WO2022110530A1 - 基于spect数据采样与噪声特性的断层图像重建方法 - Google Patents

基于spect数据采样与噪声特性的断层图像重建方法 Download PDF

Info

Publication number
WO2022110530A1
WO2022110530A1 PCT/CN2021/073324 CN2021073324W WO2022110530A1 WO 2022110530 A1 WO2022110530 A1 WO 2022110530A1 CN 2021073324 W CN2021073324 W CN 2021073324W WO 2022110530 A1 WO2022110530 A1 WO 2022110530A1
Authority
WO
WIPO (PCT)
Prior art keywords
spect
image
projection data
sampling
data
Prior art date
Application number
PCT/CN2021/073324
Other languages
English (en)
French (fr)
Inventor
李琨
陈思
杨雪松
邓晓
Original Assignee
佛山读图科技有限公司
佛山原子医疗设备有限公司
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 佛山读图科技有限公司, 佛山原子医疗设备有限公司 filed Critical 佛山读图科技有限公司
Publication of WO2022110530A1 publication Critical patent/WO2022110530A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G06T11/003—Reconstruction from projections, e.g. tomography
    • G06T11/008—Specific post-processing after tomographic reconstruction, e.g. voxelisation, metal artifact correction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10108—Single photon emission computed tomography [SPECT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing

Definitions

  • the invention relates to the technical field of medical imaging, in particular to a tomographic image reconstruction method based on SPECT data sampling and noise characteristics.
  • SPECT Single-Photon Emission Computed Tomography, single-photon emission computed tomography equipment
  • SPECT Single-Photon Emission Computed Tomography, single-photon emission computed tomography equipment
  • Its basic working principle is as follows: use a gamma camera to rotate around the imaging target (patient), and detect radiopharmaceuticals in the target at different angles.
  • the emitted gamma photons collimated by the collimator form two-dimensional projection data, and the tomographic image reconstruction algorithm is applied to the projection data collected at all sampling angles to reconstruct the three-dimensional spatial distribution of radiopharmaceuticals in the imaging target.
  • SPECT needs to acquire a certain amount of time at each projection angle—usually between 20-60 seconds—to accumulate gamma photon counts, thereby improving the signal-to-noise ratio of the projection data. Even so, SPECT projections The noise level of the data is still much higher than that of other comparable radiological imaging devices such as PET (Positron Electron Tomography) and X-ray CT.
  • PET Positron Electron Tomography
  • X-ray CT X-ray CT.
  • traditional SPECT needs to complete the projection data sampling of 60 angles within the range of 360 degrees around the imaging target axis to complete the tomographic imaging of one bed, covering the axial field of view. About 40 cm and takes about 15-20 minutes.
  • reducing the acquisition time can be achieved by reducing the sampling time per angle, reducing the number of sampling angles, or a combination of the two.
  • Different acquisition time and number of sampling angles can lead to data and image noise and sparse sampling artifacts characteristics make a big difference.
  • the noise level of SPECT raw projection data and images is also affected by many factors: such as different types or doses of radiopharmaceuticals, the differential distribution of radiopharmaceuticals in different patients or between different parts of the same patient, As well as the parameter settings of conventional image reconstruction algorithms, etc. Therefore, a single, fixed-parameter convolutional neural network cannot perform well the functions of noise reduction and artifact removal for SPECT sparse sampling tomographic reconstruction.
  • the purpose of the present invention is to propose a tomographic image reconstruction method based on SPECT data sampling and noise characteristics, especially for sparsely sampled SPECT data, to solve one or more of the above problems, so as to shorten the SPECT tomography Collect time to improve efficiency.
  • a tomographic image reconstruction method based on SPECT data sampling and noise characteristics including steps:
  • Step A using a Poisson noise model to evaluate the noise level of the SPECT original projection data, and selecting a first convolutional neural network that matches the noise level to perform noise reduction processing on the SPECT original projection data;
  • Step B applying a statistical iterative reconstruction algorithm based on a physical model and fixed parameters to the denoised projection data to obtain a preliminary reconstructed image
  • Step C applying a second convolutional neural network that matches the number of sampling angles of the SPECT projection data to post-process the preliminary reconstructed image to remove artifacts caused by sparse sampling;
  • step D an image iterative reconstruction algorithm based on compressed sensing is further applied based on the reconstructed image after de-artifacting and SPECT original projection data to obtain a final reconstructed image.
  • the method for evaluating the noise level of the SPECT raw projection data described in step A is: calculating the median or average value of all pixel values in the SPECT raw projection data greater than zero pixel values. And classify it by noise level as follows:
  • the first convolutional neural network applied in the step A performs noise reduction for two-dimensional projection data of different angles respectively or performs joint noise reduction for three-dimensional data composed of projection data from multiple angles.
  • the statistical iterative algorithm in step B adopts the maximum likelihood iteration or the ordered subset accelerated maximum likelihood iterative reconstruction algorithm.
  • step C first classify the sampling angle number V of the SPECT original projection data according to the following method:
  • the second convolutional neural network performs de-artifact processing separately for each two-dimensional image layer in the tomographic image, or performs overall processing on the three-dimensional tomographic image.
  • the image iterative reconstruction algorithm in the step D the formula is:
  • x is the SPECT target image vector to be reconstructed
  • x p is the image vector after the de-artifact processing in step c
  • y is the SPECT original projection data vector
  • A is the physical modeling of the SPECT original projection data acquisition process.
  • system transfer matrix For the final reconstructed image, TV(x) is the total variation model, that is, the L2 norm of the bidirectional spatial gradient of the image.
  • the solution of the sparsity cost function of the difference between the target image and the de-artifacted image is used as the final reconstructed image.
  • the first and second convolutional neural networks for noise reduction and de-artifacting of adaptive parameters are applied to achieve optimal image quality.
  • the adaptive and optimized noise reduction and de-artifact algorithm ensures that the image quality basically does not change, thereby improving the efficiency of patient inspection
  • Figure 1 is a flow chart of the present invention.
  • FIG. 2 is a schematic structural diagram of a first convolutional neural network MAP-NN for denoising projection data in an embodiment of the present invention.
  • FIG. 3 is a schematic diagram of the noise reduction effect under different numbers of encoders/decoders in the first convolutional neural network MAP-NN according to an embodiment of the present invention.
  • FIG. 4 is a schematic diagram of the effect of applying MAP-NN to denoise projection data in an embodiment of the present invention.
  • FIG. 5 is a schematic diagram of a U-Net network structure in an embodiment of the present invention.
  • FIG. 6 is a schematic diagram comparing the effect of removing artifacts by applying the Unet network optimized for the preliminary reconstructed images of two sampling angles respectively and applying the conventional Unet network in removing artifacts according to an embodiment of the present invention.
  • FIG. 7 is a schematic diagram of a comparison result between a reconstructed image obtained by applying a complete reconstruction method and other comparison methods according to an embodiment of the present invention.
  • a tomographic image reconstruction method based on SPECT data sampling and noise characteristics including steps:
  • Step A using a Poisson noise model to evaluate the noise level of the SPECT original projection data, and selecting a first convolutional neural network that matches the noise level to perform noise reduction processing on the SPECT original projection data;
  • Step B applying a statistical iterative reconstruction algorithm based on a physical model and fixed parameters to the denoised projection data to obtain a preliminary reconstructed image
  • Step C applying a second convolutional neural network that matches the number of sampling angles of the SPECT projection data to post-process the preliminary reconstructed image to remove artifacts caused by sparse sampling;
  • step D an image iterative reconstruction algorithm based on compressed sensing is further applied based on the reconstructed image after de-artifacting and SPECT original projection data to obtain a final reconstructed image.
  • the tomographic image reconstruction method proposed in this application aiming at the differentiated SPECT data and noise characteristics in practical clinical applications, applies noise reduction and de-artifact convolutional neural networks with adaptive parameters to achieve optimized image quality.
  • the two convolutional neural networks deal with noise and sparse sampling artifacts respectively.
  • the complexity of network parameters and the difficulty of training are reduced.
  • compressed sensing technology and original projection data are applied.
  • the reconstructed image is further improved to avoid the false details of the image introduced by the limitation of the convolutional neural network to the greatest extent.
  • sparsely sampled SPECT raw projection data it solves the problem that the noise of data and images and the characteristics of sparse sampling artifacts are quite different, which can shorten the acquisition time of SPECT tomography and improve the efficiency.
  • the method for evaluating the noise level of the SPECT raw projection data described in step A is: calculating the median or average value of all pixel values in the SPECT raw projection data greater than zero pixel values. And classify it by noise level as follows:
  • Denoising of SPECT raw projection data is achieved by selecting a first convolutional neural network whose network parameters match the estimated noise level. Different neural network parameters in the first convolutional neural network are trained by applying the corresponding noise levels. The dataset is obtained by parameter training.
  • the first convolutional neural network applied in the step A performs noise reduction for two-dimensional projection data of different angles respectively or performs joint noise reduction for three-dimensional data composed of projection data from multiple angles.
  • the statistical iterative algorithm in step B adopts the maximum likelihood iteration or the ordered subset accelerated maximum likelihood iterative reconstruction algorithm.
  • image iteration update number subset number * full iteration number
  • step C first classify the sampling angle number V of SPECT original projection data according to the following method:
  • the sparse artifact removal processing of the preliminary reconstructed image is realized by a second convolutional neural network whose network parameters correspond to the number of sampling angles of the SPECT original projection data. Different neural network parameters in the second convolutional neural network are obtained by The parameters are obtained by applying the training data set with the corresponding sampling angle range for parameter training.
  • the second convolutional neural network performs de-artifact processing separately for each two-dimensional image layer in the tomographic image, or performs overall processing on the three-dimensional tomographic image.
  • the image iterative reconstruction algorithm in the step D its formula is:
  • x is the image vector of the preliminary reconstructed image to be reconstructed
  • x p is the image vector after the de-artifact processing in step C
  • y is the SPECT original projection data vector
  • A is the physical reconstruction of the SPECT original projection data acquisition process.
  • TV(x) is the total variation model, that is, the L2 norm of the bidirectional spatial gradient of the image.
  • the solution of the sparsity cost function of the difference between the image and the de-artifacted image is used as the final reconstructed image.
  • MAP-NN contains multiple encoder-decoder combinations with the same structure
  • Figure 3 shows the noise reduction under different numbers of encoders/decoders. Effect.
  • noise-free simulated projection data is used as the output of the network
  • simulated projection data with Poisson noise added is used as the input of the network.
  • the number T of encoder-decoder combinations in training is 5.
  • a total of 4 noise levels of the network were trained, Represents the average count rate of pixels with counts greater than 0:
  • the number D of encoder/decoder combinations used for prediction by using the model is five.
  • the program first counts the count rate of the image, and based on this, the corresponding network weight is called to denoise the SPECT original projection data.
  • the effect after denoising can be seen in Figure 4, from left to right: (1 ) the preliminary reconstructed image without noise reduction processing, (2) is the preliminary reconstructed image after the noise reduction processing in step A.
  • the present embodiment uses the simulation reconstruction result of sparse angle sampling as the input of the network, and uses the simulation reconstruction result of foot angle sampling as the output of the network to train the first.
  • Two convolutional neural network U-Net During training, a total of four kinds of networks are trained according to the number of sampling angles used by the input image:
  • the program When using the second convolutional neural network U-Net, the program will select the corresponding network weight according to the number of reconstruction angles to remove artifacts from the initial reconstructed image. As shown in Figure 5, due to the detection technology, the input image is surrounded by highlighted areas. There are artifacts. After the output image is processed by the second convolutional neural network U-Net, the artifacts are removed to obtain a higher-quality output image, thereby preventing the artifacts in the image from interfering with subsequent disease diagnosis.
  • (1) is a conventional reconstructed image based on 12 sampling angles
  • the targeted de-sparse sampling second convolutional neural network U-Net proposed in this application can be obtained under 12 sampling angle data training or 12-20 multiple sampling angle data training based on the image obtained, which is the same as that in the conventional method.
  • the conventional reconstructed image using 60 sampling angles is very approximate, that is, under the targeted de-sparse sampling second convolutional neural network U-Net proposed in the application, reducing the number of sampling angles can still ensure better image quality, and its image quality Compared with images with multiple sampling angles, there is basically no transformation, which can reduce data acquisition time and improve patient inspection efficiency.
  • Combining (2), (3) and (4) can be obtained, after applying convolutional neural network noise reduction for projection data and applying convolutional neural network to the reconstruction results of conventional statistical iterative algorithm to remove artifacts, On this basis, the quality of the image obtained by applying compressed sensing reconstruction is obviously improved; and the image quality of (3) is basically the same as that of (4).

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (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)
  • Nuclear Medicine (AREA)
  • Image Processing (AREA)

Abstract

基于SPECT数据采样与噪声特性的断层图像重建方法，其包括步骤：步骤A，评估SPECT原始投影数据的噪声水平，选择与噪声水平相匹配的第一卷积神经网络对SPECT原始投影数据进行降噪处理；步骤B，对降噪后的投影数据应用统计迭代重建算法，得出初步重建图像；步骤C，应用与SPECT投影数据采样角度数目相匹配的第二卷积神经网络对初步重建图像进行后处理；步骤D，基于去伪影后的重建图像和SPECT原始投影数据进一步应用基于压缩感知的图像迭代重建算法，得到最终的重建图像。上述重建方法，特别针对稀疏采样的SPECT数据，可缩短SPECT断层成像采集时间，进行提升效率。

Description

基于SPECT数据采样与噪声特性的断层图像重建方法 技术领域
本发明涉及医学成像技术领域，特别是基于SPECT数据采样与噪声特性的断层图像重建方法。
背景技术
SPECT(Single-Photon Emission Computed Tomography，单光子发射计算机断层成像设备)是一种医学成像设备，其基本工作原理如下：利用伽马相机环绕成像目标(病人)旋转，在不同角度探测目标体内放射性药物发出的经准直器准直的伽马光子，形成二维投影数据，对所有采样角度采集的投影数据应用断层图像重建算法重建出成像目标体内放射性药物的三维空间分布。
受限于成像物理原理，SPECT在每个投影角度需要采集一定时间——通常在20-60秒之间——以积累伽马光子计数，从而提升投影数据的信噪比，即便如此，SPECT投影数据的噪声水平仍然远远高于PET(正电子断层成像设备)和X光CT等其它同类放射成像设备。为满足断层成像的角度采样要求，在全身成像应用中，传统SPECT需要在环绕成像目标轴向360度角的范围内完成60个角度的投影数据采样以完成一个床位的断层成像，覆盖轴向视野约40厘米，耗时约15-20分钟。以此类推，3个床位覆盖约1.2米轴向视野的断层成像需要约1个小时，从临床工作效率和病人耐受性的角度，这是无法接受的。因此一种能够在降低数据采集时间的前提下仍能基本保持图像质量不变的图像重建方法，会对SPECT断层成像临床应用的推广普及产生重要的推动作用。
近年来，以卷积神经网络为代表的深度学习技术，广泛应用于CT、PET医学图像降噪和去除稀疏采样引入的伪影等领域，取得了较好的效果。但是其局 限性在于对差异化数据的适应性较差，当实际应用数据与训练神经网络数据的噪声水平和伪影特性接近时，可以有较好的降噪和去伪影效果；否则，降噪和去伪影效果就会有不同程度的降低。
对于SPECT成像来说，降低采集时间可以通过降低每个角度采样时间、减少采样角数目以及二者结合来实现，不同的采集时间和采样角数目选择都会导致数据和图像的噪声与稀疏采样伪影的特性产生较大差异。另外，在临床实践中，SPECT原始投影数据和图像的噪声水平还受很多因素的影像响：如不同种类或剂量的放射性药物，放射性药物在不同病人或同一病人不同部位之间的差异化分布、以及常规图像重建算法的参数设置等。因此，单一的、固定参数的卷积神经网络无法很好的地完成SPECT稀疏采样断层重建的降噪和去伪影的功能。
发明内容
针对上述缺陷，本发明的目的在于提出基于SPECT数据采样与噪声特性的断层图像重建方法，特别针对稀疏采样的SPECT数据，旨在解决上述的问题中的一项或多项，以缩短SPECT断层成像采集时间，进行提升效率。
为达此目的，本发明采用以下技术方案：
基于SPECT数据采样与噪声特性的断层图像重建方法，包括步骤：
步骤A，使用泊松噪声模型评估SPECT原始投影数据的噪声水平，选择与噪声水平相匹配的第一卷积神经网络对SPECT原始投影数据进行降噪处理；
步骤B，对降噪后的投影数据应用基于物理模型、参数固定的统计迭代重建算法，得出初步重建图像；
步骤C，应用与SPECT投影数据采样角度数目相匹配的第二卷积神经网络对初步重建图像进行后处理，去除因稀疏采样引起的伪影；
步骤D，基于去伪影后的重建图像和SPECT原始投影数据进一步应用基于压缩感知的图像迭代重建算法，得到最终的重建图像。
进一步的说明，步骤A中所述评估SPECT原始投影数据的噪声水平的方法为：计算SPECT原始投影数据中所有像素值大于零像素的像素值的中值或平均值
Figure PCTCN2021073324-appb-000001
并将其按以下方法进行噪声水平分类：
第一噪声水平：
Figure PCTCN2021073324-appb-000002
第二噪声水平：
Figure PCTCN2021073324-appb-000003
第三噪声水平：
Figure PCTCN2021073324-appb-000004
第四噪声水平：
Figure PCTCN2021073324-appb-000005
进一步的说明，所述步骤A中所应用的第一卷积神经网络针对不同角度的二维投影数据分别进行降噪或者针对多个角度的投影数据组成的三维数据进行联合降噪。
更优的，步骤B中的统计迭代算法采用最大似然迭代或有序子集加速最大似然迭代重建算法。
具体的，步骤C中，先对SPECT原始投影数据的采样角数目V按以下方法进行分类：
第一采样角度范围：V＝{8,9,10}；
第二采样角度范围：V＝{12,14,16}；
第三采样角度范围：V＝{18,20,24}；
第四采样角度范围：V＝{30,32,36,40}。
具体的，所述步骤C中第二卷积神经网络针对断层图像中的每一个二维图像层分别进行去伪影处理，或针对三维断层图像进行整体处理。
具体的，所述步骤D中图像迭代重建算法，其公式为：
Figure PCTCN2021073324-appb-000006
Figure PCTCN2021073324-appb-000007
其中，x为拟重建的SPECT目标图像矢量，x p为步骤c中经去伪影处理后的图像矢量，y为SPECT原始投影数据矢量，A为对SPECT原始投影数据采集过程进行物理建模的系统传输矩阵，
Figure PCTCN2021073324-appb-000008
为最终重建图像，TV(x)为全变分模型，即图像双向空间梯度的L2范数，当x为尺寸M*N的二维图像的矢量表示时，其公式为：
Figure PCTCN2021073324-appb-000009
所述M和N为自然数，n＝1，2，…，N-1，
Figure PCTCN2021073324-appb-000010
为图像矢量x的L1范数，即所有像素值的绝对值之和，α为权重参数。
进一步的，步骤D中所述图像迭代重建算法为：在满足SPECT重建目标图像与SPECT原始投影数据的一致性条件，即Ax＝y的前提下，求解能够最小化公式(1)所述SPECT重建目标图像与去伪影后图像之差的稀疏性代价函数的解，作为最终的重建图像。
优选的，对所述步骤D中的求解公式(2)进行优化，具体是以去伪影后图像为初始估计，通过迭代交替应用公式(2)中的一致性条件Ax＝y和公式(1)中稀疏性最小化条件对SPECT目标重建图像进行更新，待其收敛后，即得到最终的重建图像。
本发明可以达到以下有益效果：
1、针对临床实际应用中差异化的SPECT原始投影数据与其噪声特性，应用自适应参数的降噪与去伪影的第一卷积神经网络和第二卷积神经网络，实现最优化的图像质量；
2、在适当降低数据采集时间情况下，通过自适应优化的降噪和去伪影算法，确保图像质量基本不发生变化换，从而提升病人检查效率；
3、采用两个卷积神经网络分别处理噪声和稀疏采样伪影问题，与采用单一网络的方案相比，降低了网络参数的复杂度和训练难度。
4、应用压缩感知技术和原始投影数据，在两个卷积神经网络的基础上进一步对重建图像进行完善，最大程度规避因为卷积神经网络的局限性引入图像的虚假细节。
附图说明
图1是本发明的流程图。
图2是本发明的一个实施例中，对投影数据降噪的第一卷积神经网络MAP-NN的结构示意图。
图3是本发明的一个实施例中，第一卷积神经网络MAP-NN中不同编码/解码器数目下的降噪效果示意图。
图4是本发明一个实施例中，应用MAP-NN对投影数据进行降噪的效果示意图。
图5是本发明的一个实施例中，U-Net网络结构示意图。
图6是本发明的一个实施例中，应用分别针对两种采样角度的初步重建图像进行优化的Unet网络去除伪影的效果和应用常规Unet网络去除伪影效果的对比示意图。
图7是本发明的一个实施例中，应用完整的重建方法得出的重建图像与其它对比方法的比较结果示意图。
具体实施方式
下面详细描述本发明的实施例，所述实施例的示例在附图中示出，其中自 始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施例是示例性的，仅用于解释本发明，而不能理解为对本发明的限制。
此外，术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。
基于SPECT数据采样与噪声特性的断层图像重建方法，包括步骤：
步骤A，使用泊松噪声模型评估SPECT原始投影数据的噪声水平，选择与噪声水平相匹配的第一卷积神经网络对SPECT原始投影数据进行降噪处理；
步骤B，对降噪后的投影数据应用基于物理模型、参数固定的统计迭代重建算法，得出初步重建图像；
步骤C，应用与SPECT投影数据采样角度数目相匹配的第二卷积神经网络对初步重建图像进行后处理，去除因稀疏采样引起的伪影；
步骤D，基于去伪影后的重建图像和SPECT原始投影数据进一步应用基于压缩感知的图像迭代重建算法，得到最终的重建图像。
本申请提出的断层图像重建方法，针对临床实际应用中差异化的SPECT数据与噪声特性，应用自适应参数的降噪与去伪影卷积神经网络，实现最优化的图像质量，具体的，采用两个卷积神经网络分别处理噪声和稀疏采样伪影问题，与采用单一网络的方案相比，降低了网络参数的复杂度和训练难度，同时应用压缩感知技术和原始投影数据，在卷积神经网络的基础上进一步对重建图像进行完善，最大程度规避因为卷积神经网络的局限性引入图像的虚假细节。特别针对稀疏采样的SPECT原始投影数据，解决了数据和图像的噪声与稀疏采样伪影的特性产生较大差异的问题，可以缩短SPECT断层成像采集时间，从而提升效率。
进一步的说明，步骤A中所述评估SPECT原始投影数据的噪声水平的方法为：计算SPECT原始投影数据中所有像素值大于零像素的像素值的中值或平均值
Figure PCTCN2021073324-appb-000011
并将其按以下方法进行噪声水平分类：
第一噪声水平：
Figure PCTCN2021073324-appb-000012
第二噪声水平：
Figure PCTCN2021073324-appb-000013
第三噪声水平：
Figure PCTCN2021073324-appb-000014
第四噪声水平：
Figure PCTCN2021073324-appb-000015
对SPECT原始投影数据的降噪是通过选择网络参数与评估噪声水平相匹配的第一卷积神经网络实现的，第一卷积神经网络中不同的神经网络参数是通过应用相对应噪声水平的训练数据集进行参数训练获得的。
更优的，所述步骤A中所应用的第一卷积神经网络针对不同角度的二维投影数据分别进行降噪或者针对多个角度的投影数据组成的三维数据进行联合降噪。
优选的，步骤B中的统计迭代算法采用最大似然迭代或有序子集加速最大似然迭代重建算法。
对SPECT成像过程中的伽马光子衰减、散射以及准直器探测器响应等物理过程进行建模，同时图像迭代更新次数(图像迭代更新次数＝子集数*全迭代次数)保持固定。
进一步的说明，步骤C中，先对SPECT原始投影数据的采样角数目V按以下方法进行分类：
第一采样角度范围：V＝{8,9,10}；
第二采样角度范围：V＝{12,14,16}；
第三采样角度范围：V＝{18,20,24}；
第四采样角度范围：V＝{30,32,36,40}。
对于初步重建图像的去除稀疏伪影处理，是由网络参数与SPECT原始投影数据的采样角数目相对应的第二卷积神经网络实现的，第二卷积神经网络中不同的神经网络参数是通过应用相对应的采样角度范围的训练数据集进行参数训练获得的。
优选的，所述步骤C中第二卷积神经网络针对断层图像中的每一个二维图像层分别进行去伪影处理，或针对三维断层图像进行整体处理。
进一步的说明，所述步骤D中图像迭代重建算法，其公式为：
Figure PCTCN2021073324-appb-000016
Figure PCTCN2021073324-appb-000017
其中，x为拟重建的初步重建图像的图像矢量，x p为步骤C中经去伪影处理后的图像矢量，y为SPECT原始投影数据矢量，A为对SPECT原始投影数据采集过程进行物理建模的系统传输矩阵，
Figure PCTCN2021073324-appb-000018
为最终重建图像，TV(x)为全变分模型，即图像双向空间梯度的L2范数，当x为尺寸M*N的二维图像的矢量表示时，其公式为：
Figure PCTCN2021073324-appb-000019
所述M和N为自然数，n＝1，2，…，N-1，
Figure PCTCN2021073324-appb-000020
为图像矢量x的L1范数，即所有像素值的绝对值之和，α为权重参数。
优选的，步骤D所述图像迭代重建算法为：在满足SPECT重建目标图像与SPECT原始投影数据的一致性条件，即Ax＝y的前提下，求解能够最小化公式(1)所述SPECT重建目标图像与去伪影后图像之差的稀疏性代价函数的解，作为最终的重建图像。
更优的，对所述步骤D中的求解公式(2)进行优化，具体是以去伪影后图像为初始估计，通过迭代交替应用公式(2)中的一致性条件Ax＝y和公式(1)中稀疏性最小化条件对SPECT目标重建图像进行更新，待其收敛后，即得到最终的重建图像。
具体示例如下：
参见图2所示的第一卷积神经网络MAP-NN结构，MAP-NN中包含了多个结构相同的编码-解码器组合，图3所示是其不同编码/解码器数目下的降噪效果。本实施例的第一卷积神经网络MAP-NN训练中使用无噪声的仿真投影数据作为网络的输出，以添加了泊松噪声的仿真投影数据作为网络的输入。训练中编码-解码器组合数T为5。训练时，总共训练了4种噪声水平的网络，
Figure PCTCN2021073324-appb-000021
代表计数大于0的像素的平均计数率：
第一噪声水平：
Figure PCTCN2021073324-appb-000022
第二噪声水平：
Figure PCTCN2021073324-appb-000023
第三噪声水平：
Figure PCTCN2021073324-appb-000024
第四噪声水平：
Figure PCTCN2021073324-appb-000025
在本实施例训练的MAP-NN网络的使用中，在训练完成后，利用模型进行预测是所采用的编码/解码器组合数D为5。在使用时，首先由程序统计出图像的计数率，并以此为根据调用相应的网络权重对SPECT原始投影数据进行降噪，降噪后的效果可参见图4，从左到右：(1)未经降噪处理的初步重建图像，(2)是经步骤A降噪处理后的初步重建图像。
另一实施例，在第二卷积神经网络U-Net的训练中，本实施例使用稀疏角度采样的仿真重建结果作为网络的输入，用足角度采样的仿真重建结果作为网络的输出来训练第二卷积神经网络U-Net。训练时，根据输入图像使用的采样角度数，总共训练了四种网络：
第一采样角度范围：V＝{8,9,10}；
第二采样角度范围：V＝{12,14,16}；
第三采样角度范围：V＝{18,20,24}；
第四采样角度范围：V＝{30,32,36,40}。
在使用第二卷积神经网络U-Net时，程序会根据重建角度数选择相应的网络权重对初步重建图像进行去伪影，如图5所示，由于检测技术造成输入的图像中高亮区域周围存在伪影，输出图像经过第二卷积神经网络U-Net处理后，伪影被去除从而得到更高质量的输出的图像，进而避免图像中伪影对后续疾病诊断造成干扰。
对比例组，将上述针对性去稀疏采样第二卷积神经网络U-Net和常规的同一卷积神经网络对同一图像进行后处理，其去伪影效果示意如图6所示，从左到右：
(1)是基于12个采样角的常规重建图像；
(2)是应用了基于12个采样角数据训练的卷积神经网络U-Net去伪影后的图像；
(3)是应用了基于12-20之间多种采样角数据训练的卷积神经网络U-Net去伪影后的图像；
(4)为基于60个采样角的常规重建图像(真实值)。
可得本申请提出的针对性去稀疏采样第二卷积神经网络U-Net其在12个采 样角数据训练或12-20之间多种采样角数据训练基础下获得的图像，与常规方法中采用60个采样角的常规重建图像十分近似，即在申请提出的针对性去稀疏采样第二卷积神经网络U-Net下，减少采样角的数量仍可确保较好的图像质量，其图像质量与多采样角数量的图像相比基本不发生变换，可以降低数据采集时间，从而提升病人检查效率。
使用本申请提出的完整的断层图像重建方法和常规其他方法进行对比，其比较结果如图7所示，从左到右：
(1)15个角度噪声投影数据的常规重建图像；
(2)针对投影数据应用卷积神经网络降噪，并对常规统计迭代算法的重建结果应用卷积神经网络去伪影之后的图像；
(3)使用本发明所用方法的重建图像，即再(2)的基础上应用压缩感知重建后的图像；
(4)60个角度低噪声投影数据重建图像(真实值)。
结合(2)、(3)和(4)可得，在可得针对投影数据应用卷积神经网络降噪，并对常规统计迭代算法的重建结果应用卷积神经网络去伪影之后，再在此基础上应用压缩感知重建后得到的图像的质量明显提高；且(3)的图像质量与(4)的图像质量基本一致。
在本说明书的描述中，参考术语“一个实施例”、“另一实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本发明的至少一个实施例或示例中。在本说明书中，对上述术语的示意性表述不一定指的是相同的实施例或示例。而且，描述的具体特征、结构、材料或者特点可以在任何的一个或多个实施例或示例中以合适的方式结合。
尽管已经示出和描述了本发明的实施例，本领域的普通技术人员可以理解：在不脱离本发明的原理和宗旨的情况下可以对这些实施例进行多种变化、修改、替换和变型，本发明的范围由权利要求及其等同物限定。

Claims (9)

  1. 基于SPECT数据采样与噪声特性的断层图像重建方法，其特征在于：包括步骤：
    步骤A，使用泊松噪声模型评估SPECT原始投影数据的噪声水平，选择与噪声水平相匹配的第一卷积神经网络对SPECT原始投影数据进行降噪处理；
    步骤B，对降噪后的投影数据应用基于物理模型、参数固定的统计迭代重建算法，得出初步重建图像；
    步骤C，应用与SPECT投影数据采样角度数目相匹配的第二卷积神经网络对初步重建图像进行后处理，去除因稀疏采样引起的伪影；
    步骤D，基于去伪影后的重建图像和SPECT原始投影数据进一步应用基于压缩感知的图像迭代重建算法，得到最终的重建图像。
  2. 根据权利要求1所述的基于SPECT数据采样与噪声特性的断层图像重建方法，其特征在于：步骤A中所述评估SPECT原始投影数据的噪声水平的方法为：计算SPECT原始投影数据中所有像素值大于零像素的像素值的中值或平均值
    Figure PCTCN2021073324-appb-100001
    并将其按以下方法进行噪声水平分类：
    第一噪声水平：
    Figure PCTCN2021073324-appb-100002
    第二噪声水平：
    Figure PCTCN2021073324-appb-100003
    第三噪声水平：
    Figure PCTCN2021073324-appb-100004
    第四噪声水平：
    Figure PCTCN2021073324-appb-100005
  3. 根据权利要求2所述的基于SPECT数据采样与噪声特性的断层图像重建方法，其特征在于：所述步骤A中所应用的第一卷积神经网络针对不同角度的二维投影数据分别进行降噪或者针对多个角度的投影数据组成的三维数据进行联合降噪。
  4. 根据权利要求1所述的基于SPECT数据采样与噪声特性的断层图像重 建方法，其特征在于：步骤B中的统计迭代算法采用最大似然迭代或有序子集加速最大似然迭代重建算法。
  5. 根据权利要求1所述的基于SPECT数据采样与噪声特性的断层图像重建方法，其特征在于：步骤C中，先对SPECT原始投影数据的采样角数目V按以下方法进行分类：
    第一采样角度范围：V＝{8,9,10}；
    第二采样角度范围：V＝{12,14,16}；
    第三采样角度范围：V＝{18,20,24}；
    第四采样角度范围：V＝{30,32,36,40}。
  6. 根据权利要求1所述的基于SPECT数据采样与噪声特性的断层图像重建方法，其特征在于：所述步骤C中第二卷积神经网络针对断层图像中的每一个二维图像层分别进行去伪影处理，或针对三维断层图像进行整体处理。
  7. 根据权利要求1所述的基于SPECT数据采样与噪声特性的断层图像重建方法，其特征在于：所述步骤D中图像迭代重建算法，其公式为：
    Figure PCTCN2021073324-appb-100006
    Figure PCTCN2021073324-appb-100007
    其中，x为拟重建的SPECT目标图像矢量，x p为步骤c中经去伪影处理后的图像矢量，y为SPECT原始投影数据矢量，A为对SPECT原始投影数据采集过程进行物理建模的系统传输矩阵，
    Figure PCTCN2021073324-appb-100008
    为最终重建图像，TV(x)为全变分模型，即图像双向空间梯度的L2范数，当x为尺寸M*N的二维图像的矢量表示时，其公式为：
    Figure PCTCN2021073324-appb-100009
    所述M和N为自然数，n＝1，2，…，N-1，
    Figure PCTCN2021073324-appb-100010
    为图像矢量x的L1范数，即所有像素值的绝对值之和，α为权重参数。
  8. 根据权利要求7所述的基于SPECT数据采样与噪声特性的断层图像重建方法，其特征在于：步骤D所述图像迭代重建算法为：在满足SPECT重建目标图像与SPECT原始投影数据的一致性条件，即Ax＝y的前提下，求解能够最小化公式(1)所述SPECT重建目标图像与去伪影后图像之差的稀疏性代价函数的解，作为最终的重建图像。
  9. 根据权利要求8所述的基于SPECT数据采样与噪声特性的断层图像重建方法，其特征在于：对所述步骤D中的求解公式(2)进行优化，具体是以去伪影后图像为初始估计，通过迭代交替应用公式(2)中的一致性条件Ax＝y和公式(1)中稀疏性最小化条件对SPECT目标重建图像进行更新，待其收敛后，即得到最终的重建图像。
PCT/CN2021/073324 2020-11-24 2021-01-22 基于spect数据采样与噪声特性的断层图像重建方法 WO2022110530A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011333288.6A CN112381741B (zh) 2020-11-24 2020-11-24 基于spect数据采样与噪声特性的断层图像重建方法
CN202011333288.6 2020-11-24

Publications (1)

Publication Number Publication Date
WO2022110530A1 true WO2022110530A1 (zh) 2022-06-02

Family

ID=74588367

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/073324 WO2022110530A1 (zh) 2020-11-24 2021-01-22 基于spect数据采样与噪声特性的断层图像重建方法

Country Status (2)

Country Link
CN (1) CN112381741B (zh)
WO (1) WO2022110530A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115856767A (zh) * 2023-02-08 2023-03-28 安徽大学 一种可重构智能超表面辅助的波到达方向估计方法

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11540798B2 (en) 2019-08-30 2023-01-03 The Research Foundation For The State University Of New York Dilated convolutional neural network system and method for positron emission tomography (PET) image denoising
CN113012252A (zh) * 2021-03-24 2021-06-22 苏州深透智能科技有限公司 一种spect成像预测模型创建方法、装置、设备及存储介质
CN113706419A (zh) * 2021-09-13 2021-11-26 上海联影医疗科技股份有限公司 一种图像处理方法和系统
CN115187470A (zh) * 2022-06-10 2022-10-14 成都飞机工业（集团）有限责任公司 一种基于3d打印内型腔的双域迭代降噪方法
CN115423890B (zh) * 2022-09-15 2023-09-19 京心禾(北京)医疗科技有限公司 一种断层图像迭代重建方法
CN115908204B (zh) * 2023-02-21 2023-06-06 北京唯迈医疗设备有限公司 用于放射成像设备获取的医学图像的降噪处理方法及装置

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109559359A (zh) * 2018-09-27 2019-04-02 东南大学 基于深度学习实现的稀疏角度数据重建图像的伪影去除方法
CN110151210A (zh) * 2019-05-21 2019-08-23 上海联影医疗科技有限公司 一种医学图像处理方法、系统、装置和计算机可读介质
CN110249365A (zh) * 2017-11-10 2019-09-17 上海联影医疗科技有限公司 用于图像重建的系统和方法
CN110462689A (zh) * 2017-04-05 2019-11-15 通用电气公司 基于深度学习的断层摄影重建
US20200027252A1 (en) * 2018-06-04 2020-01-23 Korea Advanced Institute Of Science And Technology Method for processing multi-directional x-ray computed tomography image using artificial neural network and apparatus therefor
CN111323437A (zh) * 2018-12-17 2020-06-23 通用电气公司 用于统计迭代重建和材料分解的系统和方法
CN111492406A (zh) * 2017-10-11 2020-08-04 通用电气公司 使用机器学习的图像生成
CN111681297A (zh) * 2020-05-28 2020-09-18 上海联影智能医疗科技有限公司 图像重建方法、计算机设备和存储介质

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104077763B (zh) * 2014-07-08 2017-01-25 中国科学院高能物理研究所 基于压缩感知理论的tof‑pet图像重建方法
US11517197B2 (en) * 2017-10-06 2022-12-06 Canon Medical Systems Corporation Apparatus and method for medical image reconstruction using deep learning for computed tomography (CT) image noise and artifacts reduction
CN108898642B (zh) * 2018-06-01 2022-11-11 安徽工程大学 一种基于卷积神经网络的稀疏角度ct成像方法
CN110660123B (zh) * 2018-06-29 2021-08-31 清华大学 基于神经网络的三维ct图像重建方法和设备以及存储介质
WO2020062262A1 (en) * 2018-09-30 2020-04-02 Shanghai United Imaging Healthcare Co., Ltd. Systems and methods for generating a neural network model for image processing
CN110264559B (zh) * 2019-04-01 2021-02-19 佛山原子医疗设备有限公司 一种骨断层图像重建方法及系统
US11315221B2 (en) * 2019-04-01 2022-04-26 Canon Medical Systems Corporation Apparatus and method for image reconstruction using feature-aware deep learning
CN111292386B (zh) * 2020-01-15 2023-05-23 中国人民解放军战略支援部队信息工程大学 一种基于U-net的CT投影金属迹线补全金属伪影校正方法

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110462689A (zh) * 2017-04-05 2019-11-15 通用电气公司 基于深度学习的断层摄影重建
CN111492406A (zh) * 2017-10-11 2020-08-04 通用电气公司 使用机器学习的图像生成
CN110249365A (zh) * 2017-11-10 2019-09-17 上海联影医疗科技有限公司 用于图像重建的系统和方法
US20200027252A1 (en) * 2018-06-04 2020-01-23 Korea Advanced Institute Of Science And Technology Method for processing multi-directional x-ray computed tomography image using artificial neural network and apparatus therefor
CN109559359A (zh) * 2018-09-27 2019-04-02 东南大学 基于深度学习实现的稀疏角度数据重建图像的伪影去除方法
CN111323437A (zh) * 2018-12-17 2020-06-23 通用电气公司 用于统计迭代重建和材料分解的系统和方法
CN110151210A (zh) * 2019-05-21 2019-08-23 上海联影医疗科技有限公司 一种医学图像处理方法、系统、装置和计算机可读介质
CN111681297A (zh) * 2020-05-28 2020-09-18 上海联影智能医疗科技有限公司 图像重建方法、计算机设备和存储介质

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115856767A (zh) * 2023-02-08 2023-03-28 安徽大学 一种可重构智能超表面辅助的波到达方向估计方法

Also Published As

Publication number Publication date
CN112381741A (zh) 2021-02-19
CN112381741B (zh) 2021-07-16

Similar Documents

Publication Publication Date Title
WO2022110530A1 (zh) 基于spect数据采样与噪声特性的断层图像重建方法
CN111325686B (zh) 一种基于深度学习的低剂量pet三维重建方法
CN108898642B (zh) 一种基于卷积神经网络的稀疏角度ct成像方法
La Riviere et al. Reduction of noise-induced streak artifacts in X-ray computed tomography through spline-based penalized-likelihood sinogram smoothing
JP2020168352A (ja) 医用装置及びプログラム
CN111429379B (zh) 一种基于自监督学习的低剂量ct图像去噪方法及系统
CN107481297A (zh) 一种基于卷积神经网络的ct图像重建方法
CN112258642B (zh) 基于深度学习的低剂量pet数据三维迭代更新重建方法
Zhang et al. Low-dose CT reconstruction via L1 dictionary learning regularization using iteratively reweighted least-squares
JP2021013725A (ja) 医用装置
CN103810734B (zh) 一种低剂量x射线ct投影数据恢复方法
Zhang et al. PET image reconstruction using a cascading back-projection neural network
Zhang et al. REDAEP: Robust and enhanced denoising autoencoding prior for sparse-view CT reconstruction
Van Slambrouck et al. Reconstruction scheme for accelerated maximum likelihood reconstruction: The patchwork structure
Zhang et al. DREAM-Net: Deep residual error iterative minimization network for sparse-view CT reconstruction
Feng et al. Dual residual convolutional neural network (DRCNN) for low-dose CT imaging
CN108596995A (zh) 一种pet-mri最大后验联合重建方法
Li et al. LU-Net: combining LSTM and U-Net for sinogram synthesis in sparse-view SPECT reconstruction
Zhao et al. Few-view CT reconstruction method based on deep learning
CN112656438A (zh) 一种基于曲面全变差的低剂量ct投影域去噪及重建方法
US20220375038A1 (en) Systems and methods for computed tomography image denoising with a bias-reducing loss function
Moriakov et al. Lire: Learned invertible reconstruction for cone beam ct
Ma et al. Low dose CT reconstruction assisted by an image manifold prior
CN110176045A (zh) 一种由单能ct图像生成双能ct图像的方法
Liang et al. A model-based unsupervised deep learning method for low-dose CT reconstruction

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: 21896050

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: 21896050

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 21896050

Country of ref document: EP

Kind code of ref document: A1