WO2020108562A1 - 一种ct图像内的肿瘤自动分割方法及系统 - Google Patents

一种ct图像内的肿瘤自动分割方法及系统 Download PDF

Info

Publication number
WO2020108562A1
WO2020108562A1 PCT/CN2019/121594 CN2019121594W WO2020108562A1 WO 2020108562 A1 WO2020108562 A1 WO 2020108562A1 CN 2019121594 W CN2019121594 W CN 2019121594W WO 2020108562 A1 WO2020108562 A1 WO 2020108562A1
Authority
WO
WIPO (PCT)
Prior art keywords
layer
data
image
convolutional
construct
Prior art date
Application number
PCT/CN2019/121594
Other languages
English (en)
French (fr)
Inventor
贾富仓
方驰华
初陈曦
Original Assignee
中国科学院深圳先进技术研究院
南方医科大学珠江医院
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 中国科学院深圳先进技术研究院, 南方医科大学珠江医院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2020108562A1 publication Critical patent/WO2020108562A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • 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
    • 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
    • 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
    • G06T2207/30096—Tumor; Lesion

Definitions

  • the invention relates to the technical field of medical image processing, in particular to a method and system for automatic tumor segmentation in CT images.
  • Liver tumors are major diseases that threaten human health. Early detection and accurate measurement of liver tumors are of great significance for clinical diagnosis and treatment.
  • CT Computer Tomography, Chinese name for electronic computer tomography
  • CT has gradually become a routine method of clinical diagnosis of liver tumors.
  • the rapid and accurate segmentation of tumor lesions from liver CT images is not only helpful for surgical planning, but also has great practical value for the precise positioning and resection of intraoperative tumor regions and the evaluation of postoperative radiotherapy and chemotherapy effects.
  • the network structure of CNN is relatively simple, and the addition of a fully connected layer makes the overall training parameters of the network relatively large, the calculation is complicated, the amount of information is large, the network training time is long, and the segmentation accuracy is poor. Afterwards, the overall segmentation accuracy of the improved FCN is still low, and the pixel-based classification does not consider the relationship between pixels and lacks spatial consistency. Since then, some segmentation network architectures have been proposed based on these classic network architectures, but the segmentation accuracy still needs to be improved.
  • the main purpose of the present invention is to provide an automatic tumor segmentation method and system in CT images, aiming to solve the technical problem of low segmentation accuracy of CT images in the prior art.
  • the first aspect of the present invention provides a method for automatic tumor segmentation in CT images, including: performing data enhancement and expansion on original image data to obtain enhanced expansion data; and performing normalization processing on the enhanced expansion data to obtain normalized data Normalize the data; put the normalized data into the trained processing network to get the segmented image; perform noise reduction on the segmented image.
  • the data enhancement and expansion of the original image data includes: enhancement and expansion of the original image data based on the translation and rotation principle, or enhancement and expansion of the original image data based on the principle of random elastic deformation.
  • the normalization processing of the enhanced expansion data includes: performing normalization processing on the enhanced expansion data and the liver tumor gold standard according to the principle of linear normalization to obtain linear normalized image data;
  • the normalized image data is subjected to normalization processing of data distribution to obtain normalized data.
  • the training method of the trained processing network includes: constructing a first convolutional layer, a first modified linear unit layer, a pooling layer, a dropout layer, and a downsampling layer to form a contraction path; according to the first volume
  • the accumulation layer, the first modified linear unit layer, the pooling layer, the dropout layer and the down-sampling layer extract and encode the noise reduction data to generate coded data; construct a second convolution layer and a second modified linear unit layer And the upsampling layer to form an expansion path; decode and divide the sample encoded data according to the second convolutional layer, the second modified linear unit layer and the upsampling layer to generate decoded data; construct a probability output layer and output the decoding data.
  • the constructing the first convolutional layer, the first modified linear unit layer, the pooling layer, the dropout layer, and the downsampling layer includes: sequentially constructing three first 3*3 convolutional layers, and constructing two first 2*2 convolutional layers, and the number of feature channels of the first 3*3 convolutional layer and the first 2*2 convolutional layer constructed in sequence starts from 64 and doubles in turn; in each first 3*3 volume Build a modified linear unit layer after stacking; between adjacent first 3*3 convolutional layers, between adjacent first 2*2 convolutional layers, and adjacent first 3*3 convolutional layers and The pooling layer is constructed between the first 2*2 convolutional layers; the first first 2*2 convolutional layer is constructed after the first 3*3 convolutional layer, and the first dropout layer is constructed.
  • the unit layer and the up-sampling layer include: after the second dropout layer is constructed, four second-volume base layers are sequentially constructed, each consisting of an up-sampling layer and two second 3*3 convolutional layers, and the second The number of feature channels of the convolutional base layer starts with the second first 2*2 convolutional layer, and the number of feature channels is sequentially reduced by a factor of two; an upsampling layer is constructed before each second 3*3 convolutional layer.
  • contraction path and the expansion path are mirror-image distributed and cascaded.
  • the training method of the trained processing network further includes: constructing an adversarial network, the constructing the adversarial network includes: constructing first data, the first data is referenced by a gold splitting standard, and Golden standard and original liver grayscale image dot product; construct second data, the second data is obtained by dot multiplying the segmented image and the original liver grayscale image; construct a loss function, and combine the first data and the second
  • the data input loss function captures the long and short spatial features of pixel levels from different levels.
  • a second aspect of the present invention provides an automatic tumor segmentation system in a CT image, including: an enhanced expansion module for enhancing and expanding original image data to obtain enhanced expansion data; and a normalization processing module for enhancing enhanced expansion data Perform normalization processing to obtain normalized data; image segmentation module, which is used to input the noise reduction data into the trained processing network to obtain segmented images; noise reduction module, which is used to perform noise reduction processing on the segmented images.
  • a third aspect of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program When any of the above methods are implemented.
  • a fourth aspect of the present invention provides a computer-readable storage medium, a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, any of the above The method.
  • the invention provides an automatic tumor segmentation method in CT images.
  • the beneficial effect lies in: enhancing and expanding the original image data, on the basis of satisfying the robust requirements of changing grayscale information and not affecting the processing of real information by the processing network , Can achieve the result of enriching the amount of information, thereby improving the generalization effect of the segmentation model, making the segmentation model applicable to a wider data set, and improving the scope of the processing network; due to the difference in gray value of the original image obtained by different patients Large, so through normalization processing, it can facilitate the processing of the image data by the segmentation processing network, thereby reducing the difference of different original images brought by different CT scans, and improving the scope and accuracy of processing network processing results.
  • FIG. 1 is a schematic block diagram of a structure of an automatic tumor segmentation method in CT images according to an embodiment of the present invention
  • FIG. 2 is a schematic structural diagram of a trained processing network in an automatic tumor segmentation method in CT images according to an embodiment of the present invention
  • FIG. 3 is a schematic structural diagram of an adversarial network in an automatic tumor segmentation method in CT images according to an embodiment of the present invention
  • FIG. 4 is a schematic block diagram of a structure of an electronic device according to an embodiment of the invention.
  • FIG. 1 is an automatic tumor segmentation method in CT images, including: S1, data enhancement and expansion of original image data to obtain enhanced expansion data; S2, normalization processing of enhanced expansion data to obtain normalization S3. Put the normalized data into the trained processing network to get the segmented image; S4. Do the noise reduction process on the segmented image.
  • Data enhancement and expansion of the original image data includes: enhancement and expansion of the original image data based on the principle of translation and rotation, or enhancement and expansion of the original image data based on the principle of random elastic deformation.
  • the original image data is monotonous, there is only one vertical positive data information, which is lacking for the training processing network information richness.
  • the monotonous data information leads to weak generalization performance of processing network learning, so the original image
  • the data is expanded and enhanced to obtain enhanced expanded data, thereby enhancing the generalization performance of processing network learning.
  • the process of data enhancement abides by the principle of immutability, and the specific operations that satisfy the principle of immutability are translation, rotation, and elastic deformation
  • the process of data enhancement abides by the principle of immutability, and the specific operations that satisfy the principle of immutability are translation, rotation, and elastic deformation
  • the normalization processing of the enhanced expansion data includes: normalization processing of the enhanced expansion data and the liver tumor gold standard according to the principle of linear normalization to obtain linear normalized image data; and data distribution of the linear normalized image data The normalization process of the normalized data.
  • the linear normalization is used to normalize the gray data to the [0, 225] interval, and the linear normalization formula is:
  • X norm is normalized data
  • X is enhanced extended data
  • X max and X min are the maximum and minimum values in the enhanced extended data set, respectively; in this embodiment, the input processing network
  • the liver tumor data with the pre-gray level of 0/255 binarization is 0-1 normalized as the gold standard for tumors. After dividing the data by 255, the threshold is 0.5, the value above 0.5 is set to 1, and the value below 0.5 is set to 0.
  • the data distribution is normalized.
  • the normalization of the data distribution is as follows: the liver data before the input processing network is gray-scaled in the range of 0-255, and the mean is normalized.
  • the enhanced extended data set is normalized to a normal distribution data set with a mean of 0 and a variance of 1.
  • the normalization formula of the data distribution is as follows:
  • ⁇ and ⁇ are the mean and standard deviation of the original data set, respectively.
  • the interference noise is removed by the filter, and the resulting filter is operated by the professional medical image processing software SimpleITK library.
  • the purpose is to make the final segmented image more accurate.
  • the training method of the trained processing network includes: constructing a first convolutional layer, a first modified linear unit layer, a pooling layer, a dropout layer, and a downsampling layer to form a contraction path; according to the first convolutional layer, the first modified linearity Unit layer, pooling layer, dropout layer and downsampling layer extract and encode sample noise reduction data to generate coded data; construct a second convolutional layer, a second modified linear unit layer and an upsampling layer to form an expansion path; according to the second The convolutional layer, the second modified linear unit layer and the upsampling layer decode and divide the encoded data to generate decoded data; construct a probability output layer to output the decoded data.
  • the method of making sample noise reduction data includes: performing data enhancement and expansion on the sample image data to obtain sample enhancement and expansion data; performing normalization processing on the sample enhancement and expansion data to obtain sample normalization data; and normalizing the sample data
  • the noise reduction process is performed to obtain sample noise reduction data, and the enhanced expansion of the sample image data is consistent with the enhanced expansion method of the original image data.
  • the sample enhanced expansion data is consistent with the normalized processing method of the enhanced expansion data, and the sample normalized data It is consistent with the noise reduction processing method of normalized data.
  • Constructing the first convolutional layer, the first modified linear unit layer, the pooling layer, the dropout layer, and the downsampling layer include: sequentially constructing three first 3*3 convolutional layers and constructing two first 2*2 convolutional layers Layers, and the number of feature channels of the first 3*3 convolutional layer and the first 2*2 convolutional layer constructed in sequence starts from 64 and doubles in turn; one is constructed after each first 3*3 convolutional layer Modified linear unit layer; between adjacent first 3*3 convolutional layers, between adjacent first 2*2 convolutional layers, and between adjacent first 3*3 convolutional layers and first 2*2 Pooling layers are constructed between the convolutional layers; the first first 2*2 convolutional layer is constructed after the first 3*3 convolutional layer, and the first dropout layer is constructed after the dropout layer is constructed.
  • Two first 2*2 convolutional layers; construct a second dropout layer after the second first 2*2 convolutional layer; construct a second convolutional layer, a second modified linear unit layer, and an upsampling layer include : After constructing the second dropout layer, build four second-volume base layers in sequence, each consisting of an up-sampling layer and two second 3*3 convolutional layers, and the number of second-volume base layer feature channels constructed sequentially Starting with the two first 2*2 convolutional layers, the number of feature channels is sequentially reduced by a factor of two; an upsampling layer is constructed before each second 3*3 convolutional layer.
  • the overall processing network has a U-shaped symmetric structure, which is divided into two stages of encoding and decoding.
  • the encoding stage is a feature extraction part, which is completed by a contraction path, which is mainly the left part of the U-shaped network.
  • the operation of the classic convolutional neural network is basically the same, mainly divided into 5 blocks, three blocks are composed of two consecutive 3x3 convolution operations for feature extraction (each convolution is followed by a RELU), a 2x2 is used It is composed of the maximum pooling operation of downsampling and starts with the number of feature channels of the starting image 64.
  • the decoding part is the part of feature restoration, which is mainly completed by the expansion path. It is mainly the right part of the U-shaped network.
  • sigmoid final probability output layer
  • These four blocks are all composed of One upsampling layer (implemented by 2x2 deconvolution operation) and two 3x3 convolutions (each convolution is followed by a RELU), and each time deconvolution upsampling is performed, the number of image feature channels is halved, and finally passed 1x1
  • the sigmoid layer is used to output the probability map to which each pixel belongs.
  • the contraction path and expansion path are mirror-image distributed and cascaded, which can supplement the missing information part through the mirror symmetry structure, and further enrich the network information.
  • the training method of the trained processing network further includes: constructing an adversarial network.
  • Constructing the adversarial network includes: constructing first data, the first data is referenced by the gold segmentation standard, and the binary gold standard for liver tumors and the original liver grayscale image Obtained by dot multiplication; constructing the second data, which is obtained by dot multiplying the segmented image and the original liver grayscale image; constructing a loss function, and inputting the first data and the second data into the loss function to capture pixel-level data from different levels Long and short spatial characteristics.
  • the input of the anti-network architecture can be strictly divided into two parts.
  • One part is the segmentation gold standard as the standard reference part of the input, which is obtained by multiplying the provided liver tumor binary gold standard (ground)truth and the original liver grayscale image (point multiplication), and labelling it as label_mask;
  • the final binary prediction segmentation result graph of the segmentation network it is multiplied with the original liver grayscale image (point multiplication) and recorded as output_mask.
  • the network structure of the adversarial network is similar to the coding part of the segmentation network.
  • the network takes label_mask and output_mask as inputs, and sets the loss function to MAE (Mean Absolute Error).
  • This loss function can well capture from different levels (including High-, medium-, and low-level information) pixel-level long and short spatial features, thereby enabling multi-level comparison and correction of image feature information.
  • the adversarial network loss function calculates the gap between the standard mask and the prediction mask, and combines the loss function of the segmentation network to jointly serve as the adjustment function of the final segmentation network, so as to realize the feedback adjustment of the antithesis network to the weight of the generated segmentation model Update to achieve further optimization effect.
  • the construction of the loss function uses the Dice coefficient as an evaluation, and its formula is:
  • s1, s2 are the actual value and the predicted value
  • smotth is a parameter used to increase the smoothness of the fitted curve, similar to a variable of infinitesimal, the introduction of smotth makes the function smoother.
  • the loss function of the adversarial network is MAE (Mean Absolute Error):
  • L is the total number of layers of the adversarial network
  • the loss function can well capture the long and short spatial features of pixel level from different levels (including high, middle and low level information), so as to achieve multi-level comparison and correction of image feature information.
  • An embodiment of the present application provides an automatic tumor segmentation system in a CT image, including: an enhanced expansion module for enhancing and expanding original image data to obtain enhanced expansion data; and a normalization processing module for performing enhanced expansion data Normalized processing to obtain normalized data; image segmentation module, used to input the noise reduction data into the trained processing network to obtain segmented images; noise reduction module, used to perform noise reduction processing on the segmented images.
  • An embodiment of the present application provides an electronic device. Please refer to FIG. 4.
  • the electronic device includes a memory 601, a processor 602, and a computer program stored on the memory 601 and executable on the processor 602.
  • the processor 602 executes the computer When the program is implemented, the generation method of the incremental kernel density estimator described in the foregoing embodiments of FIGS. 1 to 4 is implemented.
  • the electronic device further includes: at least one input device 603 and at least one output device 604.
  • the aforementioned memory 601, processor 602, input device 603, and output device 604 are connected via a bus 605.
  • the input device 603 may specifically be a camera, a touch panel, a physical button, a mouse, or the like.
  • the output device 604 may specifically be a display screen.
  • the memory 601 may be a high-speed random access memory (RAM, Random Access Memory) memory, or may be a non-volatile memory (non-volatile memory), such as a disk memory.
  • RAM Random Access Memory
  • non-volatile memory non-volatile memory
  • the memory 601 is used to store a set of executable program codes, and the processor 602 is coupled to the memory 601.
  • an embodiment of the present application further provides a computer-readable storage medium.
  • the computer-readable storage medium may be provided in the electronic device in each of the foregoing embodiments, and the computer-readable storage medium may be as shown in FIG. 4 above.
  • the memory 601 in the embodiment is shown.
  • a computer program is stored on the computer-readable storage medium, and when the program is executed by the processor 602, the generation method of the incremental kernel density estimator described in the foregoing method embodiment is implemented.
  • the computer storable medium may also be various media that can store program codes, such as a U disk, a mobile hard disk, a read-only memory 601 (ROM, Read-Only Memory), RAM, a magnetic disk, or an optical disk.
  • program codes such as a U disk, a mobile hard disk, a read-only memory 601 (ROM, Read-Only Memory), RAM, a magnetic disk, or an optical disk.
  • the disclosed method may be implemented in other ways.
  • the division of the modules is only a division of logical functions. In actual implementation, there may be other divisions. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not. carried out.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or modules, and may be in electrical, mechanical, or other forms.
  • modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist alone physically, or two or more modules may be integrated into one module.
  • the above integrated modules may be implemented in the form of hardware or software function modules.
  • the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may be stored in a computer-readable storage medium.
  • the technical solution of the present invention essentially or part of the contribution to the existing technology or all or part of the technical solution can be embodied in the form of a software product, the computer software product is stored in a storage medium , Including several instructions to enable a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present invention.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code .

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)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

本申请公开了一种CT图像内的肿瘤自动分割方法及系统，属于医学图像处理领域，用于分割CT图像内的肿瘤病灶区域，解决了CT图像的分割精度较低的问题，其包括：对原始图像数据进行数据增强扩充，得到增强扩充数据；对增强扩充数据进行归一化处理，得到归一化数据；将归一化数据入已训练的处理网络，得到分割图像；对分割图像做降噪处理；从而减少不同CT机扫描带来的不同原始图像的差异，提高处理网络处理结果的适用范围及精度。

Description

一种CT图像内的肿瘤自动分割方法及系统 技术领域
本发明涉及医学图像处理技术领域，尤其涉及一种CT图像内的肿瘤自动分割方法及系统。
背景技术
肝脏肿瘤是威胁人类健康的重大疾病，肝脏肿瘤的早期检测、精准测量对临床诊治与治疗有重要意义。CT(Computed Tomography，中文名称为电子计算机断层扫描)作为一种廉价高效的检测方式，已经逐渐成为临床上肝肿瘤诊断的常规手段。快速精准地从肝脏CT图像中分割出肿瘤病灶区域，不仅对手术计划制订有帮助，而且对术中肿瘤区域部位的精准定位与切除、以及术后放化疗效果的评估具有很大的实用价值。
如何快速精准地从肝脏CT图像中分割出肿瘤病灶区域，是医生及学者们研究的一大课题，目前已有基于深度学习架构的肝脏肿瘤自动分割方法的研究，如卷积神经网络(Convolutional Neural Network,CNN)、全卷积神经网络(Fully Convolutional Network,FCN)等。
但是，CNN的网络结构较为单一，且全连接层的加入使得网络整体训练参数较为庞大，计算较为复杂，信息量大，网络训练时间较长，分割精度较差。后基于此改进的FCN的整体分割精度依然较低，且基于像素的分类没有考虑像素之间的关系，缺乏空间一致性。此后又在这些经典网络架构基础上提出了一些分割网络架构，但分割精度仍有待提高。
技术问题
本发明的主要目的在于提供一种CT图像内的肿瘤自动分割方法及系统，旨在解决现有技术中对CT图像的分割精度较低的技术问题。
技术解决方案
为实现上述目的，本发明第一方面提供一种CT图像内的肿瘤自动分割方法，包括：对原始图像数据进行数据增强扩充，得到增强扩充数据；对增强扩充数据进行归一化处理，得到归一化数据；将归一化数据入已训练的处理网络， 得到分割图像；对分割图像做降噪处理。
进一步地，所述对原始图像数据进行数据增强扩充包括：基于平移旋转原则对原始图像数据进行增强扩充，或基于随机弹性形变原则对原始图像数据进行增强扩充。
进一步地，所述对增强扩充数据进行归一化处理包括：根据线性归一化原则对增强扩充数据及肝脏肿瘤金标准进行归一化处理，得到线性归一化图像数据；对所述线性归一化图像数据进行数据分布的归一化处理得到归一化数据。
进一步地，所述已训练的处理网络的训练方法包括：构建第一卷积层、第一修正线性单元层、池化层、dropout层及下采样层，形成收缩路径；根据所述第一卷积层、所述第一修正线性单元层、所述池化层、dropout层及下采样层提取并编码所述降噪数据，生成编码数据；构建第二卷积层、第二修正线性单元层及上采样层，形成膨胀路径；根据所述第二卷积层、所述第二修正线性单元层及上采样层解码并分割样本编码数据，生成解码数据；构建概率输出层，输出所述解码数据。
进一步地，所述构建第一卷积层、第一修正线性单元层、池化层、dropout层及下采样层包括：依次构建三个第一3*3卷积层，并构建两个第一2*2卷积层，且依次构建的第一3*3卷积层及第一2*2卷积层的特征通道数从64开始，依次增加一倍；在每个第一3*3卷积层后构建一个修正线性单元层；在相邻第一3*3卷积层之间、相邻第一2*2卷积层之间、及相邻的第一3*3卷积层和第一2*2卷积层之间构建池化层；在第一3*3的卷积层之后构建的第一个第一2*2的卷积层后构建第一个dropout层，在构建dropout层后构建第二个第一2*2的卷积层；在第二个第一2*2卷积层后构建第二个dropout层；所述构建第二卷积层、第二修正线性单元层及上采样层包括：在构建第二个dropout层后，依次构建四个第二卷基层，均由上采样层及两个第二3*3卷积层构成，且依次构建的第二卷基层特征通道数以第二个第一2*2卷积层为始，特征通道数 依次减少一倍；在每个第二3*3卷积层之前构建上采样层。
进一步地，所述收缩路径及所述膨胀路径之间呈镜像分布，并进行级联。
进一步地，所述已训练的处理网络的训练方法还包括：构建对抗网络，所述构建对抗网络包括：构建第一数据，所述第一数据由分割金标准作为标准参照，并由肝脏肿瘤二值金标准及原始肝脏灰度图像点乘得到；构建第二数据，所述第二数据由所述分割图像与原始肝脏灰度图像点乘得到；构建损失函数，并将第一数据及第二数据输入损失函数，捕捉来自不同层级的像素级的长短空间特征。
本发明第二方面提供一种CT图像内的肿瘤自动分割系统，包括：增强扩充模块，用于对原始图像数据进行增强扩充，得到增强扩充数据；归一化处理模块，用于对增强扩充数据进行归一化处理，得到归一化数据；图像分割模块，用于将降噪数据输入已训练的处理网络，得到分割图像；降噪模块，用于对分割图像做降噪处理。
本发明第三方面提供一种电子装置，包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时，实现上述中的任意一项所述方法。
本发明第四方面提供一种计算机可读存储介质，一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时，实现上述中的任意一项所述方法。
有益效果
本发明提供一种CT图像内的肿瘤自动分割方法，有益效果在于：通过对原始图像数据进行增强扩充，在满足变化灰度信息的鲁棒要求的基础及不影响处理网络对真实信息的处理上，能够达到丰富信息量的结果，从而提高分割模型的泛化效果，使分割模型能够适用于更广泛的数据集，提高处理网络的适用范围；由于不同病人得到的原始图像的灰度值差异较大，故通过归一化处理，能够方便分割处理网络对图像数据的处理，从而减少不同CT机扫描带来的不同原始图像的差异，提高处理网络处理结果的适用范围及精度。
附图说明
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本发明实施例CT图像内的肿瘤自动分割方法的结构示意框图；
图2为本发明实施例CT图像内的肿瘤自动分割方法中已训练的处理网络的结构示意图；
图3为本发明实施例CT图像内的肿瘤自动分割方法中对抗网络的结构示意图；
图4为本发明实施例电子装置的结构示意框图。
本发明的实施方式
为使得本发明的发明目的、特征、优点能够更加的明显和易懂，下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而非全部实施例。基于本发明中的实施例，本领域技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
请参阅图1，为一种CT图像内的肿瘤自动分割方法，包括：S1、对原始图像数据进行数据增强扩充，得到增强扩充数据；S2、对增强扩充数据进行归一化处理，得到归一化数据；S3、将归一化数据入已训练的处理网络，得到分割图像；S4、对分割图像做降噪处理。
对原始图像数据进行数据增强扩充包括：基于平移旋转原则对原始图像数据进行增强扩充，或基于随机弹性形变原则对原始图像数据进行增强扩充。
由于原始图像数据较为单调，只有一个垂直正方向的数据信息，对于已训练的处理网络信息丰富度来说较为欠缺，单调的数据信息导致处理网络学习的泛化性能较弱，因此需要对原始图像数据进行扩充及增强，得到增强扩充数据，从而增强处理网络学习的泛化性能，在本实施例中，数据增强的过程遵守不变 性原则，满足不变性原则的具体操作为平移、旋转及弹性形变，通过不变性原则处理原始数据后，不仅使原始图像的训练数据灰度信息的鲁棒性，还能达到信息量丰富的目的，又不影响处理网络对真实信息的学习，增强了处理网络的泛化性能，使处理网络能够适用于更广泛的数据集；在本实施例中，通过对原始图像数据进行平移、旋转及弹性形变的方式对原始图像数据进行增强扩充，得到增强扩充数据。
对增强扩充数据进行归一化处理包括：根据线性归一化原则对增强扩充数据及肝脏肿瘤金标准进行归一化处理，得到线性归一化图像数据；对线性归一化图像数据进行数据分布的归一化处理得到归一化数据。
由于CT设备的不同，对病人进行扫描得到的CT图像的灰度值差异较大，在本实施例中，通过对增强扩充数据及肝脏肿瘤金标准进行归一化处理，从而方便处理网络的训练，减少不同CT设备扫描带来的CT图像的灰度值差异；具体地，使用线性归一化将灰度数据归一化到[0，225]区间，线性归一化公式为：
Figure PCTCN2019121594-appb-000001
在线性归一化公式中，X norm为归一化数据，X为增强扩充数据，X max、X min分别为增强扩充数据集中的最大值和最小值；在本实施例中，对输入处理网络前灰度为0/255二值化的肝脏肿瘤数据进行0-1标准化作为肿瘤金标准，数据除以255后以0.5为阈值，高于0.5设定为1，低于0.5设定为0。
在将增强扩充数据进行线性归一化后，进行数据分布的归一化处理，数据分布的归一化处理如下：对输入处理网络前灰度在0-255区间的肝脏数据进行0均值标准化，将增强扩充数据集归一化为均值为0、方差为1的正态分布的数据集，数据分布的归一化公式如下：
Figure PCTCN2019121594-appb-000002
在数据分布的归一化公式中，μ和σ分别为原始数据集的均值和标准差。
在处理原始图像数据的过程中存在一定程度上的噪声干扰，而这些噪声的图形特征较为明显，并且在原始图像上占比例较小，因此需要对分割图像做降噪处理，在本实施例中，通过滤波器将干扰噪声去掉，造成的过滤采用专业医学图像处理软件SimpleITK库进行操作，对噪声基于形态学信息的数据统计显示当费特雷(Feret)直径小于7时，此时对应的周长及所占有像素点数量都是噪点范围内影响最小，可以滤过而不影响其他肿瘤的识别，因此以FeretDiameter＝7作为阈值，小于该值的部分将被过滤掉，从而实现降噪滤波的目的，使得最终得到的分割图像能够更加准确。
已训练的处理网络的训练方法包括：构建第一卷积层、第一修正线性单元层、池化层、dropout层及下采样层，形成收缩路径；根据第一卷积层、第一修正线性单元层、池化层、dropout层及下采样层提取并编码样本降噪数据，生成编码数据；构建第二卷积层、第二修正线性单元层及上采样层，形成膨胀路径；根据第二卷积层、第二修正线性单元层及上采样层解码并分割编码数据，生成解码数据；构建概率输出层，输出解码数据。
其中，制作样本降噪数据的方法包括：对样本图像数据进行数据增强扩充，得到样本增强扩充数据；对样本增强扩充数据进行归一化处理，得到样本归一化数据；对样本归一化数据做降噪处理，得到样本降噪数据，且样本图像数据的增强扩充与原始图像数据的增强扩充方法一致，样本增强扩充数据与增强扩充数据的的归一化处理方法一致，样本归一化数据与归一化数据的降噪处理方法一致。
构建第一卷积层、第一修正线性单元层、池化层、dropout层及下采样层包括：依次构建三个第一3*3卷积层，并构建两个第一2*2卷积层，且依次构建的第一3*3卷积层及第一2*2卷积层的特征通道数从64开始，依次增加一倍；在每个第一3*3卷积层后构建一个修正线性单元层；在相邻第一3*3卷积 层之间、相邻第一2*2卷积层之间、及相邻的第一3*3卷积层和第一2*2卷积层之间构建池化层；在第一3*3的卷积层之后构建的第一个第一2*2的卷积层后构建第一个dropout层，在构建dropout层后构建第二个第一2*2的卷积层；在第二个第一2*2卷积层后构建第二个dropout层；构建第二卷积层、第二修正线性单元层及上采样层包括：在构建第二个dropout层后，依次构建四个第二卷基层，均由上采样层及两个第二3*3卷积层构成，且依次构建的第二卷基层特征通道数以第二个第一2*2卷积层为始，特征通道数依次减少一倍；在每个第二3*3卷积层之前构建上采样层。
请参阅图2，具体地，处理网络整体呈U形对称结构，分为编码-解码两个阶段，编码阶段是特征提取部分，由收缩路径完成，主要为U形网络左侧部分，收缩路径与经典卷积神经网络操作基本相同，主要分为5个block，其中三个block由两个连续的3x3的用于特征提取的卷积操作(每个卷积后面接一个RELU)、一个2x2的用于下采样的最大池化操作组成，且以起始图像特征通道数64开始，每进行一次下采样操作后，图像的特征通道数都进行加倍；后两个block则在前面block的基础上引入了dropout层防止网络训练出现过拟合。同时，在收缩路径中，在每个block的卷积特征提取操作部分同时引用残差网络的思路，通过使用一个基于像素叠加的shortcut，将原始信息通过1x1的卷积操作和卷积操作后的信息直接连接，保持二者通道数一致，可以补充增加网络信息传递量，提高网络学习特征的能力。解码部分是特征复原的部分，主要由膨胀路径完成，主要为该U型网络的右侧部分，主要可分为4个block+1个最后的概率输出层(sigmoid)，这四个block都由一个上采样层(2x2反卷积操作实现)和两个3x3卷积(每个卷积后面接一个RELU)，且每进行一次反卷积上采样时，图像特征通道数减半，最后通过1x1的sigmoid层来对每个像素所属的概率图进行输出。
收缩路径及膨胀路径之间呈镜像分布，并进行级联，能够通过镜像对称结 构补充信息缺失部分，进一步对网络信息进行了丰富。
已训练的处理网络的训练方法还包括：构建对抗网络，构建对抗网络包括：构建第一数据，第一数据由分割金标准作为标准参照，并由肝脏肿瘤二值金标准及原始肝脏灰度图像点乘得到；构建第二数据，第二数据由分割图像与原始肝脏灰度图像点乘得到；构建损失函数，并将第一数据及第二数据输入损失函数，捕捉来自不同层级的像素级的长短空间特征。
请参阅图3，具体地，对抗网络架构的输入可严格分为两部分。一部分为分割金标准作为标准参照部分输入，由提供的肝脏肿瘤二值金标准(ground truth)和原始肝脏灰度图像相乘(点乘)得到，记为label_mask；一部分为分割网络预测部分输入，根据分割网络最终的二值预测分割结果图，与原始肝脏灰度图相乘(点乘)，记为output_mask。对抗网络的网络结构类似于分割网络的编码部分，该网络以label_mask和output_mask作为输入，设置损失函数为MAE(Mean Absolute Error，平均绝对误差)，该损失函数可以很好地捕捉来自不同层级(包括高、中、低层的信息)的像素级的长短空间特征，从而可以实现图像特征信息的多层次比对矫正。对抗网络损失函数通过计算标准掩膜和预测掩膜之间的差距，并结合分割网络的损失函数，共同作为最终分割网络的调节函数，从而实现对抗网络对已生成的分割模型权值的反馈调节更新，达到进一步优化效果。
损失函数的构建使用Dice系数作为评估，其公式为：
Figure PCTCN2019121594-appb-000003
其中，s1、s2分别为实际值与预测值，smotth是一个用于增加拟合曲线平滑度的参数，类似于无穷小的一个变量，smotth的引入使函数更加平滑。
对抗网络的损失函数为MAE(Mean Absolute Error，平均绝对误差)：
Figure PCTCN2019121594-appb-000004
其中L为对抗网络总层数，
Figure PCTCN2019121594-appb-000005
是来自于输入金标准掩膜第i层网络的特征提取图像，
Figure PCTCN2019121594-appb-000006
是来自于输入分割预测掩膜第i层网络的特征提取图像。该损失函数可以很好地捕捉来自不同层级(包括高、中、低层的信息)的像素级的长短空间特征，从而可以实现图像特征信息的多层次比对矫正。
整体损失函数为：loss＝l mae-l dice。
本申请实施例提供一种CT图像内的肿瘤自动分割系统，包括：增强扩充模块，用于对原始图像数据进行增强扩充，得到增强扩充数据；归一化处理模块，用于对增强扩充数据进行归一化处理，得到归一化数据；图像分割模块，用于将降噪数据输入已训练的处理网络，得到分割图像；降噪模块，用于对分割图像做降噪处理。
本申请实施例提供一种电子装置，请参阅图4，该电子装置包括：存储器601、处理器602及存储在存储器601上并可在处理器602上运行的计算机程序，处理器602执行该计算机程序时，实现前述附图1至附图4的实施例中描述的增量核密度估计器的生成方法。
进一步的，该电子装置还包括：至少一个输入设备603以及至少一个输出设备604。
上述存储器601、处理器602、输入设备603以及输出设备604，通过总线605连接。
其中，输入设备603具体可为摄像头、触控面板、物理按键或者鼠标等等。输出设备604具体可为显示屏。
存储器601可以是高速随机存取记忆体(RAM，Random Access Memory)存储器，也可为非不稳定的存储器(non-volatile memory)，例如磁盘存储器。存储器601用于存储一组可执行程序代码，处理器602与存储器601耦合。
进一步的，本申请实施例还提供了一种计算机可读存储介质，该计算机可读存储介质可以是设置于上述各实施例中的电子装置中，该计算机可读存储介 质可以是前述图4所示实施例中的存储器601。该计算机可读存储介质上存储有计算机程序，该程序被处理器602执行时实现前述方法实施例中描述的增量核密度估计器的生成方法。
进一步的，该计算机可存储介质还可以是U盘、移动硬盘、只读存储器601(ROM，Read-Only Memory)、RAM、磁碟或者光盘等各种可以存储程序代码的介质。
在本申请所提供的几个实施例中，应该理解到，所揭露的方法，可以通过其它的方式实现。例如，所述模块的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个模块或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或模块的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的模块可以是或者也可以不是物理上分开的，作为模块显示的部件可以是或者也可以不是物理模块，即可以位于一个地方，或者也可以分布到多个网络模块上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。
另外，在本发明各个实施例中的各功能模块可以集成在一个处理模块中，也可以是各个模块单独物理存在，也可以两个或两个以上模块集成在一个模块中。上述集成的模块既可以采用硬件的形式实现，也可以采用软件功能模块的形式实现。
所述集成的模块如果以软件功能模块的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器， 或者网络设备等)执行本发明各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
需要说明的是，对于前述的各方法实施例，为了简便描述，故将其都表述为一系列的动作组合，但是本领域技术人员应该知悉，本发明并不受所描述的动作顺序的限制，因为依据本发明，某些步骤可以采用其它顺序或者同时进行。其次，本领域技术人员也应该知悉，说明书中所描述的实施例均属于优选实施例，所涉及的动作和模块并不一定都是本发明所必须的。
在上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述的部分，可以参见其它实施例的相关描述。
以上为对本发明所提供的一种CT图像内的肿瘤自动分割方法及系统的描述，对于本领域的技术人员，依据本发明实施例的思想，在具体实施方式及应用范围上均会有改变之处，综上，本说明书内容不应理解为对本发明的限制。

Claims (10)

  1. 一种CT图像内的肿瘤自动分割方法，其特征在于，包括：
    对原始图像数据进行数据增强扩充，得到增强扩充数据；
    对增强扩充数据进行归一化处理，得到归一化数据；
    将归一化数据入已训练的处理网络，得到分割图像；
    对分割图像做降噪处理。
  2. 根据权利要求1所述的CT图像内的肿瘤自动分割方法，其特征在于，
    所述对原始图像数据进行数据增强扩充包括：
    基于平移旋转原则对原始图像数据进行增强扩充，或基于随机弹性形变原则对原始图像数据进行增强扩充。
  3. 根据权利要求1所述的CT图像内的肿瘤自动分割方法，其特征在于，
    所述对增强扩充数据进行归一化处理包括：
    根据线性归一化原则对增强扩充数据及肝脏肿瘤金标准进行归一化处理，得到线性归一化图像数据；
    对所述线性归一化图像数据进行数据分布的归一化处理得到归一化数据。
  4. 根据权利要求1所述的CT图像内的肿瘤自动分割方法，其特征在于，
    所述已训练的处理网络的训练方法包括：
    构建第一卷积层、第一修正线性单元层、池化层、dropout层及下采样层，形成收缩路径；
    根据所述第一卷积层、所述第一修正线性单元层、所述池化层、dropout层及下采样层提取并编码样本降噪数据，生成编码数据；
    构建第二卷积层、第二修正线性单元层及上采样层，形成膨胀路径；
    根据所述第二卷积层、所述第二修正线性单元层及上采样层解码并分割所述编码数据，生成解码数据；
    构建概率输出层，输出所述解码数据。
  5. 根据权利要求4所述的CT图像内的肿瘤自动分割方法，其特征在于，
    所述构建第一卷积层、第一修正线性单元层、池化层、dropout层及下采 样层包括：
    依次构建三个第一3*3卷积层，并构建两个第一2*2卷积层，且依次构建的第一3*3卷积层及第一2*2卷积层的特征通道数从64开始，依次增加一倍；
    在每个第一3*3卷积层后构建一个修正线性单元层；
    在相邻第一3*3卷积层之间、相邻第一2*2卷积层之间、及相邻的第一3*3卷积层和第一2*2卷积层之间构建池化层；
    在第一3*3的卷积层之后构建的第一个第一2*2的卷积层后构建第一个dropout层，在构建dropout层后构建第二个第一2*2的卷积层；
    在第二个第一2*2卷积层后构建第二个dropout层；
    所述构建第二卷积层、第二修正线性单元层及上采样层包括：
    在构建第二个dropout层后，依次构建四个第二卷基层，均由上采样层及两个第二3*3卷积层构成，且依次构建的第二卷基层特征通道数以第二个第一2*2卷积层为始，特征通道数依次减少一倍；
    在每个第二3*3卷积层之前构建上采样层。
  6. 根据权利要求5所述的CT图像内的肿瘤自动分割方法，其特征在于，
    所述收缩路径及所述膨胀路径之间呈镜像分布，并进行级联。
  7. 根据权利要求4所述的CT图像内的肿瘤自动分割方法，其特征在于，
    所述已训练的处理网络的训练方法还包括：
    构建对抗网络，所述构建对抗网络包括：
    构建第一数据，所述第一数据由分割金标准作为标准参照，并由肝脏肿瘤二值金标准及原始肝脏灰度图像点乘得到；
    构建第二数据，所述第二数据由所述分割图像与原始肝脏灰度图像点乘得到；
    构建损失函数，并将第一数据及第二数据输入损失函数，捕捉来自不同层级的像素级的长短空间特征。
  8. 一种CT图像内的肿瘤自动分割系统，其特征在于，包括：
    增强扩充模块，用于对原始图像数据进行增强扩充，得到增强扩充数据；
    归一化处理模块，用于对增强扩充数据进行归一化处理，得到归一化数据；
    图像分割模块，用于将降噪数据输入已训练的处理网络，得到分割图像；
    降噪模块，用于对分割图像做降噪处理。
  9. 一种电子装置，包括：存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时，实现权利要求1至7中的任意一项所述方法。
  10. 一种计算机可读存储介质，一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时，实现权利要求1至7中的任意一项所述方法。
PCT/CN2019/121594 2018-11-29 2019-11-28 一种ct图像内的肿瘤自动分割方法及系统 WO2020108562A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811440970.8A CN109754403A (zh) 2018-11-29 2018-11-29 一种ct图像内的肿瘤自动分割方法及系统
CN201811440970.8 2018-11-29

Publications (1)

Publication Number Publication Date
WO2020108562A1 true WO2020108562A1 (zh) 2020-06-04

Family

ID=66402563

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/121594 WO2020108562A1 (zh) 2018-11-29 2019-11-28 一种ct图像内的肿瘤自动分割方法及系统

Country Status (2)

Country Link
CN (1) CN109754403A (zh)
WO (1) WO2020108562A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114494266A (zh) * 2020-10-26 2022-05-13 中国人民解放军空军军医大学 一种层级空洞金字塔卷积的宫颈及周边多器官分割方法
CN116898455A (zh) * 2023-07-06 2023-10-20 湖北大学 一种基于深度学习模型的睡眠脑电信号检测方法及系统
CN117765532A (zh) * 2024-02-22 2024-03-26 中国科学院宁波材料技术与工程研究所 基于共聚焦显微图像的角膜朗格汉斯细胞分割方法和装置
CN117765532B (zh) * 2024-02-22 2024-05-31 中国科学院宁波材料技术与工程研究所 基于共聚焦显微图像的角膜朗格汉斯细胞分割方法和装置

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109754403A (zh) * 2018-11-29 2019-05-14 中国科学院深圳先进技术研究院 一种ct图像内的肿瘤自动分割方法及系统
CN110751627B (zh) * 2019-09-19 2024-01-26 上海联影智能医疗科技有限公司 图像处理方法、装置、计算机设备和存储介质
US11367181B2 (en) 2018-12-29 2022-06-21 Shanghai United Imaging Intelligence Co., Ltd. Systems and methods for ossification center detection and bone age assessment
CN110197716B (zh) * 2019-05-20 2022-05-20 广东技术师范大学 医学影像的处理方法、装置及计算机可读存储介质
CN110717060B (zh) * 2019-09-04 2023-08-18 平安科技（深圳）有限公司 图像mask的过滤方法、装置及存储介质
CN111028242A (zh) * 2019-11-27 2020-04-17 中国科学院深圳先进技术研究院 一种肿瘤自动分割系统、方法及电子设备
CN113111684B (zh) * 2020-01-10 2024-05-21 字节跳动有限公司 神经网络模型的训练方法、装置和图像处理系统
CN111652886B (zh) * 2020-05-06 2022-07-22 哈尔滨工业大学 一种基于改进U-net网络的肝肿瘤分割方法
CN111696082A (zh) * 2020-05-20 2020-09-22 平安科技（深圳）有限公司 图像分割方法、装置、电子设备及计算机可读存储介质
CN111739008B (zh) * 2020-06-23 2024-04-12 北京百度网讯科技有限公司 图像处理方法、装置、设备及可读存储介质
CN111754530B (zh) * 2020-07-02 2023-11-28 广东技术师范大学 一种前列腺超声图像分割分类方法
CN112529909A (zh) * 2020-12-08 2021-03-19 北京安德医智科技有限公司 一种基于图像补全的肿瘤图像脑区分割方法及系统
CN113705320A (zh) * 2021-05-24 2021-11-26 中国科学院深圳先进技术研究院 手术动作识别模型的训练方法、介质和设备
CN114066871A (zh) * 2021-11-19 2022-02-18 江苏科技大学 一种新冠肺炎病灶区域分割模型训练的方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160166220A1 (en) * 2014-12-12 2016-06-16 General Electric Company Method and system for defining a volume of interest in a physiological image
CN107862695A (zh) * 2017-12-06 2018-03-30 电子科技大学 一种基于全卷积神经网络的改进型图像分割训练方法
CN108109152A (zh) * 2018-01-03 2018-06-01 深圳北航新兴产业技术研究院 医学图像分类和分割方法和装置
CN108171711A (zh) * 2018-01-17 2018-06-15 深圳市唯特视科技有限公司 一种基于完全卷积网络的婴幼儿脑部磁共振图像分割方法
CN108776969A (zh) * 2018-05-24 2018-11-09 复旦大学 基于全卷积网络的乳腺超声图像肿瘤分割方法
CN109754403A (zh) * 2018-11-29 2019-05-14 中国科学院深圳先进技术研究院 一种ct图像内的肿瘤自动分割方法及系统

Family Cites Families (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106408562B (zh) * 2016-09-22 2019-04-09 华南理工大学 基于深度学习的眼底图像视网膜血管分割方法及系统
CN106683104B (zh) * 2017-01-06 2019-07-05 西北工业大学 基于集成深度卷积神经网络的前列腺磁共振图像分割方法
CN107680678B (zh) * 2017-10-18 2020-12-01 北京航空航天大学 基于多尺度卷积神经网络甲状腺超声图像结节诊断系统
CN107945204B (zh) * 2017-10-27 2021-06-25 西安电子科技大学 一种基于生成对抗网络的像素级人像抠图方法
CN108346145B (zh) * 2018-01-31 2020-08-04 浙江大学 一种病理切片中非常规细胞的识别方法
CN108492286B (zh) * 2018-03-13 2020-05-05 成都大学 一种基于双通路u型卷积神经网络的医学图像分割方法
CN108596915A (zh) * 2018-04-13 2018-09-28 深圳市未来媒体技术研究院 一种基于无标注数据的医疗图像分割方法
CN108876793A (zh) * 2018-04-13 2018-11-23 北京迈格威科技有限公司 语义分割方法、装置和系统及存储介质
CN108596884B (zh) * 2018-04-15 2021-05-18 桂林电子科技大学 一种胸部ct图像中的食管癌分割方法
CN108806793A (zh) * 2018-04-17 2018-11-13 平安科技（深圳）有限公司 病变监测方法、装置、计算机设备和存储介质
CN108537793A (zh) * 2018-04-17 2018-09-14 电子科技大学 一种基于改进的u-net网络的肺结节检测方法
CN108830912B (zh) * 2018-05-04 2021-04-16 北京航空航天大学 一种深度特征对抗式学习的交互式灰度图像着色方法
CN108629784A (zh) * 2018-05-08 2018-10-09 上海嘉奥信息科技发展有限公司 一种基于深度学习的ct图像颅内血管分割方法及系统
CN108765422A (zh) * 2018-06-13 2018-11-06 云南大学 一种视网膜图像血管自动分割方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160166220A1 (en) * 2014-12-12 2016-06-16 General Electric Company Method and system for defining a volume of interest in a physiological image
CN107862695A (zh) * 2017-12-06 2018-03-30 电子科技大学 一种基于全卷积神经网络的改进型图像分割训练方法
CN108109152A (zh) * 2018-01-03 2018-06-01 深圳北航新兴产业技术研究院 医学图像分类和分割方法和装置
CN108171711A (zh) * 2018-01-17 2018-06-15 深圳市唯特视科技有限公司 一种基于完全卷积网络的婴幼儿脑部磁共振图像分割方法
CN108776969A (zh) * 2018-05-24 2018-11-09 复旦大学 基于全卷积网络的乳腺超声图像肿瘤分割方法
CN109754403A (zh) * 2018-11-29 2019-05-14 中国科学院深圳先进技术研究院 一种ct图像内的肿瘤自动分割方法及系统

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114494266A (zh) * 2020-10-26 2022-05-13 中国人民解放军空军军医大学 一种层级空洞金字塔卷积的宫颈及周边多器官分割方法
CN114494266B (zh) * 2020-10-26 2024-05-28 中国人民解放军空军军医大学 一种层级空洞金字塔卷积的宫颈及周边多器官分割方法
CN116898455A (zh) * 2023-07-06 2023-10-20 湖北大学 一种基于深度学习模型的睡眠脑电信号检测方法及系统
CN116898455B (zh) * 2023-07-06 2024-04-16 湖北大学 一种基于深度学习模型的睡眠脑电信号检测方法及系统
CN117765532A (zh) * 2024-02-22 2024-03-26 中国科学院宁波材料技术与工程研究所 基于共聚焦显微图像的角膜朗格汉斯细胞分割方法和装置
CN117765532B (zh) * 2024-02-22 2024-05-31 中国科学院宁波材料技术与工程研究所 基于共聚焦显微图像的角膜朗格汉斯细胞分割方法和装置

Also Published As

Publication number Publication date
CN109754403A (zh) 2019-05-14

Similar Documents

Publication Publication Date Title
WO2020108562A1 (zh) 一种ct图像内的肿瘤自动分割方法及系统
US11308587B2 (en) Learning method of generative adversarial network with multiple generators for image denoising
US10706333B2 (en) Medical image analysis method, medical image analysis system and storage medium
WO2022199143A1 (zh) 一种基于u型网络的医学图像分割方法
Liang et al. MCFNet: Multi-layer concatenation fusion network for medical images fusion
CN111028242A (zh) 一种肿瘤自动分割系统、方法及电子设备
CN113012172B (zh) 一种基于AS-UNet的医学图像分割方法及系统
CN107492071A (zh) 医学图像处理方法及设备
US11816870B2 (en) Image processing method and device, neural network and training method thereof, storage medium
CN111373439B (zh) 使用cnn进行图像分割的方法
WO2021168920A1 (zh) 基于多剂量等级的低剂量图像增强方法、系统、计算机设备及存储介质
Zhang et al. A novel denoising method for low-dose CT images based on transformer and CNN
CN113538209A (zh) 一种多模态医学影像配准方法、配准系统、计算设备和存储介质
Gao et al. LEGAN: A Light and Effective Generative Adversarial Network for medical image synthesis
Yin et al. CoT-UNet++: A medical image segmentation method based on contextual Transformer and dense connection
WO2024104035A1 (zh) 基于长短期记忆自注意力模型的三维医学图像分割方法及系统
Yang et al. Low‐dose CT denoising with a high‐level feature refinement and dynamic convolution network
Ni et al. SSCA‐Net: Simultaneous Self‐and Channel‐Attention Neural Network for Multiscale Structure‐Preserving Vessel Segmentation
Yuan et al. Explore double-opponency and skin color for saliency detection
CN116912268A (zh) 一种皮肤病变图像分割方法、装置、设备及存储介质
Huang et al. HST-MRF: heterogeneous Swin transformer with multi-receptive field for medical image segmentation
CN111932486A (zh) 一种基于3d卷积神经网络的脑胶质瘤分割方法
Zhang et al. MHA‐Net: A Multibranch Hybrid Attention Network for Medical Image Segmentation
Xue Deep Generative Models for Medical Images and Beyond
Yu et al. Texture Classification Network Integrating Adaptive Wavelet Transform

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: 19890317

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 03.11.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19890317

Country of ref document: EP

Kind code of ref document: A1