WO2020118829A1 - 基于决策树的pet图像超分辨重建方法、装置、设备及介质 - Google Patents

基于决策树的pet图像超分辨重建方法、装置、设备及介质 Download PDF

Info

Publication number
WO2020118829A1
WO2020118829A1 PCT/CN2019/071109 CN2019071109W WO2020118829A1 WO 2020118829 A1 WO2020118829 A1 WO 2020118829A1 CN 2019071109 W CN2019071109 W CN 2019071109W WO 2020118829 A1 WO2020118829 A1 WO 2020118829A1
Authority
WO
WIPO (PCT)
Prior art keywords
image block
decision tree
resolution
image
training
Prior art date
Application number
PCT/CN2019/071109
Other languages
English (en)
French (fr)
Inventor
胡战利
杨永峰
汪影
梁栋
刘新
郑海荣
Original Assignee
深圳先进技术研究院
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 深圳先进技术研究院 filed Critical 深圳先进技术研究院
Publication of WO2020118829A1 publication Critical patent/WO2020118829A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting

Definitions

  • the invention belongs to the technical field of medical imaging, and particularly relates to a method, device, equipment and medium for PET image super-resolution reconstruction based on a decision tree.
  • PET Positron emission tomography
  • ET emission imaging technology
  • PET technology is a new type of imaging technology applied to the clinic after Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), and its clinical value is unquestionable.
  • CT Computed Tomography
  • MRI Magnetic Resonance Imaging
  • High-quality PET images can improve the diagnostic accuracy of doctors, and an important indicator to measure image quality is image resolution. Therefore, in the field of image processing, improving image resolution has become an important goal in the industry.
  • Image resolution describes the number of pixels contained in an image. More pixels means that the image contains more information and the image can carry more scene details. Such an image has higher application value.
  • Super-Resolution super-resolution
  • the reconstruction-based method is to obtain high-resolution images by adding a priori knowledge.
  • the resolution of images reconstructed by such methods is not significantly improved, especially for feature structures.
  • Inconsistent PET images but the sample set required by the learning-based method is too large, and the reconstruction takes a long time. It is also necessary to manually segment the cluster image sample set manually, and the clustering features are not completely consistent. Therefore, the reconstruction of the PET image The effect is not ideal.
  • the object of the present invention is to provide a method, device, equipment and medium for PET image super-resolution reconstruction based on a decision tree, aiming to solve the problem that the existing technology cannot provide an effective PET image super-resolution reconstruction method, resulting in the reconstructed PET The problem of poor image quality of the image.
  • the present invention provides a method for super-resolution reconstruction of PET images based on decision trees.
  • the method includes the following steps:
  • the high-resolution PET image corresponding to the target PET image is reconstructed according to the high-resolution image block.
  • the present invention provides a decision tree-based PET image super-resolution reconstruction device, which includes:
  • the image block extraction unit is used for extracting the image block of the target PET image to obtain the corresponding PET image block when the super-resolution reconstruction request of the target PET image is received;
  • the mapping matrix obtaining unit is used to cluster the PET image blocks with similar feature clustering of the PET image blocks through each decision tree in the pre-trained random forest, and obtain the PET image blocks to make decisions in each decision tree ,
  • An image block prediction unit for predicting a high-resolution image block corresponding to the PET image block according to the corresponding image block mapping matrix determined by each decision tree;
  • the image reconstruction unit is configured to reconstruct a high-resolution PET image corresponding to the target PET image according to the high-resolution image block.
  • the present invention also provides an image processing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, when the processor executes the computer program.
  • the present invention also provides a computer-readable storage medium that stores a computer program, and when the computer program is executed by a processor, super-resolution reconstruction of the PET image based on the decision tree as described above is realized The steps described in the method.
  • the present invention performs image block extraction on the target PET image to be super-resolution reconstructed to obtain the corresponding PET image block, and clusters the PET image block with similar feature of the image block through each decision tree in the random forest to obtain the PET image block in
  • the corresponding image block mapping matrix determined in each decision tree, the high-resolution image block corresponding to the PET image block is predicted according to the corresponding image block mapping matrix decided in each decision tree, and reconstructed according to the high-resolution image block
  • the high-resolution PET image corresponding to the target PET image is obtained, which improves the accuracy of super-resolution reconstruction of the target PET image, and at the same time improves the resolution and clarity of the high-resolution PET image obtained by reconstruction, and thus improves the The accuracy of clinical diagnosis and treatment of PET images.
  • FIG. 1 is an implementation flowchart of a decision tree-based super-resolution reconstruction method for PET images provided in Embodiment 1 of the present invention
  • FIG. 2 is a flowchart of an implementation of a super-resolution reconstruction method for PET images based on a decision tree according to Embodiment 2 of the present invention
  • FIG. 3 is a schematic structural diagram of a PET-based super-resolution reconstruction device based on a decision tree provided by Embodiment 3 of the present invention
  • FIG. 4 is a schematic structural view of a decision tree-based PET image super-resolution reconstruction device provided in Embodiment 4 of the present invention.
  • Embodiment 5 is a schematic structural diagram of an image processing device according to Embodiment 5 of the present invention.
  • FIG. 1 shows an implementation process of a decision tree-based super-resolution reconstruction method for PET images provided in Embodiment 1 of the present invention. For convenience of description, only the parts related to the embodiment of the present invention are shown. The details are as follows:
  • step S101 when the super-resolution reconstruction request of the target PET image is received, the target PET image is subjected to image block extraction to obtain the corresponding PET image block.
  • the embodiments of the present invention are applicable to medical image processing platforms, systems or devices, such as personal computers and servers.
  • the target PET image can be collected on-site by the PET imaging device or downloaded from a pre-established medical database.
  • the image collected by the PET imaging device will be more or less noisy, affecting the quality of the target PET image, and the network download will inevitably
  • the original PET image is compressed to reduce the image volume, resulting in a reduction in the image resolution of the target PET image.
  • the target PET image needs to be super-resolution reconstructed.
  • the PET image is subjected to image block extraction based on the image size of the target PET image and the preset number of image blocks to obtain the corresponding number of PET image blocks.
  • step S102 the PET image blocks are clustered with similar feature of the PET image blocks through each decision tree in the pre-trained random forest to obtain the corresponding image block mapping matrix determined by the PET image blocks in each decision tree .
  • PET image blocks are input into each decision tree of a pre-trained random forest to cluster image block similar features, and each decision tree decides the corresponding image block mapping matrix according to the PET image block , That is, starting from the root node of the decision tree, the PET image blocks are recursively allocated to the nodes in the decision tree until reaching the leaf node of the decision tree, the image block mapping matrix corresponding to the leaf node is the PET image block in the decision The image block mapping matrix decided in the tree.
  • the random forest is composed of multiple independent decision trees.
  • the decision tree is composed of nodes, including root nodes, non-leaf nodes, and leaf nodes, and non-leaf nodes include left and right child nodes.
  • the number of decision trees in the random forest is set to 10, thereby improving the accuracy of the random forest decision result.
  • step S103 the high-resolution image block corresponding to the PET image block is predicted according to the corresponding image block mapping matrix decided by each decision tree.
  • W l is the image block mapping matrix of the PET image block classified into the lth leaf node in the decision tree
  • y is the PET image block
  • T is the decision tree in the random forest
  • the number, l(t) is the PET image block y is split into the lth leaf node of the tth decision tree, It is the predicted high-resolution image block.
  • step S104 a high-resolution PET image corresponding to the target PET image is reconstructed according to the high-resolution image block.
  • all predicted high-resolution image blocks Stacking makes the image dimension larger, and then reconstructs the high-resolution PET image corresponding to the target PET image.
  • the target PET image to be super-resolution-reconstructed is subjected to image block extraction to obtain the corresponding PET image block, and the image block similar feature clustering is performed on the PET image block through each decision tree in the random forest.
  • the resolution image block reconstructs the high-resolution PET image corresponding to the target PET image, thereby improving the accuracy of super-resolution reconstruction of the target PET image, and at the same time improving the resolution and clarity of the reconstructed high-resolution PET image , And then improve the accuracy of clinical diagnosis and treatment based on PET images.
  • FIG. 2 shows an implementation process of a decision tree-based super-resolution reconstruction method for PET images provided in Embodiment 2 of the present invention.
  • FIG. 2 shows an implementation process of a decision tree-based super-resolution reconstruction method for PET images provided in Embodiment 2 of the present invention.
  • the details are as follows:
  • step S201 recursive splitting is started from the root node of the decision tree according to the preset training samples and the preset splitting function, until the splitting depth reaches the preset decision tree depth threshold, the construction of the decision tree is ended to achieve Image tree similar feature clustering training of decision tree.
  • the samples satisfying the left splitting result in the training samples are allocated to the left child node, satisfying the right
  • the samples of the split result are distributed to the right child node, and then the left and right child node samples corresponding to the left and right child nodes are split again according to the split result of the split function, and then the recursive split is performed in turn until the sub training samples corresponding to the child nodes are inseparable , That is, when the leaf node is generated or the splitting depth reaches the preset decision tree depth threshold, the decision tree stops growing, that is, the construction of the decision tree is ended, and the left and right child nodes of the recursive splitting are split clustering all training samples, which is completed The clustering training of image block similarity features of decision tree is introduced.
  • the training process of all decision trees in the random forest is independent of each other, thereby improving the stability of the random forest model composed of decision trees.
  • the training samples of the decision tree in the random forest are selected by resampling, that is, the decision tree has repeated samples.
  • the number of training samples of all decision trees is the same, and the training samples are composed of high-resolution training image blocks and the The training image block pair composed of the low-resolution training image block corresponding to the high-resolution training image block, thereby improving the training time of the decision tree.
  • the high-resolution PET training image is degraded by a preset image degradation formula to obtain the corresponding PET Low-resolution training images, extracting image blocks from PET high-resolution training images and PET low-resolution training images, respectively, to obtain corresponding high-resolution training image blocks and low-resolution training image blocks, and high-resolution training image blocks
  • a training image block pair set composed of a low-resolution training image block corresponding to the high-resolution training image block is set as a training sample, thereby improving the accuracy of subsequent decision tree prediction results.
  • the splitting function is The low-resolution training image block y whose split function value is 0 is split to the left child node, and the high-resolution training image block corresponding to the low-resolution training image block is also split to the left child node, expressed as Conversely, it is split into the right child node
  • r ⁇ (y) y[ ⁇ ]- ⁇ th
  • ⁇ is the feature of the low-resolution training image block y in the training sample
  • r ⁇ (y) is the response function of the feature ⁇
  • y[ ⁇ ] is the y
  • the one-dimensional vector of the feature matrix ⁇ th is the preset threshold
  • ⁇ H,L ⁇ is expressed as a pair of training image blocks composed of the high-resolution training image block H and the corresponding low-resolution training image block L, thereby improving the training sample
  • the recursive splitting is implemented through the following steps:
  • the image block features of the current node training sample corresponding to the current node are traversed and extracted until the extracted image block features meet the preset feature selection Formula, and then select the image block feature as a node split feature, where the current node training sample belongs to the training sample.
  • the feature selection formula is To determine whether the extracted feature ⁇ is the optimal feature according to the feature selection formula, that is, to minimize the calculation error according to the feature ⁇ , where, Is the mean value of the nth training sample y n (that is, the mean value of pixels in the same spatial position), k is the pre-set hyperparameter, N is the number of training samples, and X C and Y C are split into the child nodes of the decision tree High-resolution training image block and low-resolution training image block, m(y n ) is the prediction result of the decision tree to predict y n , x n is the high-resolution image block corresponding to the training sample y n to be reconstructed, Le For the left child node, Ri is the right child node.
  • the splitting function splits the left and right child nodes of the current node training sample according to the selected and optimal node splitting feature.
  • the current node of the decision tree is split, thereby improving the similarity of the clustered image block features in the training sample.
  • the step of extracting image block features from the low-resolution training image block in the current node training sample corresponding to the current node of the decision tree to be split preferably, when the depth of the current node reaches a preset decision tree depth threshold, Or when the number of low-resolution training image blocks in the current node training sample is less than the preset threshold of the number of image blocks (for example, one for each of the high-resolution and low-resolution image blocks), stop splitting the left and right sub-nodes of the current node training sample.
  • the current node is set as the leaf node of the decision tree, and the image block mapping matrix is learned through the current node training samples, thereby reducing the probability of overfitting of the decision tree.
  • the decision tree depth threshold is set to 15, so as to reduce the probability of overfitting of the decision tree.
  • step S202 when the super-resolution reconstruction request of the target PET image is received, the target PET image is subjected to image block extraction to obtain the corresponding PET image block.
  • step S203 the PET image blocks are clustered with similar feature of the PET image blocks through each decision tree in the pre-trained random forest to obtain the corresponding image block mapping matrix determined by the PET image blocks in each decision tree .
  • step S204 the high-resolution image block corresponding to the PET image block is predicted according to the corresponding image block mapping matrix decided by each decision tree.
  • step S205 a high-resolution PET image corresponding to the target PET image is reconstructed according to the high-resolution image block.
  • step S202-step S205 for the specific implementation of step S202-step S205, reference may be made to the description of step S101-step S104 in Embodiment 1, and details are not described herein again.
  • recursive splitting is started from the root node of the decision tree according to a preset training sample and a preset splitting function, until the splitting depth reaches a preset decision tree depth threshold to complete the decision tree
  • Build and implement clustering training of similar features of image blocks of decision tree and then super-reconstruct the target PET image according to the trained random forest to obtain the corresponding high-resolution PET image, thereby improving the training time of the decision tree, and
  • the stability of the random forest model composed of decision trees is improved, which in turn improves the resolution and clarity of the high-resolution PET images reconstructed.
  • FIG. 3 shows the structure of a decision tree-based PET image super-resolution reconstruction device provided in Embodiment 3 of the present invention. For ease of description, only parts related to the embodiment of the present invention are shown, including:
  • the image block extraction unit 31 is configured to extract the image block of the target PET image to obtain the corresponding PET image block when the super-resolution reconstruction request of the target PET image is received;
  • the mapping matrix obtaining unit 32 is used to cluster the image blocks with similar features of the PET image blocks through each decision tree in the pre-trained random forest to obtain the corresponding images of the PET image blocks decided in each decision tree Block mapping matrix;
  • the image block prediction unit 33 is configured to predict the high-resolution image block corresponding to the PET image block according to the corresponding image block mapping matrix decided by each decision tree;
  • the image reconstruction unit 34 is configured to reconstruct a high-resolution PET image corresponding to the target PET image according to the high-resolution image block.
  • each unit of the PET tree super-resolution reconstruction device based on the decision tree may be implemented by a corresponding hardware or software unit, and each unit may be an independent software and hardware unit, or may be integrated into one software and hardware unit. It is not used here to limit the invention. Specifically, for the implementation of each unit, reference may be made to the foregoing description of Embodiment 1, and details are not described herein again.
  • FIG. 4 shows the structure of a decision tree-based super-resolution reconstruction device for PET images provided in Embodiment 4 of the present invention. For ease of explanation, only parts related to the embodiment of the present invention are shown, including:
  • the decision tree construction unit 41 is used to perform recursive splitting from the root node of the decision tree according to the preset training samples and the preset split function, until the splitting depth reaches the preset decision tree depth threshold, the construction of the decision tree is ended To achieve clustering training of image block similarity features of decision tree;
  • the image block extraction unit 42 is configured to extract the image block of the target PET image to obtain the corresponding PET image block when the super-resolution reconstruction request of the target PET image is received;
  • the mapping matrix obtaining unit 43 is used for clustering the image block similar features of the PET image blocks through each decision tree in the pre-trained random forest to obtain the corresponding images determined by the PET image blocks in each decision tree Block mapping matrix;
  • the image block prediction unit 44 is used to predict the high-resolution image block corresponding to the PET image block according to the corresponding image block mapping matrix decided by each decision tree;
  • the image reconstruction unit 45 is configured to reconstruct a high-resolution PET image corresponding to the target PET image according to the high-resolution image block.
  • the decision tree construction unit 41 includes:
  • the feature selection unit 411 is used to extract image block features from the low-resolution training image blocks in the current node training samples corresponding to the current node to be split in the decision tree until the extracted image block features meet the preset feature selection formula, Select the image block feature as the node split feature, the current node training sample belongs to the training sample;
  • the node splitting unit 412 is used to split the left and right child nodes of the current node training sample according to the node splitting feature and the splitting function;
  • the node split stop unit 413 is used to stop the training sample of the current node when the depth of the current node reaches the depth threshold of the decision tree, or the number of low-resolution training image blocks in the current node training sample is less than the preset threshold value of the number of image blocks Split left and right child nodes;
  • the mapping matrix learning unit 414 is configured to set the current node as a leaf node of the decision tree, and learn the image block mapping matrix through the current node training samples.
  • each unit of the PET tree super-resolution reconstruction device based on the decision tree may be implemented by a corresponding hardware or software unit, and each unit may be an independent software and hardware unit, or may be integrated into one software and hardware unit. It is not used here to limit the invention. Specifically, for the implementation of each unit, reference may be made to the description of the foregoing method embodiments, and details are not described herein again.
  • FIG. 5 shows the structure of the image processing apparatus provided in Embodiment 5 of the present invention. For convenience of description, only parts related to the embodiment of the present invention are shown.
  • the image processing apparatus 5 of the embodiment of the present invention includes a processor 50, a memory 51, and a computer program 52 stored in the memory 51 and executable on the processor 50.
  • the processor 50 executes the computer program 52, the steps in the embodiment of the above-described decision tree-based PET image super-resolution reconstruction method are implemented, for example, steps S101 to S104 shown in FIG. 1.
  • the processor 50 executes the computer program 52, the functions of the units in the foregoing device embodiments are realized, for example, the functions of the units 31 to 34 shown in FIG.
  • the target PET image to be super-resolution-reconstructed is subjected to image block extraction to obtain the corresponding PET image block, and the image block similar feature clustering is performed on the PET image block through each decision tree in the random forest.
  • the resolution image block reconstructs the high-resolution PET image corresponding to the target PET image, thereby improving the accuracy of super-resolution reconstruction of the target PET image, and at the same time improving the resolution and clarity of the reconstructed high-resolution PET image , And then improve the accuracy of clinical diagnosis and treatment based on PET images.
  • the image processing device in the embodiment of the present invention may be a personal computer or a server.
  • the processor 50 in the image processing device 5 executes the computer program 52 to realize the PET tree super-resolution reconstruction method based on the decision tree, reference may be made to the description of the foregoing method embodiments, which will not be repeated here.
  • a computer-readable storage medium stores a computer program that, when executed by a processor, implements the above-described decision tree-based PET image super-resolution reconstruction method embodiment
  • the computer program when executed by the processor, the functions of the units in the foregoing device embodiments are realized, for example, the functions of the units 31 to 34 shown in FIG. 3.
  • the target PET image to be super-resolution-reconstructed is subjected to image block extraction to obtain the corresponding PET image block, and the image block similar feature clustering is performed on the PET image block through each decision tree in the random forest.
  • the resolution image block reconstructs the high-resolution PET image corresponding to the target PET image, thereby improving the accuracy of super-resolution reconstruction of the target PET image, and at the same time improving the resolution and clarity of the reconstructed high-resolution PET image , And then improve the accuracy of clinical diagnosis and treatment based on PET images.
  • the computer-readable storage medium in the embodiments of the present invention may include any entity or device capable of carrying computer program code, and a recording medium, such as ROM/RAM, magnetic disk, optical disk, flash memory, and other memories.

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

一种基于决策树的PET图像超分辨重建方法、装置、设备及介质，该方法包括：对待进行超分辨率重建的目标PET图像进行图像块提取，得到对应的PET图像块，通过随机森林中的每棵决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每棵决策树中决策出的、对应的图像块映射矩阵，根据每棵决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块，根据高分辨率图像块重建出目标PET图像对应的高分辨率PET图像，从而提高了对目标PET图像进行超分辨率重建的精确度，同时提高了重建得到的高分辨率PET图像的分辨率和清晰度，进而提高了基于PET图像的临床诊疗的准确性和精确度。

Description

基于决策树的PET图像超分辨重建方法、装置、设备及介质 技术领域
本发明属于医学成像技术领域，尤其涉及一种基于决策树的PET图像超分辨重建方法、装置、设备及介质。
背景技术
正电子发射断层成像(Positron Emission Tomography，简称PET)是一种发射型成像技术(Emission Tomography，简称ET)，它通过把放射性药物注入体内的方法来显示不同组织的新陈代谢情况。PET技术是继计算机断层成像(Computed Tomography，简称CT)和磁共振成像(Magnetic Resonance Imaging，简称MRI)之后应用于临床的一种新型影像技术，其临床价值是毋庸置疑的。高质量的PET图像可以提高医生的诊断精确度，而衡量图像质量的一个重要指标是图像分辨率，因此，在图像处理领域里，提高图像分辨率成了业界一个重要目标。
图像分辨率描述的是图像包含的像素点数，更多的像素意味着图像包含的信息量更大，图像可以承载更多的场景细节，这样的图像有着更高的应用价值。在20世纪60年代，超分辨率(Super‐Resolution)重建技术的概念首次出现，这些技术主张从一幅图像或者序列图像帧中恢复高分辨率图像，在Tasi和Huang的积极推动下得到迅速而广泛的发展。在医学成像领域，例如PET成像，超分辨率重建算法可以弥补PET物理成像设备的局限，增强PET图像细节信息，提高医学影像的质量，从而辅助医生更好的对病情进行检测和分析，以做出正确的诊断，具有广泛的重要的应用价值。
然而，在现有的超分辨率重建算法中，基于重建的方法是通过加入先验知识来得到高分辨率图像，这类方法重建得到的图像分辨率提高的不太明显，特 别是对于特征结构不一致的PET图像，而基于学习的方法所需样本集过大，重建耗时较长，还需人工手动分割聚类图像样本集，聚类特征也不完全一致，因此，重建得到的PET图像的效果也不太理想。
发明内容
本发明的目的在于提供一种基于决策树的PET图像超分辨重建方法、装置、设备及介质，旨在解决由于现有技术无法提供一种有效的PET图像超分辨重建方法，导致重建得到的PET图像的图像质量不佳的问题。
一方面，本发明提供了一种基于决策树的PET图像超分辨重建方法，所述方法包括下述步骤：
当接收到目标PET图像的超分辨率重建请求时，将所述目标PET图像进行图像块提取，得到对应的PET图像块；
通过预先训练好的随机森林中的每颗决策树对所述PET图像块进行图像块相似特征聚类，获得所述PET图像块在所述每颗决策树中决策出的、对应的图像块映射矩阵，其中，所述随机森林由多颗独立的决策树组成；
根据所述每颗决策树决策出的、对应的所述图像块映射矩阵预测所述PET图像块对应的高分辨率图像块；
根据所述高分辨率图像块重建出所述目标PET图像对应的高分辨率PET图像。
另一方面，本发明提供了一种基于决策树的PET图像超分辨重建装置，所述装置包括：
图像块提取单元，用于当接收到目标PET图像的超分辨率重建请求时，将所述目标PET图像进行图像块提取，得到对应的PET图像块；
映射矩阵获得单元，用于通过预先训练好的随机森林中的每颗决策树对所述PET图像块进行图像块相似特征聚类，获得所述PET图像块在所述每颗决策树中决策出的、对应的图像块映射矩阵，其中，所述随机森林由多颗独立的决 策树组成；
图像块预测单元，用于根据所述每颗决策树决策出的、对应的所述图像块映射矩阵预测所述PET图像块对应的高分辨率图像块；以及
图像重建单元，用于根据所述高分辨率图像块重建出所述目标PET图像对应的高分辨率PET图像。
另一方面，本发明还提供了一种图像处理设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如上述基于决策树的PET图像超分辨重建方法所述的步骤。
另一方面，本发明还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现如上述基于决策树的PET图像超分辨重建方法所述的步骤。
本发明对待进行超分辨率重建的目标PET图像进行图像块提取，得到对应的PET图像块，通过随机森林中的每颗决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每颗决策树中决策出的、对应的图像块映射矩阵，根据每颗决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块，根据高分辨率图像块重建出目标PET图像对应的高分辨率PET图像，从而提高了对目标PET图像进行超分辨率重建的精确度，同时提高了重建得到的高分辨率PET图像的分辨率和清晰度，进而提高了基于PET图像的临床诊疗的准确性。
附图说明
图1是本发明实施例一提供的基于决策树的PET图像超分辨重建方法的实现流程图；
图2是本发明实施例二提供的基于决策树的PET图像超分辨重建方法的实现流程图；
图3是本发明实施例三提供的基于决策树的PET图像超分辨重建装置的结构示意图；
图4是本发明实施例四提供的基于决策树的PET图像超分辨重建装置的结构示意图；以及
图5是本发明实施例五提供的图像处理设备的结构示意图。
具体实施方式
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明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仅用以解释本发明，并不用于限定本发明。
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示出了本发明实施例一提供的基于决策树的PET图像超分辨重建方法的实现流程，为了便于说明，仅示出了与本发明实施例相关的部分，详述如下：
在步骤S101中，当接收到目标PET图像的超分辨率重建请求时，将目标PET图像进行图像块提取，得到对应的PET图像块。
本发明实施例适用于医学图像处理平台、系统或设备，例如个人计算机、服务器等。目标PET图像可以是通过PET成像设备现场采集的或者从预先建立的医学资料库中下载的，PET成像设备采集的图像或多或少都会出现噪声，影响目标PET图像质量，而网络下载难免会对原始PET图像进行压缩以缩小图像体积，从而造成目标PET图像的图像分辨率降低，为了不影响诊治，需要对目标PET图像进行超分辨率重建。当接收到目标PET图像的超分辨率重建请求时，根据目标PET图像的图像尺寸和预设的图像块数量，对PET图像进行图像块提取，得到对应的数量个PET图像块。
在步骤S102中，通过预先训练好的随机森林中的每颗决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每颗决策树中决策出的、对应的 图像块映射矩阵。
在本发明实施例中，将PET图像块分别输入到预先训练好的随机森林的每颗决策树中进行图像块相似特征聚类，每颗决策树根据PET图像块决策出对应的图像块映射矩阵，即从决策树的根节点开始，将PET图像块递归地分配到决策树中的节点，直到到达该决策树的叶子节点，该叶子节点对应的图像块映射矩阵即为PET图像块在该决策树中决策出的图像块映射矩阵。其中，随机森林由多颗独立的决策树组成，决策树由各节点组成，包括根节点、非叶子节点、以及叶子节点，非叶子节点又包括左、右子节点。
在本发明实施例中，优选地，将随机森林中的决策树的数量设置为10，从而提高随机森林决策结果的准确性。
在步骤S103中，根据每颗决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块。
在本发明实施例中，根据预设的叶子节点映射模型m l(y)＝W l·φ(y)和每颗决策树决策出的、对应的图像块映射矩阵对PET图像块进行预测，获得PET图像块在每颗决策树中预测得到的、对应的预测结果(该预测结果为初始高分辨率图像块)，再根据模型
Figure PCTCN2019071109-appb-000001
综合所有决策树的预测结果，即对所有决策树的预测结果取平均，得到最终的高分辨率图像块
Figure PCTCN2019071109-appb-000002
其中，W l为PET图像块分类到决策树中第l个叶子节点的图像块映射矩阵，φ(y)＝y为预设的基函数，y为PET图像块，T为随机森林中决策树的数量，l(t)为PET图像块y被分裂到第t棵决策树的第l个叶子节点，
Figure PCTCN2019071109-appb-000003
是预测得到的高分辨率图像块。
在步骤S104中，根据高分辨率图像块重建出目标PET图像对应的高分辨率PET图像。
在本发明实施例中，将预测得到的所有高分辨率图像块
Figure PCTCN2019071109-appb-000004
进行堆叠，使得图像维度变大，进而重建出目标PET图像对应的高分辨率PET图像。
在本发明实施例中，对待进行超分辨率重建的目标PET图像进行图像块提 取，得到对应的PET图像块，通过随机森林中的每颗决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每颗决策树中决策出的、对应的图像块映射矩阵，根据每颗决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块，根据高分辨率图像块重建出目标PET图像对应的高分辨率PET图像，从而提高了对目标PET图像进行超分辨率重建的精确度，同时提高了重建得到的高分辨率PET图像的分辨率和清晰度，进而提高了基于PET图像的临床诊疗的准确性。
实施例二：
图2示出了本发明实施例二提供的基于决策树的PET图像超分辨重建方法的实现流程，为了便于说明，仅示出了与本发明实施例相关的部分，详述如下：
在步骤S201中，根据预设的训练样本和预设的分裂函数从决策树的根节点开始进行递归分裂，直至分裂的深度达到预设的决策树深度阈值时，结束决策树的构建，以实现决策树的图像块相似特征聚类训练。
在本发明实施例中，从决策树的第一个节点(即根节点)开始，根据预设的分裂函数的分裂结果将训练样本中满足左分裂结果的样本分配到左子节点中，满足右分裂结果的样本分配到右子节点中，然后再分别对左右子节点对应的左右子节点样本按照分裂函数的分裂结果进行再次分裂，依此依次进行递归分裂，直至子节点对应的子训练样本不可分，即产生叶子节点或者分裂的深度达到预设的决策树深度阈值时，决策树就停止生长，即结束决策树的构建，同时递归分裂的左右子节点即分裂聚类所有训练样本，也即完成了决策树的图像块相似特征聚类训练。随机森林中所有决策树训练过程互相独立，从而提高了由决策树组成的随机森林模型的稳定性。随机森林中决策树的训练样本采用重采样方式进行抽取，即决策树存在重复样本的情况，所有决策树的训练样本的数量是相同的，且训练样本是由高分辨率训练图像块和与该高分辨率训练图像块对应的低分辨率训练图像块组成的训练图像块对组成的，从而提高了决策树的训练时间。
在根据预设的训练样本和预设的分裂函数从决策树的根节点开始进行递归分裂之前，优选地，通过预设的图像退化公式将PET高分辨率训练图像进行退化处理，得到对应的PET低分辨率训练图像，对PET高分辨率训练图像和PET低分辨率训练图像分别进行图像块提取，得到对应的高分辨率训练图像块和低分辨率训练图像块，将高分辨率训练图像块和与该高分辨率训练图像块对应的低分辨率训练图像块组成的训练图像块对集合设置为训练样本，从而提高后续决策树预测结果的准确性。
在本发明实施例中，优选地，分裂函数为
Figure PCTCN2019071109-appb-000005
分裂函数值为0的低分辨率训练图像块y被分裂到左子节点，低分辨率训练图像块对应的高分辨率训练图像块也被分裂到左子节点，表示为
Figure PCTCN2019071109-appb-000006
反之被分裂到右子节点，表示为
Figure PCTCN2019071109-appb-000007
其中，r θ(y)＝y[θ]-θ th，θ是训练样本中低分辨率训练图像块y的特征，r θ(y)是特征θ的响应函数，y[θ]是y的特征矩阵的一维向量，θ th是预设阈值，{H,L}表示为高分辨率训练图像块H和与其对应的低分辨率训练图像块L组成的训练图像块对，从而提高训练样本中聚类的图像块特征的相似性。
在根据预设的训练样本和预设的分裂函数从决策树的根节点开始进行递归分裂时，优选地，通过下述步骤实现递归分裂：
1)对决策树待分裂的当前节点对应的当前节点训练样本中的低分辨率训练图像块进行图像块特征提取，直至提取到的图像块特征满足预设的特征选取公式，选取图像块特征作为节点分裂特征，其中，当前节点训练样本属于训练样本。
在本发明实施例中，当需要对决策树的当前节点进行分裂时，先对当前节点对应的当前节点训练样本的图像块特征进行遍历提取，直至提取到的图像块特征满足预设的特征选取公式，再选取该图像块特征作为节点分裂特征，其中， 当前节点训练样本属于训练样本。
在本发明实施例中，优选地，特征选取公式为
Figure PCTCN2019071109-appb-000008
以根据该特征选取公式判断提取到的特征θ是不是最优特征，即根据该特征θ使得计算误差最小，其中，
Figure PCTCN2019071109-appb-000009
是第n个训练样本y n的均值(即同一个空间位置上的像素均值)，k是预先设置的超参数，N是训练样本的数量，X C、Y C分裂到决策树中子节点的高分辨率训练图像块和低分辨率训练图像块，m(y n)是决策树对y n进行预测的预测结果，x n是待重建的训练样本y n对应的高分辨率图像块，Le为左子节点，Ri为右子节点。
2)根据节点分裂特征和分裂函数对当前节点训练样本进行左、右子节点分裂。
在本发明实施例中，分裂函数根据选取的、最优的节点分裂特征对当前节点训练样本进行左、右子节点分裂。
通过上述步骤1)和2)实现决策树当前节点的分裂，从而提高训练样本中聚类的图像块特征的相似性。
在对决策树待分裂的当前节点对应的当前节点训练样本中的低分辨率训练图像块进行图像块特征提取的步骤之前，优选地，当当前节点所在的深度达到预设的决策树深度阈值，或者当前节点训练样本中低分辨率训练图像块的数量小于预设的图像块数量阈值(例如，高、低分辨率图像块各一块)时，停止对当前节点训练样本进行左右子节点分裂，将当前节点设置为决策树的叶子节点，并通过当前节点训练样本学习图像块映射矩阵，从而降低决策树过拟合的概率。
在本发明实施例中，优选地，将决策树深度阈值设置为15，从而降低决策树过拟合的概率。
在将当前节点设置为决策树的叶子节点，并通过当前节点训练样本学习图像块映射矩阵时，优选地，在决策树的每一个叶子节点，根据叶子节点对应的 当前节点训练样本，使用公式
Figure PCTCN2019071109-appb-000010
学习叶子节点对应的图像块映射矩阵，其中，W l为叶子节点对应的图像块映射矩阵，I是预设的单位矩阵，λ是预设的正则化参数(例如，0.01)，φ(Y)为预设的基函数，Y为低分辨率训练图像块，从而使得学习到的图像块映射矩阵能更准确的反映高低分辨率训练图像块之间的映射关系。
在步骤S202中，当接收到目标PET图像的超分辨率重建请求时，将目标PET图像进行图像块提取，得到对应的PET图像块。
在步骤S203中，通过预先训练好的随机森林中的每颗决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每颗决策树中决策出的、对应的图像块映射矩阵。
在步骤S204中，根据每颗决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块。
在步骤S205中，根据高分辨率图像块重建出目标PET图像对应的高分辨率PET图像。
在本发明实施例中，步骤S202‐步骤S205的具体实施方式可参考实施例一的步骤S101‐步骤S104的描述，在此不再赘述。
在本发明实施例中，首先，根据预设的训练样本和预设的分裂函数从决策树的根节点开始进行递归分裂，直至分裂的深度达到预设的决策树深度阈值，以完成决策树的构建，实现决策树的图像块相似特征聚类训练，再根据训练好的随机森林对目标PET图像进行超分辨率重建，得到对应的高分辨率PET图像，从而提高了决策树的训练时间，以及提高了由决策树组成的随机森林模型的稳定性，进而提高了重建得到的高分辨率PET图像的分辨率和清晰度。
实施例三：
图3示出了本发明实施例三提供的基于决策树的PET图像超分辨重建装置的结构，为了便于说明，仅示出了与本发明实施例相关的部分，其中包括：
图像块提取单元31，用于当接收到目标PET图像的超分辨率重建请求时， 将目标PET图像进行图像块提取，得到对应的PET图像块；
映射矩阵获得单元32，用于通过预先训练好的随机森林中的每颗决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每颗决策树中决策出的、对应的图像块映射矩阵；
图像块预测单元33，用于根据每颗决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块；以及
图像重建单元34，用于根据高分辨率图像块重建出目标PET图像对应的高分辨率PET图像。
在本发明实施例中，基于决策树的PET图像超分辨重建装置的各单元可由相应的硬件或软件单元实现，各单元可以为独立的软、硬件单元，也可以集成为一个软、硬件单元，在此不用以限制本发明。具体地，各单元的实施方式可参考前述实施例一的描述，在此不再赘述。
实施例四：
图4示出了本发明实施例四提供的基于决策树的PET图像超分辨重建装置的结构，为了便于说明，仅示出了与本发明实施例相关的部分，其中包括：
决策树构建单元41，用于根据预设的训练样本和预设的分裂函数从决策树的根节点开始进行递归分裂，直至分裂的深度达到预设的决策树深度阈值时，结束决策树的构建，以实现决策树的图像块相似特征聚类训练；
图像块提取单元42，用于当接收到目标PET图像的超分辨率重建请求时，将目标PET图像进行图像块提取，得到对应的PET图像块；
映射矩阵获得单元43，用于通过预先训练好的随机森林中的每颗决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每颗决策树中决策出的、对应的图像块映射矩阵；
图像块预测单元44，用于根据每颗决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块；以及
图像重建单元45，用于根据高分辨率图像块重建出目标PET图像对应的高 分辨率PET图像。
其中，优选地，决策树构建单元41包括：
特征选取单元411，用于对决策树待分裂的当前节点对应的当前节点训练样本中的低分辨率训练图像块进行图像块特征提取，直至提取到的图像块特征满足预设的特征选取公式，选取图像块特征作为节点分裂特征，当前节点训练样本属于训练样本；
节点分裂单元412，用于根据节点分裂特征和分裂函数对当前节点训练样本进行左、右子节点分裂；
节点分裂停止单元413，用于当当前节点所在的深度达到决策树深度阈值，或者当前节点训练样本中低分辨率训练图像块的数量小于预设的图像块数量阈值时，停止对当前节点训练样本进行左右子节点分裂；以及
映射矩阵学习单元414，用于将当前节点设置为决策树的叶子节点，并通过当前节点训练样本学习图像块映射矩阵。
在本发明实施例中，基于决策树的PET图像超分辨重建装置的各单元可由相应的硬件或软件单元实现，各单元可以为独立的软、硬件单元，也可以集成为一个软、硬件单元，在此不用以限制本发明。具体地，各单元的实施方式可参考前述方法实施例的描述，在此不再赘述。
实施例五：
图5示出了本发明实施例五提供的图像处理设备的结构，为了便于说明，仅示出了与本发明实施例相关的部分。
本发明实施例的图像处理设备5包括处理器50、存储器51以及存储在存储器51中并可在处理器50上运行的计算机程序52。该处理器50执行计算机程序52时实现上述基于决策树的PET图像超分辨重建方法实施例中的步骤，例如图1所示的步骤S101至S104。或者，处理器50执行计算机程序52时实现上述各装置实施例中各单元的功能，例如图3所示单元31至34的功能。
在本发明实施例中，对待进行超分辨率重建的目标PET图像进行图像块提 取，得到对应的PET图像块，通过随机森林中的每颗决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每颗决策树中决策出的、对应的图像块映射矩阵，根据每颗决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块，根据高分辨率图像块重建出目标PET图像对应的高分辨率PET图像，从而提高了对目标PET图像进行超分辨率重建的精确度，同时提高了重建得到的高分辨率PET图像的分辨率和清晰度，进而提高了基于PET图像的临床诊疗的准确性。
本发明实施例的图像处理设备可以为个人计算机、服务器。该图像处理设备5中处理器50执行计算机程序52时实现基于决策树的PET图像超分辨重建方法时实现的步骤可参考前述方法实施例的描述，在此不再赘述。
实施例六：
在本发明实施例中，提供了一种计算机可读存储介质，该计算机可读存储介质存储有计算机程序，该计算机程序被处理器执行时实现上述基于决策树的PET图像超分辨重建方法实施例中的步骤，例如，图1所示的步骤S101至S104。或者，该计算机程序被处理器执行时实现上述各装置实施例中各单元的功能，例如图3所示单元31至34的功能。
在本发明实施例中，对待进行超分辨率重建的目标PET图像进行图像块提取，得到对应的PET图像块，通过随机森林中的每颗决策树对PET图像块进行图像块相似特征聚类，获得PET图像块在每颗决策树中决策出的、对应的图像块映射矩阵，根据每颗决策树决策出的、对应的图像块映射矩阵预测PET图像块对应的高分辨率图像块，根据高分辨率图像块重建出目标PET图像对应的高分辨率PET图像，从而提高了对目标PET图像进行超分辨率重建的精确度，同时提高了重建得到的高分辨率PET图像的分辨率和清晰度，进而提高了基于PET图像的临床诊疗的准确性。
本发明实施例的计算机可读存储介质可以包括能够携带计算机程序代码的任何实体或装置、记录介质，例如，ROM/RAM、磁盘、光盘、闪存等存储器。
以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种基于决策树的PET图像超分辨重建方法，其特征在于，所述方法包括下述步骤：
    当接收到目标PET图像的超分辨率重建请求时，将所述目标PET图像进行图像块提取，得到对应的PET图像块；
    通过预先训练好的随机森林中的每颗决策树对所述PET图像块进行图像块相似特征聚类，获得所述PET图像块在所述每颗决策树中决策出的、对应的图像块映射矩阵，其中，所述随机森林由多颗独立的决策树组成；
    根据所述每颗决策树决策出的、对应的所述图像块映射矩阵预测所述PET图像块对应的高分辨率图像块；
    根据所述高分辨率图像块重建出所述目标PET图像对应的高分辨率PET图像。
  2. 如权利要求1所述的方法，其特征在于，在将所述目标PET图像进行图像块提取的步骤之前，所述方法还包括：
    根据预设的训练样本和预设的分裂函数从所述决策树的根节点开始进行递归分裂，直至所述分裂的深度达到预设的决策树深度阈值时，结束所述决策树的构建，以实现所述决策树的图像块相似特征聚类训练，其中，所述训练样本由高分辨率训练图像块和与所述高分辨率训练图像块对应的低分辨率训练图像块组成。
  3. 如权利要求2所述的方法，其特征在于，根据预设的训练样本和预设的分裂函数从所述决策树的根节点开始进行递归分裂的步骤，包括：
    对所述决策树待分裂的当前节点对应的当前节点训练样本中的低分辨率训练图像块进行图像块特征提取，直至提取到的图像块特征满足预设的特征选取公式，选取所述图像块特征作为节点分裂特征，所述当前节点训练样本属于所述训练样本；
    根据所述节点分裂特征和所述分裂函数对所述当前节点训练样本进行左、 右子节点分裂。
  4. 如权利要求3所述的方法，其特征在于，在对所述决策树待分裂的当前节点对应的当前节点训练样本中的低分辨率训练图像块进行图像块特征提取的步骤之前，所述方法还包括：
    当所述当前节点所在的深度达到预设的决策树深度阈值，或者所述当前节点训练样本中低分辨率训练图像块的数量小于预设的图像块数量阈值时，停止对所述当前节点训练样本进行左右子节点分裂；
    将所述当前节点设置为所述决策树的叶子节点，并通过所述当前节点训练样本学习所述图像块映射矩阵。
  5. 一种基于决策树的PET图像超分辨重建装置，其特征在于，所述装置包括：
    图像块提取单元，用于当接收到目标PET图像的超分辨率重建请求时，将所述目标PET图像进行图像块提取，得到对应的PET图像块；
    映射矩阵获得单元，用于通过预先训练好的随机森林中的每颗决策树对所述PET图像块进行图像块相似特征聚类，获得所述PET图像块在所述每颗决策树中决策出的、对应的图像块映射矩阵，其中，所述随机森林由多颗独立的决策树组成；
    图像块预测单元，用于根据所述每颗决策树决策出的、对应的所述图像块映射矩阵预测所述PET图像块对应的高分辨率图像块；以及
    图像重建单元，用于根据所述高分辨率图像块重建出所述目标PET图像对应的高分辨率PET图像。
  6. 如权利要求5所述的装置，其特征在于，所述装置还包括：
    决策树构建单元，用于根据预设的训练样本和预设的分裂函数从所述决策树的根节点开始进行递归分裂，直至所述分裂的深度达到预设的决策树深度阈值时，结束所述决策树的构建，以实现所述决策树的图像块相似特征聚类训练，其中，所述训练样本由高分辨率训练图像块和与所述高分辨率训练图像块对应 的低分辨率训练图像块组成。
  7. 如权利要求6所述的装置，其特征在于，所述决策树构建单元包括：
    特征选取单元，用于对所述决策树待分裂的当前节点对应的当前节点训练样本中的低分辨率训练图像块进行图像块特征提取，直至提取到的图像块特征满足预设的特征选取公式，选取所述图像块特征作为节点分裂特征，所述当前节点训练样本属于所述训练样本；以及
    节点分裂单元，用于根据所述节点分裂特征和所述分裂函数对所述当前节点训练样本进行左、右子节点分裂。
  8. 如权利要求5所述的装置，其特征在于，所述装置还包括：
    节点分裂停止单元，用于当所述当前节点所在的深度达到所述决策树深度阈值，或者所述当前节点训练样本中低分辨率训练图像块的数量小于预设的图像块数量阈值时，停止对所述当前节点训练样本进行左右子节点分裂；以及
    映射矩阵学习单元，用于将所述当前节点设置为所述决策树的叶子节点，并通过所述当前节点训练样本学习所述图像块映射矩阵。
  9. 一种图像处理设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如权利要求1至4任一项所述方法的步骤。
  10. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1至4任一项所述方法的步骤。
PCT/CN2019/071109 2018-12-12 2019-01-10 基于决策树的pet图像超分辨重建方法、装置、设备及介质 WO2020118829A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811519205.5 2018-12-12
CN201811519205.5A CN109741255A (zh) 2018-12-12 2018-12-12 基于决策树的pet图像超分辨重建方法、装置、设备及介质

Publications (1)

Publication Number Publication Date
WO2020118829A1 true WO2020118829A1 (zh) 2020-06-18

Family

ID=66358982

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/071109 WO2020118829A1 (zh) 2018-12-12 2019-01-10 基于决策树的pet图像超分辨重建方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN109741255A (zh)
WO (1) WO2020118829A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113128467A (zh) * 2021-05-11 2021-07-16 临沂大学 基于面部先验知识的低分辨率人脸超分辨与识别的方法
CN116188805A (zh) * 2023-04-26 2023-05-30 青岛尘元科技信息有限公司 海量图像的图像内容分析方法、装置和图像信息网络

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110580689A (zh) * 2019-08-19 2019-12-17 深圳先进技术研究院 一种图像重建方法及装置
US11133100B2 (en) * 2019-11-21 2021-09-28 GE Precision Healthcare LLC System and methods for reconstructing medical images using deep neural networks and recursive decimation of measurement data
CN113538242B (zh) * 2021-07-22 2024-06-21 苏州大学 一种基于残差引导策略的图像插值模型训练方法
CN113674335B (zh) * 2021-08-19 2022-05-31 合肥的卢深视科技有限公司 深度成像方法、电子设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107092884A (zh) * 2017-04-20 2017-08-25 宋佳颖 一种快速粗精级联行人检测方法
US20180075581A1 (en) * 2016-09-15 2018-03-15 Twitter, Inc. Super resolution using a generative adversarial network
CN107977930A (zh) * 2017-12-09 2018-05-01 北京花开影视制作有限公司 一种图像超分辨方法及其系统
CN108416734A (zh) * 2018-02-08 2018-08-17 西北大学 基于边缘驱动的文本图像超分辨率重建方法和装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110268334A1 (en) * 2010-04-30 2011-11-03 Korean Advanced Institute Of Science And Technology Apparatus for Improving Image Resolution and Apparatus for Super-Resolution Photography Using Wobble Motion and Point Spread Function (PSF), in Positron Emission Tomography
CN107680146A (zh) * 2017-09-13 2018-02-09 深圳先进技术研究院 Pet图像的重建方法、装置、设备及存储介质
CN107767337A (zh) * 2017-09-29 2018-03-06 深圳先进技术研究院 一种图像超分辨率重建方法、装置、移动终端及存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180075581A1 (en) * 2016-09-15 2018-03-15 Twitter, Inc. Super resolution using a generative adversarial network
CN107092884A (zh) * 2017-04-20 2017-08-25 宋佳颖 一种快速粗精级联行人检测方法
CN107977930A (zh) * 2017-12-09 2018-05-01 北京花开影视制作有限公司 一种图像超分辨方法及其系统
CN108416734A (zh) * 2018-02-08 2018-08-17 西北大学 基于边缘驱动的文本图像超分辨率重建方法和装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LU, HUAN: "Single Image Super-Resolution Reconstruction Based on Random Forests", ELECTRONIC TECHNOLOGY & INFORMATION SCIENCE CHINA MASTER’S THESES FULL-TEXT DATABASE, no. 2, 15 February 2018 (2018-02-15), ISSN: 1674-0246, DOI: 20190831173846X *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113128467A (zh) * 2021-05-11 2021-07-16 临沂大学 基于面部先验知识的低分辨率人脸超分辨与识别的方法
CN113128467B (zh) * 2021-05-11 2022-03-29 临沂大学 基于面部先验知识的低分辨率人脸超分辨与识别的方法
CN116188805A (zh) * 2023-04-26 2023-05-30 青岛尘元科技信息有限公司 海量图像的图像内容分析方法、装置和图像信息网络
CN116188805B (zh) * 2023-04-26 2023-08-04 青岛尘元科技信息有限公司 海量图像的图像内容分析方法、装置和图像信息网络

Also Published As

Publication number Publication date
CN109741255A (zh) 2019-05-10

Similar Documents

Publication Publication Date Title
WO2020118829A1 (zh) 基于决策树的pet图像超分辨重建方法、装置、设备及介质
CN110689038B (zh) 神经网络模型的训练方法、装置和医学图像处理系统
EP3999999A1 (en) Object detection and instance segmentation of 3d point clouds based on deep learning
WO2020118844A1 (zh) Pet图像的重建方法、计算机存储介质、计算机设备
US10258304B1 (en) Method and system for accurate boundary delineation of tubular structures in medical images using infinitely recurrent neural networks
WO2020114329A1 (zh) 磁共振快速参数成像方法及装置
Du et al. Accelerated super-resolution MR image reconstruction via a 3D densely connected deep convolutional neural network
WO2020135014A1 (zh) 医学成像模型的建立方法、装置、设备及存储介质
US20210241037A1 (en) Data processing apparatus and method
KR102428725B1 (ko) 영상 개선 방법 및 이를 수행하는 컴퓨터 프로그램
CN111325695B (zh) 基于多剂量等级的低剂量图像增强方法、系统及存储介质
Yang et al. Super-resolution of medical image using representation learning
EP3617948A1 (en) Capsule networks for analyzing medical image data
CN108038840B (zh) 一种图像处理方法、装置、图像处理设备及存储介质
Li et al. Learning non-local perfusion textures for high-quality computed tomography perfusion imaging
WO2022193379A1 (zh) 图像重建模型生成及装置、图像重建方法及装置、设备、介质
US11693919B2 (en) Anatomy-aware motion estimation
JP7459243B2 (ja) １以上のニューラルネットワークとしての画像形成のモデル化による画像再構成
KR101948701B1 (ko) 피검체의 뇌 구조를 기술하는 잠재 변수에 기반하여 상기 피검체의 뇌질환을 판정하는 방법 및 이를 이용한 장치
US9794521B2 (en) Apparatus and method of resampling event data for quantitative improvement of pet image
EP4343680A1 (en) De-noising data
US12045958B2 (en) Motion artifact correction using artificial neural networks
WO2021159236A1 (zh) 基于非衰减校正pet图像生成合成pet-ct图像的方法和系统
WO2021031069A1 (zh) 一种图像重建方法及装置
Cui et al. PET denoising and uncertainty estimation based on NVAE model

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: 19895359

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

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 02.11.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19895359

Country of ref document: EP

Kind code of ref document: A1