WO2026016159A1 - 一种基于生成对抗网络模型的fMRI生成方法及系统 - Google Patents

一种基于生成对抗网络模型的fMRI生成方法及系统

Info

Publication number
WO2026016159A1
WO2026016159A1 PCT/CN2024/106346 CN2024106346W WO2026016159A1 WO 2026016159 A1 WO2026016159 A1 WO 2026016159A1 CN 2024106346 W CN2024106346 W CN 2024106346W WO 2026016159 A1 WO2026016159 A1 WO 2026016159A1
Authority
WO
WIPO (PCT)
Prior art keywords
fmri
mri
generator
discriminator
generative adversarial
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2024/106346
Other languages
English (en)
French (fr)
Inventor
谢桃枫
王海峰
周谊航
梁栋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Institute of Advanced Technology of CAS
Original Assignee
Shenzhen Institute of Advanced Technology of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen Institute of Advanced Technology of CAS filed Critical Shenzhen Institute of Advanced Technology of CAS
Publication of WO2026016159A1 publication Critical patent/WO2026016159A1/zh
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T12/00—Tomographic reconstruction from projections
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0475—Generative networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning

Definitions

  • This invention relates to the field of magnetic resonance imaging technology, and in particular to an fMRI generation method, system, terminal, and computer-readable storage medium based on a generative adversarial network model.
  • Magnetic resonance imaging (MRI) provides high-resolution anatomical images, while functional magnetic resonance imaging (fMRI) reveals the brain's functional activity.
  • fMRI data is often difficult to obtain due to time and resource limitations associated with MRI acquisition. This results in a situation where only MRI or fMRI data may be available for the same patient, and very few patients possess both types of data simultaneously.
  • fMRI acquisition is costly, time-consuming, and requires specific patient expertise, making it generally not feasible to perform this procedure on every patient.
  • the main objective of this invention is to provide an fMRI generation method, system, terminal, and computer-readable storage medium based on a generative adversarial network model, aiming to solve the problems of high cost, long time, and high requirements for the acquisition objects in the prior art for acquiring fMRI images.
  • this invention provides an fMRI generation method based on a generative adversarial network (GAN) model, which includes the following steps:
  • the generator and discriminator in the generative adversarial network model are trained using the training dataset, and the network parameters of the generator and the discriminator are optimized to obtain a trained generative adversarial network model.
  • the MRI image of the target object is acquired, and then the MRI image of the target object is input into the trained generative adversarial network model to directly generate the fMRI image of the target object.
  • the collection of multiple target ...
  • the acquired subjects' paired MRI and fMRI data were used to build a training dataset, specifically including:
  • paired MRI and fMRI data from multiple target acquisition objects, wherein the paired MRI and fMRI data includes MRI images and real fMRI images;
  • the paired MRI and fMRI data are preprocessed, including denoising, standardization, and registration.
  • a training dataset was created based on preprocessed paired MRI and fMRI data.
  • the fMRI generation method based on a generative adversarial network model includes a generator and a discriminator.
  • the generator is used to generate fMRI images based on MRI images
  • the discriminator is used to determine the difference between the fMRI image generated by the generator and the real fMRI image
  • the generator uses the ResNet architecture to capture multi-scale features
  • the discriminator uses a CNN architecture to improve the fine-grainedness of discrimination.
  • the fMRI generation method based on a generative adversarial network (GAN) model, wherein training the generator and discriminator in the GAN model using the training dataset, and optimizing the network parameters of the generator and the discriminator to obtain a trained GAN model, specifically includes:
  • MRI images and random Gaussian noise are input into the network structure of the generator for training to generate fMRI images;
  • the fMRI image generated by the generator is input into the network structure of the discriminator along with the real fMRI image;
  • the discriminator distinguishes between generated fMRI images and real fMRI images, and outputs a true/false classification result.
  • backpropagation is used to optimize the network parameters of the generator and the discriminator, so that the fMRI image generated by the generator is as close as possible to the real fMRI image, and the discriminator can distinguish between the real fMRI image and the generated fMRI image as accurately as possible.
  • the fMRI generation method based on a generative adversarial network model further includes:
  • the generator and the discriminator are optimized by using the loss function of the generative adversarial network model, so that the fMRI images generated by the generator are as similar as possible to the real fMRI data, while at the same time enhancing the discrimination ability of the discriminator.
  • the fMRI generation method based on a generative adversarial network model includes an adversarial loss and a generation loss in the loss function of the generative adversarial network model.
  • the adversarial loss is used to optimize the adversarial learning between the generator and the discriminator, so that the fMRI images generated by the generator... Gradually approximating the distribution of real data;
  • the generation loss is used to reduce the pixel difference between the generated fMRI image and the real fMRI image, thereby improving the quality of the generated image.
  • the loss function of the GAN model is:
  • G represents the generator
  • D represents the discriminator
  • x represents the MRI image
  • y represents the real fMRI image
  • z represents random Gaussian noise
  • D(x,y) represents the true/false classification result obtained by the discriminator between the MRI image and the real fMRI image.
  • G(x,z) represent the expectation of logD(x,y)
  • G(x,z) represent the result obtained by the generator after the MRI image and random Gaussian noise are fed into it
  • D(x,G(x,z)) represent the result obtained by the discriminator after the MRI image and the fMRI image generated by the generator are passed through it. This represents the expectation of log(1-D(x,G(x,z))).
  • the present invention also provides an fMRI generation system based on a generative adversarial network (GAN) model, wherein the fMRI generation system based on the GAN model comprises:
  • the data acquisition module is used to collect paired MRI and fMRI data from multiple target acquisition objects and to build a training dataset based on the paired MRI and fMRI data.
  • the model training module is used to train the generator and discriminator in the generative adversarial network model using the training dataset, optimize the network parameters of the generator and the discriminator, and obtain a trained generative adversarial network model.
  • the image generation module is used to acquire MRI images of the target object, input the MRI images of the target object into the trained generative adversarial network model, and directly generate fMRI images of the target object.
  • the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an fMRI generation program based on a generative adversarial network model stored in the memory and executable on the processor, wherein when the fMRI generation program based on the generative adversarial network model is executed by the processor, it implements the steps of the fMRI generation method based on the generative adversarial network model as described above.
  • the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an fMRI generation program based on a generative adversarial network model, and the fMRI generation program based on the generative adversarial network model, when executed by a processor, implements the steps of the fMRI generation method based on the generative adversarial network model as described above.
  • paired MRI and fMRI data of multiple target subjects are collected, and a training dataset is established based on the paired MRI and fMRI data.
  • the generator and discriminator in a generative adversarial network (GAN) model are trained using the training dataset, and the network parameters of the generator and discriminator are optimized to obtain a trained GAN model.
  • MRI images of the target subjects are acquired and input into the trained GAN model to directly generate fMRI images of the target subjects.
  • This invention utilizes the MRI data of the acquired subjects to generate corresponding fMRI data.
  • the MRI data of each acquired subject can be converted into fMRI data, thus providing each acquired subject with both MRI and fMRI data. This not only provides accurate and reliable information for downstream tasks but also maximizes the use of existing single-modal data, avoiding data waste, reducing reliance on fMRI data acquisition, and lowering the complexity and cost of clinical procedures.
  • Figure 1 is a flowchart of a preferred embodiment of the fMRI generation method based on a generative adversarial network model of the present invention
  • Figure 2 is a schematic diagram of the training of the generative adversarial network model in a preferred embodiment of the fMRI generation method based on the generative adversarial network model of the present invention.
  • Figure 3 is a schematic diagram of a preferred embodiment of the fMRI generation system based on a generative adversarial network model of the present invention
  • Figure 4 is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention.
  • GAN Geneative Adversarial Networks
  • SRGAN Geneative Adversarial Networks
  • SRGAN Geneative Adversarial Networks
  • CycleGAN is a generative adversarial network that does not require paired training data and is widely used for the conversion between different modalities of medical images.
  • CycleGAN can perform modality transfer without paired images, but the quality and consistency of the generated images still need improvement.
  • the pix2pix model has been used for various image-to-image conversion tasks, including CT to MRI conversion. High-quality image translation is achieved through a conditional generative adversarial network, relying on paired training data.
  • the preferred embodiment of the fMRI generation method based on a generative adversarial network model of the present invention includes the following steps:
  • Step S10 Collect paired MRI and fMRI data of multiple target acquisition objects, and establish a training dataset based on the paired MRI and fMRI data.
  • paired MRI and fMRI data are collected from multiple target subjects (e.g., Parkinson's disease patients).
  • This paired MRI and fMRI data includes MRI images and actual fMRI images (i.e., MRI data and fMRI data).
  • the paired MRI and fMRI data undergo preprocessing, including denoising, normalization, and registration operations, to ensure image quality and alignment.
  • 3Dslicer software is used to register the MRI and fMRI data to ensure the accuracy and consistency of the training data.
  • the registration process involves aligning the MRI and fMRI images to ensure they have the same spatial resolution and geometry.
  • a training dataset is then built based on the preprocessed paired MRI and fMRI data.
  • Step S20 Use the training dataset to train the generator and discriminator in the generative adversarial network model, optimize the network parameters of the generator and the discriminator, and obtain the trained generative adversarial network model.
  • the generative adversarial network model includes a generator G and a discriminator D.
  • the generator G is used to generate fMRI images based on MRI images.
  • the discriminator D is used to distinguish between the fMRI images generated by the generator and real fMRI images.
  • the generator G uses the ResNet architecture (residual network), which is known for its powerful performance in image generation tasks. ResNet effectively solves the gradient vanishing problem in deep networks through residual blocks, improving the quality of generated images.
  • the generator G mainly learns the mapping between MRI and fMRI to capture multi-scale features.
  • the discriminator D uses the CNN architecture (convolutional neural network), which is known for its excellent performance in image classification tasks. CNN can extract high-level features of images and accurately distinguish between generated fMRI and real fMRI, thereby improving the fine-grainedness of discrimination.
  • the training process specifically includes:
  • MRI images and random Gaussian noise are input into the network structure of the generator G for training to generate fMRI images.
  • the fMRI images generated by the generator G are then input together with real fMRI images into the network structure of the discriminator D.
  • the discriminator D distinguishes between the generated fMRI images and the real fMRI images, outputting a true/false classification result.
  • backpropagation is used to optimize the network parameters of the generator G and the discriminator D, so that the fMRI images generated by the generator G approximate the real fMRI images as closely as possible, and the discriminator D distinguishes between the real fMRI images and the generated fMRI images as accurately as possible.
  • the entire training process is adversarial training.
  • the goal of the generator G is to minimize the difference between the generated fMRI image and the real fMRI image, making the generated fMRI image as close to the real fMRI image as possible.
  • the goal of the discriminator is to maximize the accuracy of distinguishing between the generated fMRI image and the real fMRI image.
  • the generator and the discriminator are optimized using the loss function of the generative adversarial network model, so that the fMRI images generated by the generator are as similar as possible to the real fMRI data, while simultaneously enhancing the discrimination ability of the discriminator.
  • the loss function of the generative adversarial network model includes adversarial loss and generation loss.
  • the adversarial loss optimizes the adversarial learning between the generator and the discriminator, enabling the generated fMRI images to gradually approximate the distribution of real data.
  • the generation loss reduces the pixel differences between the generated fMRI images and real fMRI images, further improving the quality of the generated images.
  • adversarial loss and generation loss i.e., L1 loss
  • the loss function of the generative adversarial network model is:
  • G represents the generator
  • D represents the discriminator
  • x represents the MRI image
  • y represents the real fMRI image
  • z represents random Gaussian noise
  • D(x,y) represents the true/false classification result obtained by the discriminator between the MRI image and the real fMRI image.
  • G(x,z) represent the expectation of logD(x,y)
  • G(x,z) represent the result obtained by the generator after the MRI image and random Gaussian noise are fed into it
  • D(x,G(x,z)) represent the result obtained by the discriminator after the MRI image and the fMRI image generated by the generator are passed through it. This represents the expectation of log(1-D(x,G(x,z))).
  • Step S30 Obtain the MRI image of the target object to be acquired. Input the MRI image of the target object to be acquired into the trained generative adversarial network model to directly generate the fMRI image of the target object to be acquired.
  • a well-trained GAN model (which has learned the mapping relationship between MRI and fMRI images) can be obtained. Then, for a new target object (e.g., a new Parkinson's patient), the MRI image of the target object can be directly input into the well-trained GAN model. The well-trained GAN model can then directly generate a high-resolution fMRI image (high fidelity and realism) of the target object based on the MRI image.
  • This allows for the simultaneous acquisition of both MRI and fMRI images for the target object, compensating for the missing fMRI data of patients (e.g., Parkinson's patients) and providing comprehensive multimodal data support.
  • patients e.g., Parkinson's patients
  • This further provides more evidence (information) to improve the accuracy of diagnosis and treatment, such as generating high-resolution functional MRI images for Parkinson's patients, thereby supplementing and enhancing the original fMRI data.
  • fMRI images generated by GAN models have higher resolution and clarity, and can more accurately reflect the details of brain functional activities, which helps to provide more reliable evidence for improving the diagnostic accuracy of Parkinson's disease.
  • the method of the present invention reduces reliance on expensive and time-consuming fMRI data acquisition, thereby reducing the complexity and cost of clinical procedures.
  • By generating high-quality functional MRI images complete multimodal data can be obtained for each patient.
  • High-quality functional MRI images provide clinicians with more valuable information, helping to diagnose Parkinson's disease more accurately and develop more effective treatment plans. This not only improves the accuracy of diagnosis but also allows for better monitoring of disease progression and treatment effectiveness.
  • the method of the present invention can be widely applied to different clinical scenarios, not limited to the diagnosis and research of Parkinson's disease, but can also be extended to the research of other neurodegenerative diseases, thereby improving the application value of multimodal data in the medical field.
  • this invention effectively solves the problems of difficult fMRI data acquisition and low image resolution by utilizing GAN models to generate high-quality functional MRI images, providing strong support for the diagnosis and treatment of Parkinson's disease.
  • This innovative method not only improves diagnostic accuracy but also reduces the complexity and cost of clinical procedures, demonstrating broad application prospects and significant clinical value.
  • This invention alleviates the problem of insufficient paired data through data augmentation techniques and transfer learning. Simultaneously, it explores semi-supervised learning methods to further reduce reliance on paired data. During the training process, this invention employs... Several techniques, such as spectral normalization and experience replay, were used to improve training stability and reduce the occurrence of mode collapse and gradient vanishing.
  • this invention also provides an fMRI generation system based on a GAN model, wherein the fMRI generation system based on a GAN model includes:
  • Data acquisition module 51 is used to collect paired MRI and fMRI data of multiple target acquisition objects and to build a training dataset based on the paired MRI and fMRI data;
  • the model training module 52 is used to train the generator and discriminator in the generative adversarial network model using the training dataset, optimize the network parameters of the generator and the discriminator, and obtain a trained generative adversarial network model.
  • the image generation module 53 is used to acquire the MRI image of the target object to be acquired, input the MRI image of the target object to be acquired into the trained generative adversarial network model, and directly generate the fMRI image of the target object to be acquired.
  • the present invention also provides a terminal, which includes a processor 10, a memory 20, and a display 30.
  • Figure 4 only shows some components of the terminal; however, it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.
  • the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices.
  • the memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal.
  • the memory 20 can also be used to temporarily store data that has been output or will be output.
  • the memory 20 stores an fMRI generation program 40 based on a generative adversarial network (GAN) model, which can be executed by the processor 10 to implement the fMRI generation method based on a GAN model in this application.
  • GAN generative adversarial network
  • the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the fMRI generation method based on the generative adversarial network model.
  • CPU central processing unit
  • microprocessor or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the fMRI generation method based on the generative adversarial network model.
  • the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen.
  • the display 30 is used to display information on the terminal and to display a visual user interface.
  • the components 10-30 of the terminal are...
  • the system buses communicate with each other.
  • the processor 10 executes the fMRI generation program 40 based on the generative adversarial network model in the memory 20, the following steps are performed:
  • the generator and discriminator in the generative adversarial network model are trained using the training dataset, and the network parameters of the generator and the discriminator are optimized to obtain a trained generative adversarial network model.
  • the MRI image of the target object is acquired, and then the MRI image of the target object is input into the trained generative adversarial network model to directly generate the fMRI image of the target object.
  • the process of collecting paired MRI and fMRI data from multiple target acquisition objects and establishing a training dataset based on the paired MRI and fMRI data specifically includes:
  • paired MRI and fMRI data from multiple target acquisition objects, wherein the paired MRI and fMRI data includes MRI images and real fMRI images;
  • the paired MRI and fMRI data are preprocessed, including denoising, standardization, and registration.
  • a training dataset was created based on preprocessed paired MRI and fMRI data.
  • the generative adversarial network model includes a generator and a discriminator
  • the generator is used to generate fMRI images based on MRI images
  • the discriminator is used to determine the difference between the fMRI image generated by the generator and the real fMRI image
  • the generator uses the ResNet architecture to capture multi-scale features
  • the discriminator uses a CNN architecture to improve the fine-grainedness of discrimination.
  • the step of training the generator and discriminator in the generative adversarial network model using the training dataset, optimizing the network parameters of the generator and the discriminator, and obtaining the trained generative adversarial network model includes:
  • MRI images and random Gaussian noise are input into the network structure of the generator for training to generate fMRI images;
  • the fMRI image generated by the generator is input into the network structure of the discriminator along with the real fMRI image;
  • the discriminator distinguishes between generated fMRI images and real fMRI images, and outputs a true/false classification result.
  • backpropagation optimizes the network parameters of the generator and the discriminator, so that the fMRI image generated by the generator is as close as possible to the real fMRI image, and the discriminator can distinguish between the real fMRI image and the generated fMRI image as accurately as possible.
  • the fMRI generation method based on the generative adversarial network model further includes:
  • the generator and the discriminator are optimized by using the loss function of the generative adversarial network model, so that the fMRI images generated by the generator are as similar as possible to the real fMRI data, while at the same time enhancing the discrimination ability of the discriminator.
  • the loss function of the generative adversarial network model includes adversarial loss and generative loss
  • the adversarial loss is used to optimize the adversarial learning between the generator and the discriminator, so that the fMRI images generated by the generator gradually approximate the distribution of the real data.
  • the generation loss is used to reduce the pixel difference between the generated fMRI image and the real fMRI image, thereby improving the quality of the generated image.
  • the loss function of the generative adversarial network model is:
  • G represents the generator
  • D represents the discriminator
  • x represents the MRI image
  • y represents the real fMRI image
  • z represents random Gaussian noise
  • D(x,y) represents the true/false classification result obtained by the discriminator between the MRI image and the real fMRI image.
  • G(x,z) represent the expectation of logD(x,y)
  • G(x,z) represent the result obtained by the generator after the MRI image and random Gaussian noise are fed into it
  • D(x,G(x,z)) represent the result obtained by the discriminator after the MRI image and the fMRI image generated by the generator are passed through it. This represents the expectation of log(1-D(x,G(x,z))).
  • the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an fMRI generation program based on a generative adversarial network model, and the fMRI generation program based on the generative adversarial network model, when executed by a processor, implements the steps of the fMRI generation method based on the generative adversarial network model as described above.
  • this invention provides an fMRI generation method, system, terminal, and computer-readable storage medium based on a generative adversarial network (GAN) model.
  • the method includes: collecting paired MRI and fMRI data of multiple target objects; establishing a training dataset based on the paired MRI and fMRI data; training the generator and discriminator in the GAN model using the training dataset; optimizing the network parameters of the generator and discriminator to obtain a trained GAN model; acquiring MRI images of the target objects to be acquired; and inputting the MRI images of the target objects into...
  • the trained generative adversarial network model directly generates fMRI images of the target object.
  • This invention utilizes the MRI data of the acquired object to generate corresponding fMRI data.
  • the MRI data of each acquired object can be converted into fMRI data, thus enabling each acquired object to possess both MRI and fMRI data. This not only provides accurate and reliable information for downstream tasks but also maximizes the use of existing single-modal data, avoids data waste, reduces reliance on fMRI data acquisition, and lowers the complexity and cost of clinical procedures.
  • the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.).
  • the program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments.
  • the computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)

Abstract

本发明公开了一种基于生成对抗网络模型的fMRI生成方法及系统，所述方法包括：收集多个目标采集对象的MRI和fMRI配对数据，建立训练数据集；使用训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化生成器和鉴别器的网络参数；获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。本发明利用采集对象的MRI数据生成相应的fMRI数据，每个采集对象的MRI数据都可以转换为fMRI数据，使每个采集对象都拥有MRI和fMRI数据，不仅能够给下游任务提供准确可靠的信息，还能最大限度地利用现有的单模态数据，避免数据浪费，可以减少对fMRI数据采集的依赖，降低临床操作的复杂性和成本。

Description

一种基于生成对抗网络模型的fMRI生成方法及系统 技术领域
本发明涉磁共振成像技术领域，尤其涉及一种基于生成对抗网络模型的fMRI生成方法、系统、终端及计算机可读存储介质。
背景技术
磁共振成像(MRI，Magnetic Resonance Imaging)提供了高分辨率的解剖图像，而功能磁共振成像(fMRI，functional Magnetic Resonance Imaging)则揭示了大脑的功能活动。然而，实际临床中，由于MRI采集的时间和资源限制，fMRI数据往往难以获得。这导致了在同一患者中可能只有MRI或fMRI数据，而同时拥有这两种数据的患者非常少，由于fMRI的采集成本高、时间长且对患者要求高，通常无法在每个患者身上进行。
这种情况带来了两个主要问题：第一，如果仅使用单模态数据给诊断提供信息，尽管可以避免数据不足的问题，但会显著降低诊断的准确率。第二，如果仅考虑多模态数据给诊断提供信息，则必须丢弃那些仅有单模态数据的患者信息，这不仅是对宝贵医学数据的浪费，也会减少可用的样本量，进一步限制多模态数据的实际应用。
因此，现有技术还有待于改进和发展。
发明内容
本发明的主要目的在于提供一种基于生成对抗网络模型的fMRI生成方法、系统、终端及计算机可读存储介质，旨在解决现有技术中采集fMRI图像的成本高，时间长且对采集对象要求高的问题。
为实现上述目的，本发明提供一种基于生成对抗网络模型的fMRI生成方法，所述基于生成对抗网络模型的fMRI生成方法包括如下步骤：
收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集；
使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型；
获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。
可选地，所述的基于生成对抗网络模型的fMRI生成方法，其中，所述收集多个目标 采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集，具体包括：
收集多个目标采集对象的MRI和fMRI配对数据，所述MRI和fMRI配对数据包括MRI图像和真实的fMRI图像；
对所述MRI和fMRI配对数据进行预处理，所述预处理包括去噪、标准化和配准；
根据预处理后的MRI和fMRI配对数据建立训练数据集。
可选地，所述的基于生成对抗网络模型的fMRI生成方法，其中，所述生成对抗网络模型包括生成器和鉴别器；
所述生成器用于根据MRI图像生成fMRI图像；
所述鉴别器用于判别所述生成器生成的fMRI图像与真实的fMRI图像之间的差异；
所述生成器使用ResNet架构，用于捕捉多尺度特征；
所述鉴别器使用CNN架构，用于提高判别的细粒度。
可选地，所述的基于生成对抗网络模型的fMRI生成方法，其中，所述使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型，具体包括：
将MRI图像和随机高斯噪声输入到所述生成器的网络结构中进行训练，生成fMRI图像；
将所述生成器生成的fMRI图像与真实的fMRI图像一起输入到所述鉴别器的网络结构中；
所述鉴别器区分生成的fMRI图像和真实的fMRI图像，输出真假分类结果；
根据所述鉴别器的所述真假分类结果，反向传播优化所述生成器和所述鉴别器的网络参数，使得所述生成器生成的fMRI图像尽可能逼近真实的fMRI图像，使得所述鉴别器尽可能准确地区分真实的fMRI图像和生成的fMRI图像。
可选地，所述的基于生成对抗网络模型的fMRI生成方法，其中，所述基于生成对抗网络模型的fMRI生成方法还包括：
通过所述生成对抗网络模型的损失函数对所述生成器和所述鉴别器进行优化，使所述生成器生成的fMRI图像与真实的fMRI数据尽量相似，并同时增强所述鉴别器的辨别能力。
可选地，所述的基于生成对抗网络模型的fMRI生成方法，其中，所述生成对抗网络模型的损失函数包括对抗损失和生成损失；
所述对抗损失用于优化生成器和鉴别器之间的对抗学习，使生成器生成的fMRI图像 逐步逼近真实数据的分布；
所述生成损失用于减少生成的fMRI图像与真实的fMRI图像之间的像素差异，以提高生成图像的质量。
可选地，所述的基于生成对抗网络模型的fMRI生成方法，其中，所述生成对抗网络模型的损失函数为：
其中，表示生成对抗网络模型的损失函数，表示对抗损失，表示生成损失；
所述对抗损失为：
其中，G表示生成器，D表示鉴别器，x表示MRI图像，y表示真实的fMRI图像，z表示随机高斯噪声，D(x,y)表示MRI图像和真实的fMRI图像经过鉴别器得到的真假分类结果，表示logD(x,y)的期望，G(x,z)表示MRI图像和随机高斯噪声进入生成器得到的结果，D(x,G(x,z))表示MRI图像和生成器生成的fMRI图像经过鉴别器得到的结果，表示log(1-D(x,G(x,z)))的期望，
所述生成损失为：
其中，λ表示一个给定的参数，表示||y-G(x,z)||1的期望。
此外，为实现上述目的，本发明还提供一种基于生成对抗网络模型的fMRI生成系统，其中，所述基于生成对抗网络模型的fMRI生成系统包括：
数据获取模块，用于收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集；
模型训练模块，用于使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型；
图像生成模块，用于获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。
此外，为实现上述目的，本发明还提供一种终端，其中，所述终端包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的基于生成对抗网络模型的fMRI生成程序，所述基于生成对抗网络模型的fMRI生成程序被所述处理器执行时实现如上所述的基于生成对抗网络模型的fMRI生成方法的步骤。
此外，为实现上述目的，本发明还提供一种计算机可读存储介质，其中，所述计算机可读存储介质存储有基于生成对抗网络模型的fMRI生成程序，所述基于生成对抗网络模型的fMRI生成程序被处理器执行时实现如上所述的基于生成对抗网络模型的fMRI生成方法的步骤。
本发明中，收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集；使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型；获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。本发明利用采集对象的核磁共振数据生成相应的fMRI数据，每个采集对象的MRI数据都可以转换为fMRI数据，从而使每个采集对象都拥有MRI和fMRI数据，不仅能够给下游任务提供准确可靠的信息，还能最大限度地利用现有的单模态数据，避免数据浪费，可以减少对fMRI数据采集的依赖，降低临床操作的复杂性和成本。
附图说明
图1是本发明基于生成对抗网络模型的fMRI生成方法的较佳实施例的流程图；
图2是本发明基于生成对抗网络模型的fMRI生成方法的较佳实施例中生成对抗网络模型进行训练的示意图；
图3是本发明基于生成对抗网络模型的fMRI生成系统的较佳实施例的原理示意图；
图4为本发明终端的较佳实施例的运行环境示意图。
具体实施方式
为使本发明的目的、技术方案及优点更加清楚、明确，以下参照附图并举实施例对本发明进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
目前，使用GAN(Generative adversarial network，生成对抗网络)生成医学影像的研究已经取得了一定的进展。例如SRGAN被用于提升医学图像的分辨率，虽然该方法不直接用于模态转换，但它展示了GANs在提高图像质量方面的潜力。CycleGAN是一种不需要成对训练数据的生成对抗网络，广泛用于不同模态医学图像之间的转换，CycleGAN可以在不需要成对图像的情况下进行模态转换，但其生成的图像质量和一致性仍然需要改进。例如pix2pix模型被用于多种图像到图像的转换任务，包括CT到MRI的转换，这一方法 依赖于成对的训练数据，通过条件生成对抗网络实现了高质量的图像转换。
尽管现有的基于GANs的方法在医学图像生成和转换方面取得了一定的进展，但它们在应用过程中仍然存在一些显著的缺点。第一，现有的方法，如CycleGAN，在没有成对训练数据的情况下进行模态转换时，生成的图像质量和一致性仍然存在不足。生成的图像可能会出现伪影或细节缺失，影响其在临床中的实际应用。第二，使用GAN模型等方法需要大量的成对训练数据，而在实际临床中，获取足够数量的配对MRI和fMRI数据是非常困难的。数据不足会影响模型的泛化能力和生成图像的质量。第三，使用GAN模型等方法需要大量的成对训练数据，而在实际临床中，获取足够数量的配对MRI和fMRI数据是非常困难的。数据不足会影响模型的泛化能力和生成图像的质量。
本发明较佳实施例所述的基于生成对抗网络模型的fMRI生成方法，如图1所示，所述基于生成对抗网络模型的fMRI生成方法包括以下步骤：
步骤S10、收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集。
具体地，收集多个目标采集对象(例如帕金森病患者)的MRI和fMRI配对数据(同一个患者的MRI和fMRI数据，这两个数据的空间位置是对齐的)，所述MRI和fMRI配对数据包括MRI图像和真实的fMRI图像(即MRI数据和fMRI数据)；对所述MRI和fMRI配对数据进行预处理，所述预处理包括去噪、标准化和配准等操作，确保图像的质量和对齐，例如利用3Dslicer软件对MRI和fMRI数据进行配准，以确保训练数据的准确性和一致性，配准过程包括对齐MRI和fMRI图像，使它们具有相同的空间分辨率和几何形状。根据预处理后的MRI和fMRI配对数据建立训练数据集。
步骤S20、使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型。
具体地，所述生成对抗网络模型包括生成器G和鉴别器D；所述生成器G用于根据MRI图像生成fMRI图像；所述鉴别器D用于判别所述生成器生成的fMRI图像与真实的fMRI图像之间的差异；所述生成器G使用ResNet架构(残差网络)，以其在图像生成任务中的强大性能著称，ResNet通过残差块有效解决了深层网络中的梯度消失问题，提高了生成图像的质量，生成器G主要学习MRI和fMRI之间的映射，用于捕捉多尺度特征；所述鉴别器D使用CNN架构(卷积神经网络)，以其在图像分类任务中的优异表现而闻名，CNN能够提取图像的高级特征，准确区分生成的fMRI和真实的fMRI，用于提高判别的细粒度。
如图2所示，使用所述训练数据集对生成对抗网络模型(GAN)中的生成器和鉴别器 进行训练的过程具体包括：
将MRI图像和随机高斯噪声输入到所述生成器G的网络结构中进行训练，生成fMRI图像；将所述生成器G生成的fMRI图像与真实的fMRI图像一起输入到所述鉴别器D的网络结构中；所述鉴别器D区分生成的fMRI图像和真实的fMRI图像，输出真假分类结果；根据所述鉴别器D的所述真假分类结果，反向传播优化所述生成器G和所述鉴别器D的网络参数，使得所述生成器G生成的fMRI图像尽可能逼近真实的fMRI图像，使得所述鉴别器D尽可能准确地区分真实的fMRI图像和生成的fMRI图像。整个训练过程就是对抗训练。
即生成器G的目标是最小化生成fMRI图像与真实fMRI图像之间的差异，使得生成的fMRI图像尽量逼近真实的fMRI图像。鉴别器的目标是最大化区分生成的fMRI图像和真实的fMRI图像的准确性。
进一步地，通过所述生成对抗网络模型的损失函数对所述生成器和所述鉴别器进行优化，使所述生成器生成的fMRI图像与真实的fMRI数据尽量相似，并同时增强所述鉴别器的辨别能力。
其中，所述生成对抗网络模型的损失函数包括对抗损失和生成损失；所述对抗损失用于优化生成器和鉴别器之间的对抗学习，使生成器生成的fMRI图像逐步逼近真实数据的分布；所述生成损失用于减少生成的fMRI图像与真实的fMRI图像之间的像素差异，进一步提高生成图像的质量。通过结合对抗损失和生成损失(即L1损失)，提升生成图像的细节和一致性。
所述生成对抗网络模型的损失函数为：
其中，表示生成对抗网络模型的损失函数，表示对抗损失，表示生成损失；
所述对抗损失为：
其中，G表示生成器，D表示鉴别器，x表示MRI图像，y表示真实的fMRI图像，z表示随机高斯噪声，D(x,y)表示MRI图像和真实的fMRI图像经过鉴别器得到的真假分类结果，表示logD(x,y)的期望，G(x,z)表示MRI图像和随机高斯噪声进入生成器得到的结果，D(x,G(x,z))表示MRI图像和生成器生成的fMRI图像经过鉴别器得到的结果，表示log(1-D(x,G(x,z)))的期望，
所述生成损失为：
其中，λ表示一个给定的参数，表示||y-G(x,z)||1的期望。
步骤S30、获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。
具体地，当使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练后，就可以得到训练好的生成对抗网络模型(GAN模型已学习MRI与fMRI图像之间的映射关系)，那么针对新的待采集目标对象(例如新的帕金森患者)，就可以直接将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，那么训练好的生成对抗网络模型就可以根据MRI图像直接生成待采集目标对象高分辨率的fMRI图像(高保真度和逼真度)，这样针对待采集目标对象就可以同时得到MRI图像和fMRI图像，可以弥补患者(例如帕金森患者)缺失的fMRI数据，提供全面的多模态数据支持，从而进一步为提高诊断和治疗的准确性提供更多的依据(信息)，例如针对帕金森病患者生成高分辨率的功能MRI图像，从而补充和增强原有的fMRI数据。
本发明的有益效果：
(1)、通过GAN模型生成的fMRI图像具有更高的分辨率和清晰度，能够更准确地反映大脑功能活动的细节，有助于为提高帕金森病的诊断准确性提供更多可靠依据。
(2)、本发明的方法减少了对昂贵且耗时的fMRI数据采集的依赖，降低了临床操作的复杂性和成本。通过生成高质量的功能MRI图像，使得每个患者都能够获得完整的多模态数据。
(3)、高质量的功能MRI图像为临床医生提供了更多有价值的信息，帮助更准确地诊断帕金森病，并制定更有效的治疗方案。这不仅提高了诊断的准确性，还能更好地监测疾病的进展和治疗效果。
(4)、本发明的方法可以广泛应用于不同的临床场景，不仅限于帕金森病的诊断和研究，还可以扩展到其他神经退行性疾病的研究中，提高多模态数据在医学领域的应用价值。
总之，本发明通过利用GAN模型生成高质量的功能MRI图像，有效解决了fMRI数据采集难、图像分辨率低的问题，为帕金森病的诊断和治疗提供了强有力的支持。这种创新的方法不仅提高了诊断准确性，还降低了临床操作的复杂性和成本，具有广泛的应用前景和巨大的临床价值。
本发明通过数据增强技术和迁移学习等手段，缓解了成对数据不足的问题。同时，本发明也探索了半监督学习方法，进一步降低了对成对数据的依赖。本发明在训练过程中采 用了多种技术，如谱归一化和经验重放，以提高训练的稳定性，减少模式崩溃和梯度消失的发生。
进一步地，如图3所示，基于上述基于生成对抗网络模型的fMRI生成方法，本发明还相应提供了一种基于生成对抗网络模型的fMRI生成系统，其中，所述基于生成对抗网络模型的fMRI生成系统包括：
数据获取模块51，用于收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集；
模型训练模块52，用于使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型；
图像生成模块53，用于获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。
进一步地，如图4所示，基于上述基于生成对抗网络模型的fMRI生成方法和系统，本发明还相应提供了一种终端，所述终端包括处理器10、存储器20及显示器30。图4仅示出了终端的部分组件，但是应理解的是，并不要求实施所有示出的组件，可以替代的实施更多或者更少的组件。
所述存储器20在一些实施例中可以是所述终端的内部存储单元，例如终端的硬盘或内存。所述存储器20在另一些实施例中也可以是所述终端的外部存储设备，例如所述终端上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字(Secure Digital,SD)卡，闪存卡(Flash Card)等。进一步地，所述存储器20还可以既包括所述终端的内部存储单元也包括外部存储设备。所述存储器20用于存储安装于所述终端的应用软件及各类数据，例如所述安装终端的程序代码等。所述存储器20还可以用于暂时地存储已经输出或者将要输出的数据。在一实施例中，存储器20上存储有基于生成对抗网络模型的fMRI生成程序40，该基于生成对抗网络模型的fMRI生成程序40可被处理器10所执行，从而实现本申请中基于生成对抗网络模型的fMRI生成方法。
所述处理器10在一些实施例中可以是一中央处理器(Central Processing Unit,CPU)，微处理器或其他数据处理芯片，用于运行所述存储器20中存储的程序代码或处理数据，例如执行所述基于生成对抗网络模型的fMRI生成方法等。
所述显示器30在一些实施例中可以是LED显示器、液晶显示器、触控式液晶显示器以及OLED(Organic Light-Emitting Diode，有机发光二极管)触摸器等。所述显示器30用于显示在所述终端的信息以及用于显示可视化的用户界面。所述终端的部件10-30通过 系统总线相互通信。
在一实施例中，当处理器10执行所述存储器20中基于生成对抗网络模型的fMRI生成程序40时实现如下步骤：
收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集；
使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型；
获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。
其中，所述收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集，具体包括：
收集多个目标采集对象的MRI和fMRI配对数据，所述MRI和fMRI配对数据包括MRI图像和真实的fMRI图像；
对所述MRI和fMRI配对数据进行预处理，所述预处理包括去噪、标准化和配准；
根据预处理后的MRI和fMRI配对数据建立训练数据集。
其中，所述生成对抗网络模型包括生成器和鉴别器；
所述生成器用于根据MRI图像生成fMRI图像；
所述鉴别器用于判别所述生成器生成的fMRI图像与真实的fMRI图像之间的差异；
所述生成器使用ResNet架构，用于捕捉多尺度特征；
所述鉴别器使用CNN架构，用于提高判别的细粒度。
其中，所述使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型，具体包括：
将MRI图像和随机高斯噪声输入到所述生成器的网络结构中进行训练，生成fMRI图像；
将所述生成器生成的fMRI图像与真实的fMRI图像一起输入到所述鉴别器的网络结构中；
所述鉴别器区分生成的fMRI图像和真实的fMRI图像，输出真假分类结果；
根据所述鉴别器的所述真假分类结果，反向传播优化所述生成器和所述鉴别器的网络参数，使得所述生成器生成的fMRI图像尽可能逼近真实的fMRI图像，使得所述鉴别器尽可能准确地区分真实的fMRI图像和生成的fMRI图像。
其中，所述基于生成对抗网络模型的fMRI生成方法还包括：
通过所述生成对抗网络模型的损失函数对所述生成器和所述鉴别器进行优化，使所述生成器生成的fMRI图像与真实的fMRI数据尽量相似，并同时增强所述鉴别器的辨别能力。
其中，所述生成对抗网络模型的损失函数包括对抗损失和生成损失；
所述对抗损失用于优化生成器和鉴别器之间的对抗学习，使生成器生成的fMRI图像逐步逼近真实数据的分布；
所述生成损失用于减少生成的fMRI图像与真实的fMRI图像之间的像素差异，以提高生成图像的质量。
其中，所述生成对抗网络模型的损失函数为：
其中，表示生成对抗网络模型的损失函数，表示对抗损失，表示生成损失；
所述对抗损失为：
其中，G表示生成器，D表示鉴别器，x表示MRI图像，y表示真实的fMRI图像，z表示随机高斯噪声，D(x,y)表示MRI图像和真实的fMRI图像经过鉴别器得到的真假分类结果，表示logD(x,y)的期望，G(x,z)表示MRI图像和随机高斯噪声进入生成器得到的结果，D(x,G(x,z))表示MRI图像和生成器生成的fMRI图像经过鉴别器得到的结果，表示log(1-D(x,G(x,z)))的期望，
所述生成损失为：
其中，λ表示一个给定的参数，表示||y-G(x,z)||1的期望。
本发明还提供一种计算机可读存储介质，其中，所述计算机可读存储介质存储有基于生成对抗网络模型的fMRI生成程序，所述基于生成对抗网络模型的fMRI生成程序被处理器执行时实现如上所述的基于生成对抗网络模型的fMRI生成方法的步骤。
综上所述，本发明提供一种基于生成对抗网络模型的fMRI生成方法、系统、终端及计算机可读存储介质，所述方法包括：收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集；使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型；获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入 到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。本发明利用采集对象的核磁共振数据生成相应的fMRI数据，每个采集对象的MRI数据都可以转换为fMRI数据，从而使每个采集对象都拥有MRI和fMRI数据，不仅能够给下游任务提供准确可靠的信息，还能最大限度地利用现有的单模态数据，避免数据浪费，可以减少对fMRI数据采集的依赖，降低临床操作的复杂性和成本。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者终端不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者终端所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者终端中还存在另外的相同要素。
当然，本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关硬件(如处理器，控制器等)来完成，所述的程序可存储于一计算机可读取的计算机可读存储介质中，所述程序在执行时可包括如上述各方法实施例的流程。其中所述的计算机可读存储介质可为存储器、磁碟、光盘等。
应当理解的是，本发明的应用不限于上述的举例，对本领域普通技术人员来说，可以根据上述说明加以改进或变换，所有这些改进和变换都应属于本发明所附权利要求的保护范围。

Claims (10)

  1. 一种基于生成对抗网络模型的fMRI生成方法，其特征在于，所述基于生成对抗网络模型的fMRI生成方法包括：
    收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集；
    使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型；
    获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。
  2. 根据权利要求1所述的基于生成对抗网络模型的fMRI生成方法，其特征在于，所述收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集，具体包括：
    收集多个目标采集对象的MRI和fMRI配对数据，所述MRI和fMRI配对数据包括MRI图像和真实的fMRI图像；
    对所述MRI和fMRI配对数据进行预处理，所述预处理包括去噪、标准化和配准；
    根据预处理后的MRI和fMRI配对数据建立训练数据集。
  3. 根据权利要求2所述的基于生成对抗网络模型的fMRI生成方法，其特征在于，所述生成对抗网络模型包括生成器和鉴别器；
    所述生成器用于根据MRI图像生成fMRI图像；
    所述鉴别器用于判别所述生成器生成的fMRI图像与真实的fMRI图像之间的差异；
    所述生成器使用ResNet架构，用于捕捉多尺度特征；
    所述鉴别器使用CNN架构，用于提高判别的细粒度。
  4. 根据权利要求3所述的基于生成对抗网络模型的fMRI生成方法，其特征在于，所述使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型，具体包括：
    将MRI图像和随机高斯噪声输入到所述生成器的网络结构中进行训练，生成fMRI图像；
    将所述生成器生成的fMRI图像与真实的fMRI图像一起输入到所述鉴别器的网络结构中；
    所述鉴别器区分生成的fMRI图像和真实的fMRI图像，输出真假分类结果；
    根据所述鉴别器的所述真假分类结果，反向传播优化所述生成器和所述鉴别器的网络参数，使得所述生成器生成的fMRI图像尽可能逼近真实的fMRI图像，使得所述鉴别器尽可能准确地区分真实的fMRI图像和生成的fMRI图像。
  5. 根据权利要求4所述的基于生成对抗网络模型的fMRI生成方法，其特征在于，所述基于生成对抗网络模型的fMRI生成方法还包括：
    通过所述生成对抗网络模型的损失函数对所述生成器和所述鉴别器进行优化，使所述生成器生成的fMRI图像与真实的fMRI数据尽量相似，并同时增强所述鉴别器的辨别能力。
  6. 根据权利要求5所述的基于生成对抗网络模型的fMRI生成方法，其特征在于，所述生成对抗网络模型的损失函数包括对抗损失和生成损失；
    所述对抗损失用于优化生成器和鉴别器之间的对抗学习，使生成器生成的fMRI图像逐步逼近真实数据的分布；
    所述生成损失用于减少生成的fMRI图像与真实的fMRI图像之间的像素差异，以提高生成图像的质量。
  7. 根据权利要求6所述的基于生成对抗网络模型的fMRI生成方法，其特征在于，所述生成对抗网络模型的损失函数为：
    其中，表示生成对抗网络模型的损失函数，表示对抗损失，表示生成损失；
    所述对抗损失为：
    其中，G表示生成器，D表示鉴别器，x表示MRI图像，y表示真实的fMRI图像，z表示随机高斯噪声，D(x,y)表示MRI图像和真实的fMRI图像经过鉴别器得到的真假分类结果，表示logD(x,y)的期望，G(x,z)表示MRI图像和随机高斯噪声进入生成器得到的结果，d(x,G(x,z))表示MRI图像和生成器生成的fMRI图像经过鉴别器得到的 结果，表示log(1-D(x,G(x,z)))的期望，
    所述生成损失为：
    其中，λ表示一个给定的参数，表示||y-G(x,z)||1的期望。
  8. 一种基于生成对抗网络模型的fMRI生成系统，其特征在于，所述基于生成对抗网络模型的fMRI生成系统包括：
    数据获取模块，用于收集多个目标采集对象的MRI和fMRI配对数据，根据MRI和fMRI配对数据建立训练数据集；
    模型训练模块，用于使用所述训练数据集对生成对抗网络模型中的生成器和鉴别器进行训练，优化所述生成器和所述鉴别器的网络参数，得到训练好的生成对抗网络模型；
    图像生成模块，用于获取待采集目标对象的MRI图像，将待采集目标对象的MRI图像输入到训练好的生成对抗网络模型，直接生成待采集目标对象的fMRI图像。
  9. 一种终端，其特征在于，所述终端包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的基于生成对抗网络模型的fMRI生成程序，所述基于生成对抗网络模型的fMRI生成程序被所述处理器执行时实现如权利要求1-7任一项所述的基于生成对抗网络模型的fMRI生成方法的步骤。
  10. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有基于生成对抗网络模型的fMRI生成程序，所述基于生成对抗网络模型的fMRI生成程序被处理器执行时实现如权利要求1-7任一项所述的基于生成对抗网络模型的fMRI生成方法的步骤。
PCT/CN2024/106346 2024-07-15 2024-07-19 一种基于生成对抗网络模型的fMRI生成方法及系统 Pending WO2026016159A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202410942672.8 2024-07-15
CN202410942672.8A CN118799430A (zh) 2024-07-15 2024-07-15 一种基于生成对抗网络模型的fMRI生成方法及系统

Publications (1)

Publication Number Publication Date
WO2026016159A1 true WO2026016159A1 (zh) 2026-01-22

Family

ID=93029384

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2024/106346 Pending WO2026016159A1 (zh) 2024-07-15 2024-07-19 一种基于生成对抗网络模型的fMRI生成方法及系统

Country Status (2)

Country Link
CN (1) CN118799430A (zh)
WO (1) WO2026016159A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113205567A (zh) * 2021-04-25 2021-08-03 复旦大学 基于深度学习的mri影像合成ct影像的方法
CN117349230A (zh) * 2023-10-31 2024-01-05 浙江传媒学院 一种ct体数据转换mri体数据的自动生成方法
KR102631389B1 (ko) * 2023-11-13 2024-01-30 주식회사 뉴큐어엠 Ct 영상 기반의 mri 영상 생성 방법 및 시스템
CN117635418A (zh) * 2024-01-25 2024-03-01 南京信息工程大学 生成对抗网络的训练方法、双向图像风格转换方法和装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113205567A (zh) * 2021-04-25 2021-08-03 复旦大学 基于深度学习的mri影像合成ct影像的方法
CN117349230A (zh) * 2023-10-31 2024-01-05 浙江传媒学院 一种ct体数据转换mri体数据的自动生成方法
KR102631389B1 (ko) * 2023-11-13 2024-01-30 주식회사 뉴큐어엠 Ct 영상 기반의 mri 영상 생성 방법 및 시스템
CN117635418A (zh) * 2024-01-25 2024-03-01 南京信息工程大学 生成对抗网络的训练方法、双向图像风格转换方法和装置

Also Published As

Publication number Publication date
CN118799430A (zh) 2024-10-18

Similar Documents

Publication Publication Date Title
US12288327B2 (en) Image-driven brain atlas construction method, apparatus, device and storage medium
US10902588B2 (en) Anatomical segmentation identifying modes and viewpoints with deep learning across modalities
KR101857624B1 (ko) 임상 정보를 반영한 의료 진단 방법 및 이를 이용하는 장치
US9892361B2 (en) Method and system for cross-domain synthesis of medical images using contextual deep network
US8958614B2 (en) Image-based detection using hierarchical learning
US10984894B2 (en) Automated image quality control apparatus and methods
Abbasi et al. Deep learning for the harmonization of structural MRI scans: a survey
US9142030B2 (en) Systems, methods and computer readable storage media storing instructions for automatically segmenting images of a region of interest
US20110216954A1 (en) Hierarchical atlas-based segmentation
JP2024054748A (ja) 言語特徴抽出モデルの生成方法、情報処理装置、情報処理方法及びプログラム
CN120543394B (zh) 一种基于Transformer和双分支特征解耦的多模态医学图像融合方法、系统、设备及介质
CN111243052A (zh) 图像重建方法、装置、计算机设备和存储介质
EP4304470A2 (en) Systems, devices, and methods for harmonization of imaging datasets including biomarkers
CN118037615A (zh) 一种肿瘤分割引导的磁共振图像翻译方法、系统、设备及介质
CN120316281A (zh) 基于多尺度特征融合与问题增强的医学视觉问答方法
CN110490841A (zh) 计算机辅助影像分析方法、计算机设备和存储介质
CN110246566A (zh) 基于卷积神经网络的品行障碍确定方法、系统和存储介质
US20250054136A1 (en) A computer implemented method, a method and a system
Qi et al. Blind image quality assessment for MRI with a deep three-dimensional content-adaptive hyper-network
WO2026016159A1 (zh) 一种基于生成对抗网络模型的fMRI生成方法及系统
CN118762081A (zh) 图像处理方法、装置、计算设备及存储介质
Raza et al. CycleGAN with mutual information loss constraint generates structurally aligned CT images from functional EIT images
Mazher et al. Towards Generalisable Foundation Models for Brain MRI
Balasubramaniam et al. Medical Image Enhancement for Improved Diagnostic Accuracy Using Generative Adversarial Network
CN110874614A (zh) 脑图像分类方法、计算机设备和可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 24947547

Country of ref document: EP

Kind code of ref document: A1