WO2023102709A1 - 基于静态pet图像的动态参数图像合成方法、系统 - Google Patents

基于静态pet图像的动态参数图像合成方法、系统 Download PDF

Info

Publication number
WO2023102709A1
WO2023102709A1 PCT/CN2021/135956 CN2021135956W WO2023102709A1 WO 2023102709 A1 WO2023102709 A1 WO 2023102709A1 CN 2021135956 W CN2021135956 W CN 2021135956W WO 2023102709 A1 WO2023102709 A1 WO 2023102709A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
dynamic parameter
suv
image
original
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.)
Ceased
Application number
PCT/CN2021/135956
Other languages
English (en)
French (fr)
Inventor
胡战利
王海燕
郑海荣
梁栋
刘新
杨永峰
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
Priority to PCT/CN2021/135956 priority Critical patent/WO2023102709A1/zh
Publication of WO2023102709A1 publication Critical patent/WO2023102709A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • 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
    • 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—Two-dimensional [2D] image generation

Definitions

  • the invention belongs to the technical field of biomedical engineering, and in particular relates to a dynamic parameter image synthesis method and system based on static PET images.
  • Fluorodeoxyglucose (18F-FDG) positron emission tomography (PET)/computed tomography (Computed Tomography, CT) is used for the detection and staging of primary lung tumors and distant metastases Considered a first-line tool for lung cancer patients eligible for potential treatment due to high diagnostic accuracy.
  • PET positron emission tomography
  • CT Computed tomography
  • the quantitative accuracy of static PET images (also known as SUV images) commonly used in clinic may be affected by the measurement time and plasma tracer concentration changes throughout the acquisition process, and the dynamic parameter Ki image can reflect the effect of tracer on tissue. Metabolic rate for more accurate PET quantitative results.
  • the purpose of the present invention is to provide a dynamic parameter image synthesis method, system and electronic equipment based on static PET images, aiming to solve the technical problem that high-quality dynamic parameter images cannot be efficiently obtained in the prior art.
  • the present invention provides a method for synthesizing dynamic parametric images based on static PET images, comprising:
  • a dynamic parametric image is reconstructed from the dynamic parameter data.
  • the recurrent generative adversarial network model includes a generator and a discriminator; the generator is trained according to the adversarial loss and the recurrent loss; the discriminator is trained according to the adversarial loss.
  • the cyclic generation confrontation network model includes a first generator and a second generator; the first generator includes a mapping relationship between SUV data and dynamic parameter data, and the second generator includes a dynamic parameter The mapping relationship between data and SUV data.
  • the cyclic generative adversarial network model further includes a first discriminator and a second discriminator; the first discriminator is used to identify the difference between the dynamic parameter data calculated by the first generator and the original dynamic parameter data The second discriminator is used to identify the difference between the SUV data calculated by the second generator and the original SUV data.
  • the training method of the cyclically generated confrontational network model includes:
  • the cyclic generative adversarial network model is established according to the relationship.
  • the step of using a deep learning algorithm to learn the mapping relationship between the original SUV data and the original dynamic parameter data includes:
  • Adopt deep learning algorithm calculate the corresponding first synthetic dynamic parameter data and the first synthetic SUV data of described original SUV data, and the corresponding second synthetic SUV data and the second synthetic dynamic parameter data of described original dynamic parameter data;
  • a generator and a discriminator are designed, that is, the mapping relationship between the original SUV data and the original dynamic parameter data.
  • the step of calculating the adversarial loss and the loop loss according to the above data includes:
  • a cycle loss is calculated based on the raw SUV data and the first synthesized SUV data, the raw dynamic parameter data and the second synthesized dynamic parameter data.
  • the present invention also provides a dynamic parametric image synthesis system based on static PET images, including:
  • the SUV data calculation module is used to calculate the corresponding SUV data for the static PET image
  • the operation module is used to perform iterative operation on the SUV data by using the pre-trained cyclic generation confrontation network model, and calculate the corresponding dynamic parameter data;
  • An image reconstruction module configured to reconstruct a dynamic parameter image according to the dynamic parameter data.
  • the present invention also provides an electronic device, comprising:
  • the memory stores readable instructions, and when the readable instructions are executed by the processor, the method according to the first aspect is implemented.
  • the present invention provides a computer-readable storage medium, on which a computer program is stored, and the computer program implements the method in the first aspect when executed.
  • the dynamic parameter image synthesis method and system based on static PET images and the electronic equipment provided by the present invention since the pre-trained cyclic generative confrontation network model is used to directly synthesize the dynamic parameter image from the static PET image, the original 1-hour parameter The imaging is shortened to less than 1 second, which greatly shortens the scanning time of the patient, and at the same time can still provide an additional image data for the doctor, reflecting the metabolic rate of the tracer to the tissue in the patient, and improving the accuracy of PET quantification results and clinical diagnosis It effectively realizes the efficient and high-quality acquisition of dynamic parameter images.
  • Fig. 1 is a flow chart showing a method for synthesizing a dynamic parametric image based on a static PET image according to Embodiment 1.
  • Fig. 2 is a flow chart of building a recurrent generative adversarial network model according to an exemplary embodiment.
  • Fig. 3 is a schematic structural diagram of a generator and a discriminator in a recurrent generative adversarial network model according to an exemplary embodiment.
  • Fig. 4 is a schematic diagram showing experimental results of this scheme according to an exemplary embodiment.
  • Fig. 5 is a block diagram of a dynamic parametric image synthesis system based on static PET images according to Embodiment 2.
  • Fig. 1 is a flow chart of the realization of the dynamic parametric image synthesis method based on the static PET image shown in the first embodiment.
  • the dynamic parametric image synthesis method based on the static PET image shown in Embodiment 1 is applicable to electronic equipment.
  • the electronic equipment is equipped with a processor. After obtaining the static PET image, it can calculate according to the SUV data of the static PET image. Corresponding dynamic parameter data, and then reconstruct the dynamic parameter image, realize efficient and high-quality calculation of dynamic parameter image based on the static PET image.
  • Step S110 calculating corresponding SUV data for the static PET image.
  • Step S120 using the pre-trained cyclic generative adversarial network model to iteratively calculate the SUV data, and calculate the corresponding dynamic parameter data.
  • Step S130 reconstructing a dynamic parameter image according to the dynamic parameter data.
  • Routine clinical PET imaging is static PET imaging, and the quantification accuracy of the standardized uptake value (SUV) obtained by it is easily affected by changes in measurement time and plasma tracer concentration throughout the acquisition process.
  • SUV standardized uptake value
  • the corresponding SUV data will be calculated. Then, the pre-trained cyclic generative confrontation network model is used to iteratively calculate the SUV data and calculate the corresponding dynamic parameter data.
  • the dynamic parameter data can be dynamic parameter Ki data, or other types of dynamic parameter data.
  • This solution only uses the dynamic parameter Ki data as an example to describe in detail, and does not describe the specific type of dynamic parameter data. limited.
  • the recurrent generative adversarial network model is pre-built based on static PET image samples for deep learning training.
  • Fig. 2 is a flow chart of building a recurrent generative adversarial network model according to an exemplary embodiment.
  • the cycle GAN model uses the cycleGAN network algorithm, which includes a generator and a discriminator.
  • the model includes two generators (i.e. GSUV ⁇ Ki and GKi ⁇ SUV) and two discriminators (i.e. DKi and DSUV) first and second generators , for mutual constraints.
  • the first generator includes the mapping relationship between SUV data and dynamic parameter data
  • the second generator includes the mapping relationship between dynamic parameter data and SUV data
  • the first discriminator is used to identify the For the difference between the dynamic parameter data and the original dynamic parameter data
  • the second discriminator is used to identify the difference between the SUV data calculated by the second generator and the original SUV data.
  • the original SUV data and the original dynamic parameter (Ki) data are used as input, and the synthetic Ki (sKi) and synthetic SUV (sSUV) data are output through the generator GSUV ⁇ Ki/GKi ⁇ SUV, the original Ki data and Raw SUV data is used as reference.
  • the generator network is trained using adversarial loss (L2 loss function) and recurrent loss (L1 loss function), while the discriminator network uses Adversarial loss (L2 loss function) for training.
  • the adversarial loss is obtained by feeding the referenced original image and the corresponding synthetic image into the discriminator to compute the difference between the images.
  • the loop loss is computed by passing the original image through two generators and finally obtaining a synthetic image belonging to the same class as the original image.
  • the consistency loss included in the general cycleGAN network was removed when constructing the cycle GAN model.
  • the consistency loss encourages the map to preserve the color composition of the input and output images.
  • the mapping relationship between SUV data and dynamic parameter Ki data in color composition is not strong.
  • the original SUV data and original dynamic parameter data of different static PET image samples are obtained; the deep learning algorithm is used to learn the mapping relationship between the original SUV data and the original dynamic parameter data; according to the relationship Build a recurrent generative adversarial network model.
  • a deep learning algorithm is used to calculate the first synthetic dynamic parameter data and the first synthetic SUV data corresponding to the original SUV data, and the second synthetic SUV data and the second synthetic dynamic parameter data corresponding to the original dynamic parameter data; And calculate the adversarial loss and cyclic loss according to the above data; according to the adversarial loss and cyclic loss, design the generator and the discriminator, that is, the mapping relationship between the original SUV data and the original dynamic parameter data.
  • the joint loss function can be jointly determined by three loss functions, that is, the main two adversarial loss functions and one cyclic loss function, expressed as follows:
  • ⁇ 1 is the weight coefficient.
  • Adversarial loss 1 L GAN (G SUV ⁇ Ki ,D Ki ,SUV,Ki) is the Ki image generated by the judgment generator G SUV ⁇ Ki and then input into the discriminator D Ki to calculate the loss.
  • Adversarial loss 2 Similarly, L GAN (G Ki ⁇ SUV , D SUV , SUV, Ki) is the SUV image generated by the judgment generator G Ki ⁇ SUV and then input into the discriminator D SUV to calculate the loss.
  • Cyclic loss L cyc (G SUV ⁇ Ki , G Ki ⁇ SUV ): Pass the SUV image/Ki image through the generator G SUV ⁇ Ki and G Ki ⁇ SUV to finally get the generated SUV image/Ki image, thus calculating two similar Image loss.
  • the weight coefficient ⁇ 1 of the cyclic loss is generally set to 10.
  • the adversarial loss is one of its core loss functions, and its objective function is:
  • adversarial loss 1
  • N is the number of data
  • P * indicates data distribution
  • the final dynamic parameter Ki data is closer to the real dynamic parameter Ki data.
  • the Ki image is converted from G Ki ⁇ SUV to generate the second composite SUV image
  • the second Ki image is generated through G SUV ⁇ Ki .
  • the loss calculation is performed on the generated second Ki image and the original Ki image, thereby reducing the difference between the two distance.
  • both generators need to be trained properly. This also promotes the training of the generator while reducing the loss.
  • Fig. 3 is a schematic structural diagram of a generator and a discriminator in a recurrent generative adversarial network model according to an exemplary embodiment.
  • each generator consists of three parts: the encoding layer for extracting knowledge from the input image domain (composed of 3 convolution kernels, respectively 7 ⁇ 7, 3 ⁇ 3, 3 ⁇ 3 in size, The number of channels is 32, 64, and 128 in turn), used to construct the decoding layer of the target image (composed of 3 convolution kernels, respectively 3 ⁇ 3, 3 ⁇ 3, 7 ⁇ 7 in size, 64, 32, 1 number of channels) and the residual block as the intermediate transmission information (9 residual networks, each network has two convolution kernels of 3 ⁇ 3 size, both of which have 128 channels, and the first and last convolution kernels are also identical
  • the superposition of mappings to ensure that the depth of the network is increased without increasing the error at least), is used to transfer image knowledge from the input domain to the target domain.
  • Each discriminator includes 5 convolution kernels (both are 4 ⁇ 4 in size, and the number of channels is 32, 64, 128, 256, 1). The discriminator inputs a real image or a synthetic image, and then outputs a true or false binary
  • the optimizer can also be used to optimize the model, guide the parameters of the loss function to update the appropriate size in the right direction, so that the updated parameters can make the value of the loss function (objective function) approach the global minimum, effectively improving the model construction. efficiency and accuracy.
  • Adam optimizer can be used for optimization.
  • the Adam optimizer combines the advantages of two optimization algorithms, AdaGrad and RMSProp.
  • the update step size is calculated by comprehensively considering the first moment estimation of the gradient (First Moment Estimation, that is, the mean value of the gradient) and the second moment estimation (Second Moment Estimation, that is, the uncentered variance of the gradient).
  • First Moment Estimation that is, the mean value of the gradient
  • Second Moment Estimation that is, the uncentered variance of the gradient
  • the deep learning network model is used to directly learn the dynamic parameter image from the static PET image, which shortens the original parameter imaging of up to 1 hour to less than 1 second of imaging, which greatly shortens the scanning time of patients.
  • the cycleGAN network model uses three loss functions (against loss + cycle loss + consistency loss), in which the consistency loss encourages the mapping to retain the color composition of the input image and the output image, and this scheme uses an improved cycleGAN model (removing the consistency loss function) to make the generated parameter image have inconsistent color composition with the input static PET image, so as to better fit the clinical situation (in some areas of the patient's body, the intensity of SUV is very high, and the tracer reflected by Ki relatively slow flow rates).
  • Fig. 4 is a schematic diagram showing experimental results of this scheme according to an exemplary embodiment.
  • #1 represents a healthy control
  • #2 represents a lung cancer
  • #3 represents a benign lung tumor
  • the three images from top to bottom represent the original SUV image, the synthetic Ki(sKi) image and the corresponding real Ki image , with arrows pointing to participants' lesions.
  • the method of the present invention can effectively ensure the peak signal-to-noise ratio and structural similarity of the composite image, and at the same time, for lung cancer lesions and high metabolic areas, the composite image of the dynamic parameter Ki data can reach the real Ki image Comparable image quality, clear and complete lesion boundaries.
  • Fig. 5 is a block diagram of a dynamic parametric image synthesis system based on static PET images shown in Embodiment 2.
  • the system can execute all or part of the steps of any of the above-mentioned dynamic parametric image synthesis methods based on static PET images.
  • the system includes:
  • SUV data calculating module 10 for calculating corresponding SUV data for static PET image
  • the computing module 20 is used to iteratively compute the SUV data using a pre-trained cyclic generation confrontational network model to calculate corresponding dynamic parameter data;
  • An image reconstruction module 30 configured to reconstruct a dynamic parameter image according to the dynamic parameter data.
  • Embodiment 3 of the present invention provides an electronic device, which can execute all or part of the steps of any one of the methods for synthesizing dynamic parametric images based on static PET images.
  • This electronic device includes:
  • the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can execute the method described in any of the above exemplary embodiments. method, which will not be described in detail here.
  • a storage medium is also provided, which is a computer-readable storage medium, for example, a temporary or non-transitory computer-readable storage medium including instructions.
  • the storage medium for example, includes a memory of instructions, and the above instructions can be executed by a processor of the server system to complete the above-mentioned dynamic parametric image synthesis method based on static PET images.

Landscapes

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

Abstract

一种基于静态PET图像的动态参数图像合成方法及系统、电子设备，方法包括：针对静态PET图像计算相应的SUV数据（S110）；采用预先训练的循环生成对抗网络模型对SUV数据进行迭代运算，计算相应的动态参数数据（S120）；根据动态参数数据重建动态参数图像（S130）。由于采用预先训练的循环生成对抗网络模型直接从静态PET图像合成出动态参数图像，将原来长达1小时的参数成像缩短为不到1秒成像，大大缩短了患者的扫描时间，同时仍能为医生额外提供一份影像资料，反映示踪剂对患者体内组织的代谢速率，提高PET量化结果及临床诊断的准确性，有效实现了动态参数图像的高效、高质量获取。

Description

基于静态PET图像的动态参数图像合成方法、系统 技术领域
本发明属于生物医学工程技术领域，尤其涉及一种基于静态PET图像的动态参数图像合成方法、系统。
背景技术
氟代脱氧葡萄糖(18F-FDG)正电子发射断层扫描(Positron emission tomography,PET)/计算机断层扫描(Computed Tomography，CT)因其在原发性肺肿瘤和远处转移灶的检测和分期方面的高诊断准确性而被视为符合潜在治疗条件的肺癌患者的一线工具。但是，临床中常用的静态PET图像(也称SUV图像)的量化准确性可能会受到整个采集过程中测量时间和血浆示踪剂浓度变化的影响，动态参数Ki图像可以反映示踪剂对组织的代谢率，从而获得更准确的PET定量结果。
然而，动态参数Ki图像的获取需要对患者进行长时间扫描(≥60分钟)以进行参数成像和动态建模，以及连续的动脉血液采样或图像衍生的血液活动来确定输入函数，这些严重限制了参数成像的临床应用。
因此，研究和开发高效获取高质量动态参数图像，对于目前的肺癌诊断、分期领域均有着重要的科学意义与广阔的应用前景。
发明内容
本发明的目的在于提供一种基于静态PET图像的动态参数图像合成方法、系统及电子设备，旨在解决现有技术中无法高效获取高质量动态参数图像的技术问题。
第一方面，本发明提供了一种基于静态PET图像的动态参数图像合成方 法，包括：
针对静态PET图像计算相应的SUV数据；
采用预先训练的循环生成对抗网络模型对所述SUV数据进行迭代运算，计算相应的动态参数数据；
根据所述动态参数数据重建动态参数图像。
可选的，所述循环生成对抗网络模型包括生成器和鉴别器；所述生成器根据对抗性损失和循环损失进行训练而成；所述鉴别器根据对抗性损失进行训练而成。
可选的，所述循环生成对抗网络模型包括第一生成器和第二生成器；所述第一生成器包含SUV数据与动态参数数据之间的映射关系，所述第二生成器包含动态参数数据与SUV数据之间的映射关系。
可选的，所述循环生成对抗网络模型还包括第一鉴别器和第二鉴别器；所述第一鉴别器用于鉴别经所述第一生成器运算后的动态参数数据与原始动态参数数据之间的差异，所述第二鉴别器用于鉴别经所述第二生成器运算后的SUV数据与原始SUV数据之间的差异。
可选的，所述循环生成对抗网络模型的训练方法包括：
获取不同静态PET图像样本的原始SUV数据和原始动态参数数据；
采用深度学习算法学习所述原始SUV数据和所述原始动态参数数据之间的映射关系；
根据所述关系建立所述循环生成对抗网络模型。
可选的，所述采用深度学习算法学习所述原始SUV数据和所述原始动态参数数据之间的映射关系的步骤包括：
采用深度学习算法，计算所述原始SUV数据相应的第一合成动态参数数 据和第一合成SUV数据，以及所述原始动态参数数据相应的第二合成SUV数据和第二合成动态参数数据；
根据上述数据计算对抗性损失和循环损失；
根据所述对抗性损失和所述循环损失，设计生成器和鉴别器，即所述原始SUV数据和所述原始动态参数数据之间的映射关系。
可选的，所述根据上述数据计算对抗性损失和循环损失的步骤包括：
根据所述原始SUV数据和所述第二合成SUV数据、所述原始动态参数数据和第一合成动态参数数据，计算对抗性损失；
根据所述原始SUV数据和所述第一合成SUV数据、所述原始动态参数数据和第二合成动态参数数据，计算循环损失。
第二方面，本发明还提供了一种基于静态PET图像的动态参数图像合成系统，包括：
SUV数据计算模块，用于针对静态PET图像计算相应的SUV数据；
运算模块，用于采用预先训练的循环生成对抗网络模型对所述SUV数据进行迭代运算，计算相应的动态参数数据；
图像重建模块，用于根据所述动态参数数据重建动态参数图像。
第三方面，本发明还提供了一种电子设备，包括：
处理器；以及
与所述处理器通讯连接的存储器；其中，
所述存储器存储有可读性指令，所述可读性指令被所述处理器执行时实现如第一方面所述的方法。
第四方面，本发明提供了一种计算机可读性存储介质，其上存储有计算机程序，所述计算机程序在被执行时实现如第一方面的方法。
本发明提供的基于静态PET图像的动态参数图像合成方法及系统、电子设备中，由于采用预先训练的循环生成对抗网络模型直接从静态PET图像合成出动态参数图像，将原来长达1小时的参数成像缩短为不到1秒成像，大大缩短了患者的扫描时间，同时仍能为医生额外提供一份影像资料，反映示踪剂对患者体内组织的代谢速率，提高PET量化结果及临床诊断的准确性，有效实现了动态参数图像的高效、高质量获取。
附图说明
图1是根据实施例一示出的一种基于静态PET图像的动态参数图像合成方法的实现流程图。
图2是根据一示例性实施例示出的循环生成对抗网络模型构建流程图。
图3是根据一示例性实施例示出的循环生成对抗网络模型中生成器和鉴别器的结构示意图。
图4是根据一示例性实施例示出的本方案实验结果示意图。
图5是根据实施例二示出的一种基于静态PET图像的动态参数图像合成系统的框图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅 用以解释本申请，并不用于限定本申请。
以下结合具体实施例对本发明的具体实现进行详细描述：
实施例一：
图1是实施例一示出的基于静态PET图像的动态参数图像合成方法的实现流程图。实施例一示出的基于静态PET图像的动态参数图像合成方法适用于电子设备中，电子设备中设置有处理器，在获取到静态PET图像后，即可根据该静态PET图像的SUV数据，计算相应的动态参数数据，进而重建动态参数图像，实现基于静态PET图像高效、高质量计算动态参数图像。
步骤S110，针对静态PET图像计算相应的SUV数据。
步骤S120，采用预先训练的循环生成对抗网络模型对SUV数据进行迭代运算，计算相应的动态参数数据。
步骤S130，根据动态参数数据重建动态参数图像。
临床常规PET成像是静态PET成像，在整个采集过程中，其得到的标准化摄取值(standardized uptake value,SUV)的量化精度容易受到测量时间和血浆示踪剂浓度变化的影响。
本方案中，在获取到静态PET图像后，将计算相应的SUV数据。进而采用预先训练好的循环生成对抗网络模型对SUV数据进行迭代运算，计算相应的动态参数数据。
需要说明的是，该动态参数数据可以是动态参数Ki数据，也可以是其他类型的动态参数数据，本方案仅是以动态参数Ki数据为例进行详细说明，而不对动态参数数据的具体类型进行限定。
循环生成对抗网络模型是预先根据静态PET图像样本进行深度学习的训练而构建的。
图2是根据一示例性实施例示出的循环生成对抗网络模型构建流程图。如图2所示，该循环生成对抗网络模型采用cycleGAN网络算法，该模型包括生成器和鉴别器。特别的是，该模型包括第一生成器和第二生成器两个生成器(即GSUV→Ki和GKi→SUV)和第一鉴别器和第二鉴别器两个鉴别器(即DKi和DSUV)，用于相互约束。
其中，第一生成器包含SUV数据与动态参数数据之间的映射关系，第二生成器包含动态参数数据与SUV数据之间的映射关系；第一鉴别器用于鉴别经第一生成器运算后的动态参数数据与原始动态参数数据之间的差异，第二鉴别器用于鉴别经第二生成器运算后的SUV数据与原始SUV数据之间的差异。
在深度学习训练过程中，以原始SUV数据和原始动态参数(Ki)数据作为输入，经过生成器GSUV→Ki/GKi→SUV输出合成Ki(sKi)和合成SUV(sSUV)数据，原始Ki数据和原始SUV数据作为参照。
由于单一损失函数在图像生成方面存在一定局限性，为提高最终合成图像的生成质量，生成器网络使用对抗性损失(L2损失函数)和循环损失(L1损失函数)进行训练，而鉴别器网络使用对抗性损失(L2损失函数)进行训练。通过将参照的原始图像和相应的合成图像输入鉴别器以计算图像之间的差异来获得对抗性损失。将原始图像通过两个生成器，最终获得与原始图像属于同一类别的合成图像来计算循环损失。
需要特别说明的是，为进一步提高该循环生成对抗网络模型的严谨性和高效性，在构建循环生成对抗网络模型时，去除了包含在一般cycleGAN网络中的一致性损失。一致性损失鼓励映射保留输入图像和输出图像的颜色组成。实际上，SUV数据和动态参数Ki数据在颜色组成之间的映射关系并不强。
具体的，在训练生成循环生成对抗网络模型时，获取不同静态PET图像 样本的原始SUV数据和原始动态参数数据；采用深度学习算法学习原始SUV数据和原始动态参数数据之间的映射关系；根据关系建立循环生成对抗网络模型。
在构建映射关系时，采用深度学习算法，计算原始SUV数据相应的第一合成动态参数数据和第一合成SUV数据，以及原始动态参数数据相应的第二合成SUV数据和第二合成动态参数数据；并根据上述数据计算对抗性损失和循环损失；根据对抗性损失和循环损失，设计生成器和鉴别器，即原始SUV数据和原始动态参数数据之间的映射关系。
在某些区域，SUV数据的强度很高，而动态参数Ki反映的示踪剂流速可能很慢。因而，在计算对抗性损失和循环损失时，联合损失函数可通过3个损失函数共同确定，即主要的两个对抗性损失函数和一个循环损失函数，表示如下：
L total(G SUV→Ki,G Ki→SUV,D SUV,D Ki)＝L GAN(G SUV→Ki,D Ki,SUV,Ki)+L GAN(G Ki→SUV,D SUV,SUV,Ki)+λ 1L cyc(G SUV→Ki,G Ki→SUV)  (1)
其中，λ 1为权重系数。
对抗损失1：L GAN(G SUV→Ki,D Ki,SUV,Ki)是评判生成器G SUV→Ki生成的Ki图像再输入到鉴别器D Ki中去计算损失。
对抗损失2：同理，L GAN(G Ki→SUV,D SUV,SUV,Ki)是评判生成器G Ki→SUV生成的SUV图像再输入到鉴别器D SUV中去计算损失。
循环损失L cyc(G SUV→Ki,G Ki→SUV)：将SUV图像/Ki图像经过生成器G SUV→Ki和G Ki→SUV最后再得到生成的SUV图像/Ki图像，从而计算两张同类图像的损失。根据工程经验所得，循环损失的权重系数λ 1一般设置为10。
由于循环生成对抗网络模型整体基于生成对抗网络的思想，故对抗性损失 为其核心损失函数之一，其目标函数为：
Figure PCTCN2021135956-appb-000001
其中，对抗损失1：
Figure PCTCN2021135956-appb-000002
对抗损失2：
Figure PCTCN2021135956-appb-000003
为保证生成图像的峰值信噪比，我们使用L2损失函数作为对抗性损失：
Figure PCTCN2021135956-appb-000004
Figure PCTCN2021135956-appb-000005
其中，
Figure PCTCN2021135956-appb-000006
表示期望，α和β表示输入和目标数据，N为数据个数，P *表示数据分布，a，b和c表示人为设定的超参数，其中a＝0，b＝1，c＝1。
通过将对抗性损失和循环损失相互协同，使最终得到的动态参数Ki数据与真实动态参数Ki数据更为接近。例如Ki图像经G Ki→SUV生成第二合成SUV图像后，再经G SUV→Ki生成第二Ki图像，如此将生成的第二Ki图像与原始Ki图像进行损失计算，从而减少两者相差的距离。两图像若想要足够相似则需要两个生成器都训练得当。这样在降低损失时也是促进生成器的训练。
图3是根据一示例性实施例示出的循环生成对抗网络模型中生成器和鉴别器的结构示意图。
如图3所示，每个生成器包括三部分：用于从输入图像域中提取知识的编 码层(有3个卷积核组成，分别为7×7、3×3、3×3大小，依次为32、64、128通道数)，用于构建目标图像的解码层(有3个卷积核组成，分别为3×3、3×3、7×7大小，依次为64、32、1通道数)和作为中间传输信息的残差块(9个残差网络，每个网络均有两个3×3大小的卷积核，均为128通道数，卷积核首尾还进行了恒等映射的叠加，以确保增加网络深度的同时至少不会增加误差)，用于将图像知识从输入域传输到目标域。每个鉴别器包括5个卷积核(均为4×4的大小，通道数依次为32、64、128、256、1)，鉴别器输入真实图像或合成图像，然后输出真或假的二进制值。
另外，还可以采用优化器对模型进行优化，指引损失函数的各个参数往正确的方向更新合适的大小，使得更新后的各个参数让损失函数(目标函数)值不断逼近全局最小，有效提高模型构建的效率和准确性。
例如，可采用Adam优化器进行优化。
Adam优化器结合了AdaGrad和RMSProp两种优化算法的优点。对梯度的一阶矩估计(First Moment Estimation，即梯度的均值)和二阶矩估计(Second Moment Estimation，即梯度的未中心化的方差)进行综合考虑，计算出更新步长。Adam优化具体实现过程：
1.初始：学习率lr
2.初始：平滑常数(或者叫做衰减速率)，分别用于平滑m和v
3.初始：可学习参数θ 0
4.初始：m 0＝0,v 0＝0,t＝0
5.while没有停止训练do
6.训练次数更新：t＝t+1
7.计算梯度：g t(所有的可学习参数都有自己的梯度，因此表示的是全 部梯度的集合)
8.累计梯度：m t＝β 1*m t-1+(1-β 1)*g t(每个导数对应一个m，因此m也是个集合)
9.累计梯度的平方：v t＝β 2*v t-1+(1-β 2)*(g t) 2(每个导数对应一个v，因此v也是个集合)
10.偏差纠正m：
Figure PCTCN2021135956-appb-000007
11.偏差纠正v：
Figure PCTCN2021135956-appb-000008
12.更新参数：
Figure PCTCN2021135956-appb-000009
13.end while
通过本方案的动态参数图像合成方法，使用深度学习网络模型直接从静态PET图像学习出动态参数图像，将原来长达1小时的参数成像缩短为不到1秒成像，大大缩短了患者的扫描时间，同时仍能为医生额外提供一份影像资料，反映示踪剂对患者体内组织的代谢速率，提高PET量化结果及临床诊断的准确性，实现动态参数图像的高效、高质量合成；其次，一般cycleGAN网络模型使用三种损失函数(对抗损失+循环损失+一致性损失)，其中一致性损失鼓励映射保留输入图像和输出图像的颜色组成，而本方案使用改进的cycleGAN模型(去除了一致性损失函数)来使得生成出的参数图像与输入的静态PET图像有不一致的颜色组成，以此更贴合临床实际情况(在患者身体的某些区域，SUV的强度很高，而Ki反映的示踪剂流速相对较慢)。
图4是根据一示例性实施例示出的本方案实验结果示意图。如图4所示，#1代表健康对照，#2代表肺癌，#3代表肺良性肿瘤，从上至下的3个图片分别代表原始SUV图像、合成Ki(sKi)图像和相应的真实Ki图像，箭头指向参与者的病变。从图4可以看出，本发明的方法可以有效地保证合成图像的 峰值信噪比和结构相似度，同时，对于肺癌病灶和高代谢区，动态参数Ki数据的合成图像可以达到与真实Ki图像相当的图像质量，病灶边界清晰完整。
实施例二：
图5是实施例二示出的基于静态PET图像的动态参数图像合成系统的框图。该系统可执行上述任一所示的基于静态PET图像的动态参数图像合成方法的全部或者部分步骤。该系统包括：
SUV数据计算模块10，用于针对静态PET图像计算相应的SUV数据；
运算模块20，用于采用预先训练的循环生成对抗网络模型对SUV数据进行迭代运算，计算相应的动态参数数据；
图像重建模块30，用于根据动态参数数据重建动态参数图像。
实施例三：
本发明实施例三提供了一种电子设备，该电子设备可执行上述任一所示的基于静态PET图像的动态参数图像合成方法的全部或者部分步骤。该电子设备包括：
处理器；以及
与处理器通信连接的存储器；其中，
所述存储器存储有可被所述至少一个处理器执行的指令，所述指令被所述至少一个处理器执行，以使所述至少一个处理器能够执行如上述任一示例性实施例所述的方法，此处将不做详细阐述说明。
在本实施例中，还提供了一种存储介质，该存储介质为计算机可读存储介质，例如可以为包括指令的临时性和非临时性计算机可读存储介质。该存储介 质例如包括指令的存储器，上述指令可由服务器系统的处理器执行以完成上述基于静态PET图像的动态参数图像合成方法。
对所公开的实施例的上述说明，使本领域专业技术人员能够实现或使用本发明。对这些实施例的多种修改对本领域的专业技术人员来说将是显而易见的，本发明中所定义的一般原理可以在不脱离本发明的精神或范围的情况下，在其它实施例中实现。因此，本发明将不会被限制于本发明所示的这些实施例，而是要符合与本发明所公开的原理和新颖特点相一致的最宽的范围。

Claims (10)

  1. 一种基于静态PET图像的动态参数图像合成方法，其特征在于，所述方法包括：
    针对静态PET图像计算相应的SUV数据；
    采用预先训练的循环生成对抗网络模型对所述SUV数据进行迭代运算，计算相应的动态参数数据；
    根据所述动态参数数据重建动态参数图像。
  2. 如权利要求1所述的方法，其特征在于，所述循环生成对抗网络模型包括生成器和鉴别器；所述生成器根据对抗性损失和循环损失进行训练而成；所述鉴别器根据对抗性损失进行训练而成。
  3. 如权利要求1所述的方法，其特征在于，所述循环生成对抗网络模型包括第一生成器和第二生成器；所述第一生成器包含SUV数据与动态参数数据之间的映射关系，所述第二生成器包含动态参数数据与SUV数据之间的映射关系。
  4. 如权利要求3所述的方法，其特征在于，所述循环生成对抗网络模型还包括第一鉴别器和第二鉴别器；所述第一鉴别器用于鉴别经所述第一生成器运算后的动态参数数据与原始动态参数数据之间的差异，所述第二鉴别器用于鉴别经所述第二生成器运算后的SUV数据与原始SUV数据之间的差异。
  5. 如权利要求1所述的方法，其特征在于，所述循环生成对抗网络模型的训练方法包括：
    获取不同静态PET图像样本的原始SUV数据和原始动态参数数据；
    采用深度学习算法学习所述原始SUV数据和所述原始动态参数数据之间的映射关系；
    根据所述关系建立所述循环生成对抗网络模型。
  6. 如权利要求5所述的方法，其特征在于，所述采用深度学习算法学习所述原始SUV数据和所述原始动态参数数据之间的映射关系的步骤包括：
    采用深度学习算法，计算所述原始SUV数据相应的第一合成动态参数数据和第一合成SUV数据，以及所述原始动态参数数据相应的第二合成SUV数据和第二合成动态参数数据；
    根据上述数据计算对抗性损失和循环损失；
    根据所述对抗性损失和所述循环损失，设计生成器和鉴别器，即所述原始SUV数据和所述原始动态参数数据之间的映射关系。
  7. 如权利要求6所述的方法，其特征在于，所述根据上述数据计算对抗性损失和循环损失的步骤包括：
    根据所述原始SUV数据和所述第二合成SUV数据、所述原始动态参数数据和第一合成动态参数数据，计算对抗性损失；
    根据所述原始SUV数据和所述第一合成SUV数据、所述原始动态参数数据和第二合成动态参数数据，计算循环损失。
  8. 一种基于静态PET图像的动态参数图像合成系统，其特征在于，所述系统包括：
    SUV数据计算模块，用于针对静态PET图像计算相应的SUV数据；
    运算模块，用于采用预先训练的循环生成对抗网络模型对所述SUV数据进行迭代运算，计算相应的动态参数数据；
    图像重建模块，用于根据所述动态参数数据重建动态参数图像。
  9. 一种图像合成设备，其特征在于，所述图像合成设备包括：
    处理器；以及
    与所述处理器通讯连接的存储器；其中，
    所述存储器存储有可读性指令，所述可读性指令被所述处理器执行时实现如权利要求1-7任一项所述的方法。
  10. 一种计算机可读性存储介质，其上存储有计算机程序，所述计算机程序在被执行时实现如权利要求1-7任一项所述的方法。
PCT/CN2021/135956 2021-12-07 2021-12-07 基于静态pet图像的动态参数图像合成方法、系统 Ceased WO2023102709A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/135956 WO2023102709A1 (zh) 2021-12-07 2021-12-07 基于静态pet图像的动态参数图像合成方法、系统

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/135956 WO2023102709A1 (zh) 2021-12-07 2021-12-07 基于静态pet图像的动态参数图像合成方法、系统

Publications (1)

Publication Number Publication Date
WO2023102709A1 true WO2023102709A1 (zh) 2023-06-15

Family

ID=86729457

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/135956 Ceased WO2023102709A1 (zh) 2021-12-07 2021-12-07 基于静态pet图像的动态参数图像合成方法、系统

Country Status (1)

Country Link
WO (1) WO2023102709A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119313617A (zh) * 2024-09-25 2025-01-14 上海科技大学 基于跨模态生成的代谢异常检测方法、装置、终端、介质及程序产品

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108292443A (zh) * 2015-11-20 2018-07-17 皇家飞利浦有限公司 使用病变代理的pet图像重建和处理
CN110070935A (zh) * 2019-03-20 2019-07-30 中国科学院自动化研究所 基于对抗神经网络的医学图像合成方法、分类方法及装置
CN110298897A (zh) * 2019-04-03 2019-10-01 上海联影医疗科技有限公司 用于正电子发射断层成像图像重建的系统和方法
CN111770760A (zh) * 2018-02-15 2020-10-13 加利福尼亚大学董事会 使用动态pet进行肝炎症分析
WO2021063476A1 (en) * 2019-09-30 2021-04-08 Toyota Motor Europe Method for training a generative adversarial network, modified image generation module and system for detecting features in an image

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108292443A (zh) * 2015-11-20 2018-07-17 皇家飞利浦有限公司 使用病变代理的pet图像重建和处理
CN111770760A (zh) * 2018-02-15 2020-10-13 加利福尼亚大学董事会 使用动态pet进行肝炎症分析
CN110070935A (zh) * 2019-03-20 2019-07-30 中国科学院自动化研究所 基于对抗神经网络的医学图像合成方法、分类方法及装置
CN110298897A (zh) * 2019-04-03 2019-10-01 上海联影医疗科技有限公司 用于正电子发射断层成像图像重建的系统和方法
WO2021063476A1 (en) * 2019-09-30 2021-04-08 Toyota Motor Europe Method for training a generative adversarial network, modified image generation module and system for detecting features in an image

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119313617A (zh) * 2024-09-25 2025-01-14 上海科技大学 基于跨模态生成的代谢异常检测方法、装置、终端、介质及程序产品

Similar Documents

Publication Publication Date Title
Kurz et al. CBCT correction using a cycle-consistent generative adversarial network and unpaired training to enable photon and proton dose calculation
US10964071B2 (en) Systems and methods for data and model-driven image reconstruction and enhancement
CN113256753B (zh) 基于多任务学习约束的pet图像感兴趣区域增强重建方法
Tiago et al. A data augmentation pipeline to generate synthetic labeled datasets of 3D echocardiography images using a GAN
US11915401B2 (en) Apriori guidance network for multitask medical image synthesis
CN111105472B (zh) Pet图像的衰减校正方法、装置、计算机设备和存储介质
WO2022120731A1 (zh) 基于循环生成对抗网络的mri-pet图像模态转换方法及系统
WO2020113148A1 (en) Single or a few views computed tomography imaging with deep neural network
US20200043170A1 (en) System and method for image segmentation using a joint deep learning model
Cui et al. 3D point-based multi-modal context clusters GAN for low-dose PET image denoising
CN118864288A (zh) 无监督伪ct对抗扩散模型构建方法、系统、介质及设备
CN112819914A (zh) 一种pet图像处理方法
CN114332271A (zh) 基于静态pet图像的动态参数图像合成方法、系统
Hu et al. A joint learning framework for multisite CBCT-to-CT translation using a hybrid CNN-transformer synthesizer and a registration network
WO2023102709A1 (zh) 基于静态pet图像的动态参数图像合成方法、系统
CN116543070A (zh) 锥束ct图像散射伪影校正处理方法及装置
WO2024148577A1 (zh) 动态pet成像方法、装置、电子设备及存储介质
CN110270015A (zh) 一种基于多序列MRI的sCT生成方法
US20250069222A1 (en) Method for generating synthetic images
CN113469915A (zh) 一种基于去噪打分匹配网络的pet重建方法
CN111612762A (zh) Mri脑肿瘤图像生成方法及系统
Jud et al. Statistical respiratory models for motion estimation
Yu et al. UltraTwin: towards cardiac anatomical twin generation from multi-view 2D ultrasound
Li et al. Unpaired medical image translation based on WaveUNIT
CN120931487B (zh) 一种医学图像超分辨率重建方法及系统

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: 21966637

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21966637

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 21966637

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 09/04/2025)