruction method and related apparatus based on adversarial examples. Background Technology

[0002] With the development of artificial intelligence, deep learning algorithms have demonstrated astonishingly high performance in many AI systems, such as image recognition and semantic segmentation. Deep learning is increasingly being applied in the medical field. In medical diagnosis and decision support, machine learning systems appear to have achieved diagnostic capabilities comparable to those in radiology, pathology, and dermatology. However, existing machine learning systems in the medical field generally suffer from the problem of being easily fooled by adversarial examples, resulting in erroneous outputs.

[0003] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0004] The technical problem to be solved by this application is to provide a medical image reconstruction method and related apparatus based on adversarial examples, addressing the shortcomings of existing technologies.

[0005] To address the aforementioned technical problems, the first aspect of this application provides a medical image reconstruction method based on adversarial examples, the method comprising:

[0006] A medical image to be reconstructed is acquired and input into a trained recovery network model, wherein the recovery network model is trained using the training medical image and the corresponding de-interference image, and the de-interference image is generated based on adversarial examples generated from the training medical image;

[0007] The reconstructed image corresponding to the medical image is determined using the recovery network model.

[0008] The adversarial example-based medical image reconstruction method includes a recovery network model comprising several cascaded recovery modules. Each recovery module includes a convolutional unit, a 3D attention unit, and a DC unit. The input of the DC unit includes the output of the 3D attention unit and the input of the recovery network model.

[0009] The adversarial example-based medical image reconstruction method, wherein the training process of the recovery network model specifically includes:

[0010] Adversarial example images are generated based on trained medical images and a pre-trained perturbation generation model.

[0011] The adversarial example image is input into a pre-trained deperturbation model, and the deperturbation image is determined by the deperturbation model.

[0012] The training medical image and the de-perturbed image are respectively input into a preset network model. The preset network model outputs a first restored image corresponding to the training medical image and a second restored image corresponding to the de-perturbed image. The preset network model is then trained based on the first restored image and the second restored image to obtain a restoration network model.

[0013] The adversarial example-based medical image reconstruction method, wherein generating adversarial example images based on trained medical images and a pre-trained perturbation generation model specifically includes:

[0014] Obtain image noise corresponding to the training medical image, and input the training medical image and the image operation into a pre-trained perturbation generation model to generate candidate perturbations through the perturbation generation model;

[0015] After applying norm constraints to the candidate perturbations, adversarial perturbations are generated, and these perversarial perturbations are added to the training medical images to obtain adversarial sample images.

[0016] The medical image reconstruction method based on adversarial examples includes a perturbation model comprising several sequentially arranged feature extraction modules and a first adder. The input of the first adder is the output of the last feature extraction module and the input of the first feature extraction module. A second adder is arranged every two feature extraction modules in the sequentially arranged module array. The input of the second adder is the output of the preceding feature extraction module and the input of the feature extraction module following the preceding adder.

[0017] The medical image reconstruction method based on adversarial examples includes a feature extraction module comprising a cascaded dense residual attention unit, a convolutional layer, and an activation function layer. The dense residual attention unit comprises a convolutional block, an attention layer, and a first adder. The input to the attention layer is the output of the convolutional block. The input to the first adder is the input of the convolutional block and the output of the attention layer. The convolutional block comprises two cascaded sub-units, a first sub-unit and a second sub-unit. Both the first and second sub-units comprise a cascaded first convolutional layer, a normalization layer, an activation function layer, a second convolutional layer, a second adder, and a connection layer. The input to the second adder is the input of the first convolutional layer and the output of the second convolutional layer. The input to the connection layer in the first sub-unit is the input of the first convolutional layer and the output of the adder. The input to the connection layer in the second sub-unit is the input of the first convolutional layer in the first sub-unit, the input of the first convolutional layer in the second sub-unit, and the output of the adder in the second sub-unit.

[0018] The medical image reconstruction method based on adversarial examples, wherein training the preset network model based on the first restored image and the second restored image to obtain the restored network model specifically includes:

[0019] A first loss term is determined based on the first restored image and the second restored image;

[0020] A second loss term is determined based on the first restored image and the training medical image;

[0021] The preset network model is trained based on the first loss term and the second loss term to obtain the recovery network model.

[0022] A second aspect of this application provides a medical image reconstruction system based on adversarial examples, characterized in that the system comprises:

[0023] An input module is used to acquire the medical image to be reconstructed and input the medical image into a trained recovery network model, wherein the recovery network model is trained using the training medical image and the corresponding de-interference image, and the de-interference image is generated based on adversarial examples generated from the training medical image;

[0024] The determination module is used to determine the reconstructed image corresponding to the medical image through the recovery network model.

[0025] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the medical image reconstruction method based on adversarial examples as described above.

[0026] A fourth aspect of this application provides a terminal device, which includes: a processor, a memory, and a communication bus; the memory stores a computer-readable program that can be executed by the processor;

[0027] The communication bus enables communication between the processor and the memory;

[0028] When the processor executes the computer-readable program, it implements the steps in any of the above-described adversarial example-based medical image reconstruction methods.

[0029] Beneficial Effects: Compared with existing technologies, this application provides a medical image reconstruction method and related apparatus based on adversarial examples. The method includes: acquiring a medical image to be reconstructed, and inputting the medical image into a trained recovery network model, wherein the recovery network model is trained using a training medical image and a corresponding de-interference image, the de-interference image being generated based on adversarial examples generated from the training medical image; and determining the reconstructed image corresponding to the medical image through the recovery network model. This application uses a de-interference image generated based on adversarial examples corresponding to the training medical image and trains the recovery network model using the training image to reconstruct the medical image, which can improve the image quality of the reconstructed medical image. This is because using the de-interference image as a clean image can solve the problem of decreased recovery performance of adversarial training methods for clean samples, improve the model performance of the recovery network model, and thus improve the image quality of the reconstructed medical image. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart of the medical image reconstruction method based on adversarial examples provided in this application.

[0032] Figure 2 The structural principle diagram of the recovery network model in the adversarial example-based medical image reconstruction method provided in this application.

[0033] Figure 3 A flowchart illustrating the principle of the training process of the recovery network model in the adversarial example-based medical image reconstruction method provided in this application.

[0034] Figure 4 The structural principle diagram of the interference removal model in the adversarial example-based medical image reconstruction method provided in this application.

[0035] Figure 5 This is a schematic diagram of the feature extraction module in the interference removal model.

[0036] Figure 6 The structural principle diagram of the medical image reconstruction system based on adversarial examples provided in this application.

[0037] Figure 7 A schematic diagram of the terminal device provided in this application. Detailed Implementation

[0038] This application provides a medical image reconstruction method and related apparatus based on adversarial examples. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0039] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0040] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0041] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0042] The inventors discovered through research that, with the development of artificial intelligence, deep learning algorithms have demonstrated astonishingly high performance in many AI systems, such as image recognition and semantic segmentation. The application of deep learning in the medical field is becoming increasingly widespread. In medical diagnosis and decision support, machine learning systems appear to have achieved diagnostic capabilities comparable to those in radiology, pathology, and dermatology. For example, the U.S. Food and Drug Administration (FDA) approved the commercialization of the first autonomous artificial intelligence (AI) diagnostic system and stated that they are "actively developing new regulatory frameworks to foster innovation in this field." However, existing machine learning systems in the medical field generally suffer from the problem of being easily fooled by adversarial examples, resulting in erroneous outputs.

[0043] To address the aforementioned issues, adversarial training, as an effective defense mechanism, has been widely applied in image classification. However, existing adversarial training methods generally employ fixed perturbations, making them difficult to adapt to attacks in the medical field. Furthermore, they exhibit poor recovery performance for clean samples, impacting the model's robustness.

[0044] Based on this, this application provides a medical image reconstruction method based on adversarial examples. The method includes acquiring a medical image to be reconstructed and inputting the medical image into a trained recovery network model. The recovery network model is trained using the training medical image and a corresponding de-interference image, where the de-interference image is generated based on adversarial examples generated from the training medical image. The reconstructed image corresponding to the medical image is determined through the recovery network model. This application uses a de-interference image generated from adversarial examples corresponding to the training medical image and the training image to train the recovery network model for medical image reconstruction. This can improve the image quality of the reconstructed medical image. This is because using the de-interference image as a clean image can solve the problem of decreased recovery performance of adversarial training methods on clean samples, improving the model performance of the recovery network model and thus improving the image quality of the reconstructed medical image.

[0045] The application content will be further explained below with reference to the accompanying drawings and the description of the embodiments.

[0046] This embodiment provides a medical image reconstruction method based on adversarial examples, such as... Figure 1 As shown, the method includes:

[0047] S10. Obtain the medical image to be reconstructed and input the medical image into the trained recovery network model.

[0048] Specifically, medical images are images used in the medical field, such as ultrasound images and X-ray images. The recovery network model is a deep learning-based network model, and it is pre-trained. The recovery network model is trained using medical images and corresponding de-interference images. The de-interference images are generated based on adversarial examples generated from the training medical images, where the adversarial examples are formed by adding perturbations to the training images.

[0049] The restoration network model includes several cascaded restoration modules. Each restoration module includes a convolutional unit, a 3D attention unit, and a DC unit. The input of the DC unit includes the output of the 3D attention unit and the input of the restoration network model. The convolutional unit may include several cascaded convolutional layers, the 3D attention unit may include a 3D attention layer, and the DC unit includes a data consistency layer. Each restoration module shares attention weights. This embodiment uses convolutional units and 3D attention units to form a denoising module, and then iteratively expands the module by alternating between the denoising module and the DC unit. This reduces the number of training samples required by the restoration network model, thereby improving the training speed. Simultaneously, by using a 3D attention layer, the restoration network model's attention to image features is improved, thus enhancing its performance. In one implementation, such as... Figure 2 As shown, the recovery network model includes a first recovery module, a second recovery module, and a third recovery module. The input items of the DC unit in the first recovery module are the input items of the first recovery module and the output items of the 3D attention unit in the first recovery module. The input items of the DC unit in the second recovery module are the input items of the first recovery module and the output items of the 3D attention unit in the second recovery module. The input items of the DC unit in the third recovery module are the input items of the first recovery module and the output items of the 3D attention unit in the third recovery module.

[0050] In one implementation, such as Figure 3 As shown, the training process of the recovery network model specifically includes:

[0051] H10. Generate adversarial example images based on trained medical images and a pre-trained perturbation generation model;

[0052] H20. Input the adversarial example image into a pre-trained deperturbation model, and determine the deperturbation image through the deperturbation model;

[0053] H30. Input the training medical image and the de-perturbated image into the preset network model respectively. Output the first restored image corresponding to the training medical image and the second restored image corresponding to the de-perturbated image through the preset network model. Train the preset network model based on the first restored image and the second restored image to obtain the restoration network model.

[0054] Specifically, in step H10, the adversarial example image is generated by a perturbation generation model, and the adversarial example image is a training medical image carrying perturbations. That is, the perturbation example image is formed by adding perturbations to the training medical image, wherein the perturbation is an adaptive perturbation generated by the perturbation generation model based on the training medical image and the noise corresponding to the training medical image. This embodiment uses adaptive perturbations generated based on the training medical image and the noise corresponding to the training medical image to form the adversarial example image, making the adversarial example image more consistent with the medical scenario corresponding to the training medical image. This improves the model performance of the subsequent recovery network model trained based on the de-interference image generated from the adversarial training sample, and consequently improves the image quality of the reconstructed image determined by the recovery network model.

[0055] In one implementation, generating adversarial example images based on trained medical images and a pre-trained perturbation generation model specifically includes:

[0056] Obtain image noise corresponding to the training medical image, and input the training medical image and the image operation into a pre-trained perturbation generation model to generate candidate perturbations through the perturbation generation model;

[0057] After applying norm constraints to the candidate perturbations, adversarial perturbations are generated, and these perversarial perturbations are added to the training medical images to obtain adversarial sample images.

[0058] Specifically, the image noise is noise associated with the training medical images. For example, it could be fixed noise based on the medical domain in which the training medical images are located, or white noise formed by methods such as Gaussian filtering, or image noise extracted from the training medical images by a pre-built network model. In one implementation, the image noise is formed using a normal distribution.

[0059] Furthermore, the perturbation generation model pre-trained using the training medical image and the image manipulation input refers to using the joint distribution of the training medical image and image noise as input to the perturbation generation model. The perturbation generation model learns an adaptive perturbation by studying the joint distribution. Therefore, the generation process of the candidate perturbation can be expressed as: ;in, Indicates a generator network. The parameters represent the network training parameters. Represent y and The joint distribution; This represents image noise. This embodiment uses the joint distribution of training medical images and image noise as input, learning perturbations from this joint distribution, which better aligns with real-world application scenarios and improves the performance of the subsequently trained recovery network model. In practical applications, the perturbation generation model can be determined according to actual needs; for example, U-Net, ResNet, or other network models can be used, which will not be elaborated upon here.

[0060] In step H20, the deperturbation model is pre-trained and is used to remove interference from the adversarial example image to obtain a deperturbated image. It can be understood that the input to the deperturbation model is the adversarial example image, and the output is the deperturbated image. This embodiment uses the deperturbated image output by the deperturbation model as the clean image, which reduces the gap between the restored image corresponding to the clean image and the restored image corresponding to the adversarial example, improving the model's restoration performance for clean samples, thereby improving the image quality of the reconstructed image determined based on the restoration network model.

[0061] In one implementation, the deperturbation model includes several sequentially arranged feature extraction modules and a first adder. The input to the first adder is the output of the last feature extraction module and the input of the first feature extraction module. A second adder is placed every two feature extraction modules in the sequential arrangement. The input to the second adder is the output of the preceding feature extraction module and the input of the feature extraction module following the preceding adder. The deperturbation model achieves a dense residual connection method by using adders, enabling the model to extract effective features and thus reducing the cleanliness of the deperturbated image obtained by the model. For example, as... Figure 4 As shown, the perturbation removal model includes six feature extraction modules. A first adder is positioned between the second and third feature extraction modules, between the fourth and fifth feature extraction modules, and after the sixth feature extraction module, a first adder is positioned, followed by a second adder. It is worth noting that in practical applications, the number of feature extraction modules included in the perturbation removal model is preferably even.

[0062] Furthermore, such as Figure 4 and Figure 5As shown, the feature extraction module DRAM includes a dense residual attention unit, a convolutional layer, and an activation function layer cascaded in sequence. The dense residual attention unit includes a convolutional block, an attention layer, and a first adder. The input of the attention layer is the output of the convolutional block. The input of the first adder is the input of the convolutional block and the output of the attention layer. The convolutional block includes two cascaded first sub-units and second sub-units. Both the first and second sub-units include a first convolutional layer (Conv), a normalization layer (BN), an activation function layer (PReLU), a second convolutional layer (Conv), a second adder, and a connection layer (Concatenation) cascaded in sequence. The input of the second adder is the input of the first convolutional layer and the output of the second convolutional layer. The input of the connection layer in the first sub-unit is the input of the first convolutional layer and the output of the adder. The input of the connection layer in the second sub-unit is the input of the first convolutional layer in the first sub-unit, the input of the first convolutional layer in the second sub-unit, and the output of the adder in the second sub-unit. In this embodiment, the feature extraction module is equipped with dense residual attention units. By learning key features through dense residual attention units, the learning ability of the perturbation model for key features can be improved, thereby further improving the model performance of the perturbation model.

[0063] In one implementation, the deperturbation model is constrained by a high-dimensional representation loss during training. This improves the cleanliness of the deperturbated image output by the trained model. The loss term during training can be expressed as:

[0064]

[0065] in, Indicates adversarial examples, Indicates the removal of interference from the image. This represents the perturbation-free model. This represents a high-dimensional feature projection network.

[0066] In step H30, training the preset network model based on the first restored image and the second restored image to obtain the restored network model specifically includes:

[0067] A first loss term is determined based on the first restored image and the second restored image;

[0068] A second loss term is determined based on the first restored image and the training medical image;

[0069] The preset network model is trained based on the first loss term and the second loss term to obtain the recovery network model.

[0070] This embodiment trains the preset network model using a first loss term and a second loss term, achieving high-quality sample restoration without the need for labeled data, thus reducing dependence on labeled data. Specifically, the loss terms for the first and second restored images reflect the consistency loss of the restoration network model, while the second loss term reflects the data fidelity loss of the training medical images. By constraining the restoration network model with consistency and fidelity losses, the performance of the trained model can be improved, thereby enhancing the image quality of the reconstructed images determined by the restoration network model.

[0071] S20. Determine the reconstructed image corresponding to the medical image through the recovery network model.

[0072] Specifically, the reconstructed image is the reconstructed image corresponding to the medical image. The image quality of the reconstructed image is higher than that of the medical image, and the image content carried by the reconstructed image is the same as that of the medical image.

[0073] In summary, this embodiment provides a medical image reconstruction method based on adversarial examples. The method includes: acquiring a medical image to be reconstructed, and inputting the medical image into a trained recovery network model, wherein the recovery network model is trained using a training medical image and a corresponding de-interference image, and the de-interference image is generated based on adversarial examples generated from the training medical image; and determining the reconstructed image corresponding to the medical image through the recovery network model. This application uses a de-interference image generated based on adversarial examples corresponding to the training medical image and trains the recovery network model using the training image to reconstruct the medical image, which can improve the image quality of the reconstructed medical image. This is because using the de-interference image as a clean image can solve the problem of decreased recovery performance of adversarial training methods on clean samples, improve the model performance of the recovery network model, and thus improve the image quality of the reconstructed medical image.

[0074] Based on the aforementioned adversarial example-based medical image reconstruction method, this embodiment provides an adversarial example-based medical image reconstruction system, such as... Figure 6 As shown, the system includes:

[0075] The input module 100 is used to acquire the medical image to be reconstructed and input the medical image into a trained recovery network model, wherein the recovery network model is trained using the training medical image and the corresponding de-interference image, and the de-interference image is generated based on adversarial examples generated from the training medical image.

[0076] The determination module 200 is used to determine the reconstructed image corresponding to the medical image through the recovery network model.

[0077] Based on the above-described adversarial example-based medical image reconstruction method, this embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the adversarial example-based medical image reconstruction method described in the above embodiment.

[0078] Based on the aforementioned adversarial example-based medical image reconstruction method, this application also provides a terminal device, such as... Figure 7 As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.

[0079] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0080] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.

[0081] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as transient storage media.

[0082] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions 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 this application.

Claims

1. A medical image reconstruction method based on adversarial examples, characterized in that, The method includes: A medical image to be reconstructed is acquired and input into a trained recovery network model, wherein the recovery network model is trained using the training medical image and the corresponding de-interference image, and the de-interference image is generated based on adversarial examples generated from the training medical image; The reconstructed image corresponding to the medical image is determined using the recovery network model; The training process of the recovery network model specifically includes: Adversarial example images are generated based on trained medical images and a pre-trained perturbation generation model. The adversarial example image is input into a pre-trained deperturbation model, and the deperturbation image is determined by the deperturbation model. The training medical image and the de-perturbed image are respectively input into a preset network model. The preset network model outputs a first restored image corresponding to the training medical image and a second restored image corresponding to the de-perturbed image. The preset network model is then trained based on the first restored image and the second restored image to obtain a restoration network model.

2. The medical image reconstruction method based on adversarial examples according to claim 1, characterized in that, The recovery network model includes several cascaded recovery modules. Each recovery module includes a convolutional unit, a 3D attention unit, and a DC unit. The input of the DC unit includes the output of the 3D attention unit and the input of the recovery network model.

3. The medical image reconstruction method based on adversarial examples according to claim 1, characterized in that, The generation of adversarial example images based on trained medical images and a pre-trained perturbation generation model specifically includes: Obtain image noise corresponding to the training medical image, and input the training medical image and the image noise into a pre-trained perturbation generation model to generate candidate perturbations. After applying norm constraints to the candidate perturbations, adversarial perturbations are generated, and these perversarial perturbations are added to the training medical images to obtain adversarial sample images.

4. The medical image reconstruction method based on adversarial examples according to claim 1, characterized in that, The perturbation removal model includes several feature extraction modules arranged in sequence and a first adder. The input of the first adder is the output of the last feature extraction module and the input of the first feature extraction module. A second adder is arranged every two feature extraction modules in the several sequentially arranged feature extraction modules. The input of the second adder is the output of the preceding feature extraction module and the input of the feature extraction module following the preceding adder.

5. The medical image reconstruction method based on adversarial examples according to claim 4, characterized in that, The feature extraction module includes a dense residual attention unit, a convolutional layer, and an activation function layer cascaded sequentially. The dense residual attention unit includes a convolutional block, an attention layer, and a first adder. The input of the attention layer is the output of the convolutional block. The input of the first adder is the input of the convolutional block and the output of the attention layer. The convolutional block includes two cascaded first sub-units and second sub-units. Both the first and second sub-units include a first convolutional layer, a normalization layer, an activation function layer, a second convolutional layer, a second adder, and a connection layer cascaded sequentially. The input of the second adder is the input of the first convolutional layer and the output of the second convolutional layer. The input of the connection layer in the first sub-unit is the input of the first convolutional layer and the output of the adder. The input of the connection layer in the second sub-unit is the input of the first convolutional layer in the first sub-unit, the input of the first convolutional layer in the second sub-unit, and the output of the adder in the second sub-unit.

6. The medical image reconstruction method based on adversarial examples according to claim 1, characterized in that, The step of training the preset network model based on the first restored image and the second restored image to obtain the restored network model specifically includes: A first loss term is determined based on the first restored image and the second restored image; A second loss term is determined based on the first restored image and the training medical image; The preset network model is trained based on the first loss term and the second loss term to obtain the recovery network model.

7. A medical image reconstruction system based on adversarial examples, characterized in that, The system includes: An input module is used to acquire the medical image to be reconstructed and input the medical image into a trained recovery network model, wherein the recovery network model is trained using the training medical image and the corresponding de-interference image, and the de-interference image is generated based on adversarial examples generated from the training medical image; The determination module is used to determine the reconstructed image corresponding to the medical image through the recovery network model; The training process of the recovery network model specifically includes: Adversarial example images are generated based on trained medical images and a pre-trained perturbation generation model. The adversarial example image is input into a pre-trained deperturbation model, and the deperturbation image is determined by the deperturbation model. The training medical image and the de-perturbed image are respectively input into a preset network model. The preset network model outputs a first restored image corresponding to the training medical image and a second restored image corresponding to the de-perturbed image. The preset network model is then trained based on the first restored image and the second restored image to obtain a restoration network model.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the medical image reconstruction method based on adversarial examples as described in any one of claims 1-6.

9. A terminal device, characterized in that, include: Processor, memory, and communication bus; the memory stores a computer-readable program that can be executed by the processor; The communication bus enables communication between the processor and the memory; When the processor executes the computer-readable program, it implements the steps in the medical image reconstruction method based on adversarial examples as described in any one of claims 1-6.

Images