WO2021097595A1 - 图像的病变区域分割方法、装置及服务器 - Google Patents

图像的病变区域分割方法、装置及服务器 Download PDF

Info

Publication number
WO2021097595A1
WO2021097595A1 PCT/CN2019/119099 CN2019119099W WO2021097595A1 WO 2021097595 A1 WO2021097595 A1 WO 2021097595A1 CN 2019119099 W CN2019119099 W CN 2019119099W WO 2021097595 A1 WO2021097595 A1 WO 2021097595A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
feature
lesion area
similarity
module
Prior art date
Application number
PCT/CN2019/119099
Other languages
English (en)
French (fr)
Inventor
王珊珊
郑海荣
祁可翰
刘新
Original Assignee
中国科学院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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Priority to PCT/CN2019/119099 priority Critical patent/WO2021097595A1/zh
Publication of WO2021097595A1 publication Critical patent/WO2021097595A1/zh

Links

Images

Classifications

    • 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
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • 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

Definitions

  • This application relates to the technical field of image recognition, and in particular to a method, device and server for segmentation of an image lesion area.
  • Magnetic resonance images of the brain can help experts effectively evaluate stroke lesions and formulate effective treatment plans.
  • segmentation of stroke lesions is usually done manually by professional radiologists on the magnetic resonance image slices slice by slice. Obviously, this is extremely time-consuming and highly subjective.
  • image segmentation methods mostly use deep learning-based methods such as convolutional neural networks to achieve automatic segmentation of stroke lesions.
  • convolutional neural networks to achieve automatic segmentation of stroke lesions.
  • people have also introduced a hollow convolution operation and a pyramid pooling structure to obtain multi-scale feature maps for accurate prediction.
  • people use long- and short-term memory-based networks to capture complex spatial background information, and use a cavity convolution model to extract rich multi-scale context information.
  • the network based on long and short-term memory and the model based on hole convolution only collect information from a few surrounding pixels, and do not extract information at a long distance, which makes the automatic segmentation method unable to make full use of the context information between all pixels. Effective processing of the different sizes and positions of the lesions in the image results in low accuracy of the segmented lesion areas.
  • One of the objectives of the embodiments of the present application is to provide a method, device, and server for segmenting a lesion area of an image, aiming to solve the problem of low accuracy of segmentation of the lesion area in the image.
  • a method for segmentation of a lesion area of an image which includes:
  • the calculation of the correlation between each pixel in the feature image by the feature similarity module to obtain a feature image containing similarity information includes:
  • the performing high-dimensional feature extraction on the image obtained by the magnetic resonance scan to generate a feature image includes:
  • the cascaded depth separable convolutional layer in the encoder performs feature extraction on the image obtained by the magnetic resonance scan to generate a feature image.
  • the predicting the lesion area in the feature image containing the similarity information and outputting the lesion area image includes:
  • the cascaded depth separable convolutional layer in the decoder predicts the location of the lesion area of the feature image containing the similarity information, and outputs the lesion area mask.
  • the high-dimensional feature extraction is performed on the image obtained by the magnetic resonance scan to generate the feature image, it includes:
  • the gradient is calculated backward to update the cascaded depth separable convolutional layer in the encoder and the decoder and the features are similar.
  • a device for segmenting an image of a lesion area including:
  • the feature extraction module is used to perform feature extraction on the image obtained by the magnetic resonance scan to generate a feature image
  • the similarity calculation module is used to calculate the correlation between each pixel in the feature image through the feature similarity module to obtain a feature image containing similarity information;
  • the lesion area segmentation module is used to predict the lesion area in the feature image containing the similarity information, and output the lesion area image.
  • the feature similarity module includes:
  • An associated information calculation unit configured to calculate the similarity between the pixel and any other pixel on the characteristic image for each pixel of the characteristic image to obtain associated information
  • the information combination unit is used to multiply the associated information of each pixel by the characteristic image to obtain a characteristic image containing similarity information.
  • the feature extraction module includes:
  • the feature extraction unit is used to perform high-dimensional feature extraction on the image obtained by the magnetic resonance scan through the cascaded depth separable convolutional layer in the encoder to generate a feature image.
  • the lesion area segmentation module includes:
  • the lesion area prediction unit is used to predict the location of the lesion area of the feature image containing the similarity information through the cascaded depth separable convolutional layer in the decoder, and output a lesion area mask.
  • the device for segmenting the lesion area of the image further includes:
  • An encoder training module configured to input preset image training data into the cascaded depth separable convolutional layer in the encoder for feature extraction to generate first feature data
  • the feature similarity training module is used to calculate the correlation between each pixel in the first feature data through the feature similarity module to obtain second feature data containing similarity information;
  • a decoder training module configured to predict the location of the lesion area of the second feature data through the cascaded depth separable convolutional layer in the decoder, and output a lesion area prediction mask;
  • the parameter update module is used to calculate gradients inversely according to the target mask corresponding to the lesion area prediction mask and the image training data, so as to update the cascaded depth separable convolution in the encoder and the decoder The parameters of the layer and the feature similarity module.
  • a server including: a memory, a processor, and a computer program stored in the memory and capable of being run on the processor.
  • the processor executes the computer program, the computer program in the first aspect is implemented.
  • Image segmentation method of lesion area When the processor executes the computer program, the computer program in the first aspect is implemented.
  • the method, device, and server for segmenting an image lesion area provided by the embodiments of the present application have the beneficial effects of: extracting features from the image obtained by magnetic resonance scanning to generate a feature image; and calculating the feature image by the feature similarity module
  • the correlation between each pixel point is used to obtain a feature image containing similarity information; the lesion area in the feature image containing the similarity information is predicted, and the lesion area image is output.
  • the feature similarity module calculates the correlation between the pixels at any two positions in the feature image with feature values, and obtains the feature image containing the similarity information between the pixels at any two positions.
  • the similarity information between each pixel in the feature image is extracted at a long distance, so that the lesion area of different positions and sizes in the image can be effectively distinguished.
  • the image of the lesion area improves the accuracy of the lesion area in the segmented image.
  • FIG. 1 is a schematic flowchart of a method for segmenting a lesion area of an image provided in Embodiment 1 of the present application;
  • FIG. 2 is a schematic flowchart of a method for segmenting a lesion area of an image provided in Embodiment 2 of the present application;
  • FIG. 3 is a schematic structural diagram of an image lesion area segmentation device provided in Embodiment 3 of the present application.
  • Fig. 4 is a schematic diagram of the structure of a server provided in the fourth embodiment of the present application.
  • the element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be construed as a limitation of the present application.
  • the specific meaning of the above terms can be understood according to specific conditions.
  • the terms “first” and “second” are only used for ease of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features.
  • the meaning of "plurality” means two or more than two, unless otherwise specifically defined.
  • FIG. 1 it is a schematic flowchart of a method for segmenting a lesion area of an image provided in Embodiment 1 of the present application.
  • This embodiment can be applied to the application scenario of segmenting the lesion area in the image obtained by the magnetic resonance scan.
  • the method can be executed by the image lesion area segmentation device, which can be a server, a smart terminal, a tablet or a PC, etc.; in this application
  • the image lesion area segmentation device is used as the execution subject for description.
  • the method specifically includes the following steps:
  • magnetic resonance is often used to scan the patient's body to obtain various parts of the body, so that the doctor can determine the diseased condition inside the patient's body by determining the lesion area in the image obtained by the magnetic resonance. Since the artificial judgment of the lesion area in the image is subjective and extremely time-consuming, the lesion area in the image can be determined by the lesion area segmentation method of the deep learning image. Specifically, in order to realize the recognition of the lesion area in the image, it is necessary to perform feature extraction on the image obtained by the magnetic resonance scan to obtain a feature image with feature values.
  • the feature extraction is performed on the image obtained by the magnetic resonance scan
  • the process of generating the feature image may be: performing feature extraction on the image obtained by the magnetic resonance scan through the deep separable convolution layer cascaded in the encoder to generate Feature image.
  • the segmentation of the lesion area of the deep learning image can be achieved by adopting a method of a symmetric encoder-decoder architecture.
  • the encoder can be used for feature extraction of the image obtained by the magnetic resonance scan. In order to obtain a multi-scale feature image containing more information, the encoder needs to perform high-dimensional feature extraction on the image to completely extract the feature information in the image.
  • the encoder uses ordinary neural network convolutional layers such as SegNet, U-Net and 2D Dense-UNet for high-dimensional feature extraction, a large number of parameters are required to participate in the calculation when performing simultaneous multi-channel convolution operations on images obtained by magnetic resonance scanning. The calculation takes too long.
  • the pre-trained encoder can use the cascaded depth separable convolutional layer to extract high-dimensional features of the image. Specifically, through the depth separable convolution layer in the encoder, each channel of the image is convolved in multi-scale dimensions, and then a common convolution with a convolution kernel size of 1x1 is used to achieve the channel dimension. Convolution operation.
  • the output data of each depth separable convolutional layer is input to the next depth separable convolutional layer for the next channel to be convolved in multi-scale dimensions, after N cascades of depth separable convolutional layer convolution
  • a feature image with a channel number of N is obtained, and high-dimensional features are extracted from the image obtained by the magnetic resonance scan.
  • S120 Calculate the correlation between each pixel in the feature image by using the feature similarity module to obtain a feature image containing similarity information
  • the association information between pixels in the feature image is extracted through a network based on long and short-term memory and a model based on hole convolution.
  • the network based on long and short-term memory and the model based on hole convolution only collect information from a few surrounding pixels, and fail to extract related information at a long distance.
  • automatic segmentation methods cannot make full use of the context between all pixels. The information effectively processes the different sizes and positions of lesions in the image.
  • a feature similarity module can be used to calculate the correlation between each pixel in the feature image to obtain a feature image containing similarity information.
  • the feature similarity module is used to calculate the correlation information between pixels at all positions on the feature image output by the encoder, not limited to obtaining relationship information from the surroundings of each pixel, but to obtain each pixel in the feature image
  • the associated information between a point and all pixels in the feature image except the pixel point can be used to extract the associated information between pixels in the feature image at a long distance.
  • the calculated association information is combined with the feature value in the feature image to obtain a feature image that contains similarity information, so that lesion areas of different positions and sizes in the image can be effectively distinguished.
  • the process of calculating the correlation between each pixel in the feature image by the feature similarity module to obtain a feature image containing similarity information may be: for each pixel of the feature image, Calculate the similarity between the pixel and any other pixel on the characteristic image to obtain associated information; multiply the associated information of each pixel by the characteristic image to obtain a characteristic image containing similarity information .
  • the similarity between each pixel point and any pixel point on the feature image other than the pixel point on the feature image is obtained through convolution calculation, and the similarity between any two positions on the feature image is obtained.
  • the characteristic image containing the similarity information is obtained by multiplying the associated information with the originally input characteristic image.
  • w(x,y) is defined as ⁇ (x) ⁇ (y), and both ⁇ and ⁇ are convolution operations.
  • the characteristic image is obtained by matrix multiplication of the results of ⁇ (x) and ⁇ (y) The similarity between any two pixels on the above.
  • w(x, y) After obtaining the similarity w(x, y) between the pixel points at any two positions on the feature image, w(x, y) is multiplied by the feature value of the original input feature image of u(y). A feature image v(x) containing similarity information is obtained. In order to avoid the disappearance of the gradient in the feature similarity model, residual error is also introduced. The feature image v(x) containing the similarity information is added to the original input feature image to obtain the final output feature image containing the similarity information.
  • the feature similarity module calculates the correlation between each pixel in the feature image to obtain the feature image containing similarity information, the size of the feature image is not changed, and there is no requirement for the size of the feature image, so The feature similarity module can be inserted into any position of the encoder or decoder for calculation without any adjustment of the encoder and decoder.
  • the lesion area in the image can be predicted based on the comprehensive and rich multi-scale context information contained in the feature image, thereby outputting the lesion area image.
  • predicting the lesion area in the feature image containing the similarity information the specific process of outputting the lesion area image may be: the depth separable convolution layer cascaded in the decoder is used to analyze the similarity The location of the lesion area of the characteristic image of the information is predicted, and the lesion area mask is output. Since the encoder-decoder has a symmetrical structure, in order to reduce the amount of parameters, the cascaded depth separable convolution layer in the encoder performs convolution operations on the multi-channels of the image in multi-scale dimensions to extract high-dimensional features to obtain feature images.
  • the decoder performs lesion area prediction to obtain a lesion area mask based on the context information and multi-scale feature information between all pixels in the feature image containing similarity information.
  • the image obtained by the MRI scan can be segmented according to the mask of the lesion area output by the decoder to obtain the lesion area image.
  • the feature image that contains the correlation information between the pixels in the feature image is extracted at a long distance to achieve accurate prediction of the lesion area in the image, and the accuracy of the segmentation image of the lesion area is improved.
  • An image lesion area segmentation method generateds a feature image by extracting features from an image obtained by a magnetic resonance scan; a feature similarity module calculates the association between each pixel in the feature image The feature image containing similarity information is obtained; the lesion area in the feature image containing the similarity information is predicted, and the lesion area image is output.
  • the feature similarity module calculates the correlation between the pixels at any two positions in the feature image with feature values, and obtains the feature image containing the similarity information between the pixels at any two positions.
  • the similarity information between each pixel in the feature image is extracted at a long distance, so that the lesion area of different positions and sizes in the image can be effectively distinguished.
  • the image of the lesion area improves the accuracy of the lesion area in the segmented image.
  • FIG. 2 is a schematic flowchart of the method for segmenting the lesion area of the image provided in the second embodiment of the present application.
  • this embodiment also provides a process of optimizing the parameters in the method for segmenting the lesion area of the image, so as to further segment the accuracy of the lesion area in the image.
  • the method specifically includes:
  • S210 Input preset image training data into the cascaded depth separable convolutional layer in the encoder for feature extraction, and generate first feature data;
  • each channel of the image training data is convolved in multi-scale dimensions, and then a common convolution with a convolution kernel size of 1x1 is used to achieve convolution in the channel dimension Product operation.
  • the output data of each depth separable convolutional layer is input to the next depth separable convolutional layer for the next channel to be convolved in multi-scale dimensions, after N cascades of depth separable convolutional layer convolution
  • the first feature data with the number of channels N is obtained, and the high-dimensional feature is extracted from the image training data obtained by the magnetic resonance scan.
  • S220 Calculate the correlation between each pixel in the first feature data by using the feature similarity module to obtain second feature data containing similarity information;
  • the feature image obtained by feature extraction of the preset image training data by the encoder is input to the feature similarity module to calculate the correlation between each pixel in the first feature data to obtain second feature data containing similarity information , Complete the training process of relevance calculation for feature similarity module. Specifically, after the first feature data is input to the feature similarity module, the similarity between each pixel point and any pixel point on the first feature data other than the pixel point on the first feature data is obtained through convolution calculation, and the first feature data is obtained. The related information of the similarity between any two pixels on the above. The second feature data containing the similarity information is obtained by multiplying the associated information with the originally input feature image.
  • S230 Predict the location of the lesion area of the second feature data through the cascaded depth separable convolutional layer in the decoder, and output a lesion area prediction mask;
  • the prediction training of the lesion area in the image can be performed according to the comprehensive and rich multi-scale context information contained in the second feature data, so as to output the lesion area prediction mask. membrane.
  • the determined lesion area mask is used as the target mask.
  • the gradient can be inversely calculated according to the target mask corresponding to the target mask and the image training data according to the lesion area prediction mask, so as to update the cascaded depth separable convolutional layer and the parameters of the feature similarity module in the encoder and decoder according to the calculated gradient . And use the cascaded depth separable convolutional layer and feature similarity module in the encoder and decoder after parameter update to segment the diseased region of the image or perform the next iterative training.
  • an embodiment of the present application also provides an image lesion area segmentation device 3, which includes:
  • the feature extraction module 301 is configured to perform feature extraction on the image obtained by the magnetic resonance scan to generate a feature image
  • the similarity calculation module 302 is configured to calculate the correlation between each pixel in the characteristic image through the characteristic similarity module to obtain a characteristic image containing similarity information;
  • the lesion area segmentation module 303 is used to predict the lesion area in the feature image containing the similarity information, and output the lesion area image.
  • the feature similarity module includes:
  • An associated information calculation unit configured to calculate the similarity between the pixel and any other pixel on the characteristic image for each pixel of the characteristic image to obtain associated information
  • the information combination unit is used to multiply the associated information of each pixel by the characteristic image to obtain a characteristic image containing similarity information.
  • the feature extraction module 301 includes:
  • the feature extraction unit is used to perform high-dimensional feature extraction on the image obtained by the magnetic resonance scan through the cascaded depth separable convolutional layer in the encoder to generate a feature image.
  • the lesion area segmentation module 303 includes:
  • the lesion area prediction unit is used to predict the location of the lesion area of the feature image containing the similarity information through the cascaded depth separable convolutional layer in the decoder, and output a lesion area mask.
  • the device further includes:
  • An encoder training module configured to input preset image training data into the cascaded depth separable convolutional layer in the encoder for feature extraction to generate first feature data
  • the feature similarity training module is used to calculate the correlation between each pixel in the first feature data through the feature similarity module to obtain second feature data containing similarity information;
  • a decoder training module configured to predict the location of the lesion area of the second feature data through the cascaded depth separable convolutional layer in the decoder, and output a lesion area prediction mask;
  • the parameter update module is used to calculate the gradient inversely according to the target mask corresponding to the lesion area prediction mask and the image training data, so as to update the cascaded depth separable convolution in the encoder and the decoder Layer and the parameters of the feature similarity module.
  • An image lesion area segmentation device provided by an embodiment of the present application generates a feature image by extracting features from an image obtained by a magnetic resonance scan; a feature similarity module calculates the association between each pixel in the feature image The feature image containing similarity information is obtained; the lesion area in the feature image containing the similarity information is predicted, and the lesion area image is output.
  • the feature similarity module calculates the correlation between the pixels at any two positions in the feature image with feature values, and obtains the feature image containing the similarity information between the pixels at any two positions.
  • the similarity information between each pixel in the feature image is extracted at a long distance, so that the lesion area of different positions and sizes in the image can be effectively distinguished.
  • the image of the lesion area improves the accuracy of the lesion area in the segmented image.
  • Fig. 4 is a schematic diagram of the structure of a server provided in the fourth embodiment of the present application.
  • the server includes: a processor 41, a memory 42, and a computer program 43 stored in the memory 42 and running on the processor 41, such as a program for a method for segmenting a lesion area of an image.
  • the processor 41 executes the computer program 43, the steps in the embodiment of the method for segmenting the lesion area of the image are implemented, for example, steps S110 to S130 shown in FIG. 1.
  • the computer program 43 may be divided into one or more modules, and the one or more modules are stored in the memory 42 and executed by the processor 41 to complete the application.
  • the one or more modules may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program 43 in the server.
  • the computer program 43 can be divided into a feature extraction module, a similarity calculation module, and a lesion area segmentation module.
  • the specific functions of each module are as follows:
  • the feature extraction module is used to perform feature extraction on the image obtained by the magnetic resonance scan to generate a feature image
  • the similarity calculation module is used to calculate the correlation between each pixel in the feature image through the feature similarity module to obtain a feature image containing similarity information;
  • the lesion area segmentation module is used to predict the lesion area in the feature image containing the similarity information, and output the lesion area image.
  • the server may include, but is not limited to, a processor 41, a memory 42, and a computer program 43 stored in the memory 42.
  • FIG. 4 is only an example of a server, and does not constitute a limitation on the server. It may include more or less components than those shown in the figure, or a combination of certain components, or different components, such as the
  • the server may also include input and output devices, network access devices, buses, and so on.
  • the processor 41 may be a central processing unit (Central Processing Unit, CPU), other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the memory 42 may be an internal storage unit of the server, such as a hard disk or memory of the server.
  • the memory 42 may also be an external storage device, such as a plug-in hard disk equipped on a server, a smart memory card (Smart Media Card, SMC), a Secure Digital (SD) card, a flash memory card (Flash Card), etc.
  • the storage 42 may also include both an internal storage unit of the server and an external storage device.
  • the memory 42 is used to store the computer program and other programs and data required by the method for segmenting the lesion area of the image.
  • the memory 42 can also be used to temporarily store data that has been output or will be output.
  • the disclosed device/terminal device and method may be implemented in other ways.
  • the device/terminal device embodiments described above are merely illustrative.
  • the division of the modules or units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units.
  • components can be combined or integrated into another system, or some features can be omitted or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • the functional units in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.
  • the integrated module/unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the present application implements all or part of the processes in the above-mentioned embodiments and methods, and can also be completed by instructing relevant hardware through a computer program.
  • the computer program can be stored in a computer-readable storage medium. When the program is executed by the processor, it can implement the steps of the foregoing method embodiments.
  • the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file, or some intermediate forms.
  • the computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory) , Random Access Memory (RAM, Random Access Memory), electrical carrier signal, telecommunications signal, and software distribution media, etc.
  • ROM Read-Only Memory
  • RAM Random Access Memory
  • electrical carrier signal telecommunications signal
  • software distribution media etc.
  • the content contained in the computer-readable medium can be appropriately added or deleted according to the requirements of the legislation and patent practice in the jurisdiction.
  • the computer-readable medium Does not include electrical carrier signals and telecommunication signals.

Landscapes

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

Abstract

一种图像的病变区域分割方法、装置(3)及服务器(4)，方法包括：对磁共振扫描获得的图像进行特征提取，生成特征图像(S110)；通过特征相似性模块计算特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像(S120)；预测包含相似性信息的特征图像中的病变区域，输出病变区域图像(S130)。解决了分割图像中病变区域的准确率低的问题。

Description

图像的病变区域分割方法、装置及服务器 技术领域
本申请涉及图像识别的技术领域，尤其涉及一种图像的病变区域分割方法、装置及服务器。
背景技术
脑部磁共振图像可帮助专家对中风病变进行有效的评估，并制定行之有效的治疗计划。在传统方法中，脑中风病变区域的分割通常由专业放射科医师在磁共振图像切片上逐切片地去手动完成，显然这极为耗时，且有很强的主观性。目前图像分割方法多采用卷积神经网络等基于深度学习的方法，实现自动化对脑卒中病变区域进行分割。此外，人们也引入了空洞卷积运算和金字塔池化结构来获得多尺度特征图，以便进行精确的预测。为了充分利用像素之间的关联信息，人们采用了基于长短期记忆的网络来捕捉复杂的空间背景信息，采用于空洞卷积的模型来提取丰富的多尺度上下文信息。
但基于长短期记忆的网络和基于空洞卷积的模型只是从几个周围的像素中收集信息，没有在一个较远的距离上提取信息，导致自动分割方法无法充分利用所有像素之间的上下文信息对图像中病变的不同大小和位置进行有效的处理，造成分割出来的病变区域的准确率低。
发明概述
技术问题
本申请实施例的目的之一在于：提供一种图像的病变区域分割方法、装置及服务器，旨在解决解决分割图像中病变区域的准确率低的问题。
问题的解决方案
技术解决方案
为解决上述技术问题，本申请实施例采用的技术方案是：
第一方面，提供了一种图像的病变区域分割方法，包括：
对磁共振扫描获得的图像进行特征提取，生成特征图像；
通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；
预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。
在一个实施示例中，所述通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像，包括：
对于所述特征图像每一像素点，计算所述像素点与所述特征图像上其它任一像素点之间的相似性得到关联信息；
将每一所述像素点的关联信息与所述特征图像相乘得到包含相似性信息的特征图像。
在一个实施示例中，所述对磁共振扫描获得的图像进行高维特征提取，生成特征图像，包括：
通过编码器中级联的深度可分离卷积层对磁共振扫描获得的图像进行特征提取，生成特征图像。
在一个实施示例中，所述预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像，包括：
通过解码器中级联的深度可分离卷积层对所述包含相似性信息的特征图像的病变区域位置进行预测，输出病变区域掩膜。
在一个实施示例中，在对磁共振扫描获得的图像进行高维特征提取，生成特征图像之前，包括：
将预设图像训练数据输入所述编码器中级联的深度可分离卷积层进行特征提取，生成第一特征数据；
通过所述特征相似性模块计算所述第一特征数据中每一像素点之间的关联性，得到包含相似性信息的第二特征数据；
通过所述解码器中级联的深度可分离卷积层对所述第二特征数据的病变区域位置进行预测，输出病变区域预测掩膜；
根据所述病变区域预测掩膜与所述图像训练数据对应的目标掩膜反向计算梯度，以更新所述编码器和所述解码器中级联的深度可分离卷积层以及所述特征相似性模块的参数。
第二方面，提供了一种图像的病变区域分割装置，包括：
特征提取模块，用于对磁共振扫描获得的图像进行特征提取，生成特征图像；
相似性计算模块，用于通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；
病变区域分割模块，用于预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。
在一个实施示例中，所述特征相似性模块包括：
关联信息计算单元，用于对于所述特征图像每一像素点，计算所述像素点与所述特征图像上其它任一像素点之间的相似性得到关联信息；
信息组合单元，用于将每一所述像素点的关联信息与所述特征图像相乘得到包含相似性信息的特征图像。
在一个实施示例中，所述特征提取模块包括：
特征提取单元，用于通过编码器中级联的深度可分离卷积层对磁共振扫描获得的图像进行高维特征提取，生成特征图像。
在一个实施示例中，所述病变区域分割模块包括：
病变区域预测单元，用于通过解码器中级联的深度可分离卷积层对所述包含相似性信息的特征图像的病变区域位置进行预测，输出病变区域掩膜。
在一个实施例中，所述图像的病变区域分割装置还包括：
编码器训练模块，用于将预设图像训练数据输入所述编码器中级联的深度可分离卷积层进行特征提取，生成第一特征数据；
特征相似性训练模块，用于通过所述特征相似性模块计算所述第一特征数据中每一像素点之间的关联性，得到包含相似性信息的第二特征数据；
解码器训练模块，用于通过所述解码器中级联的深度可分离卷积层对所述第二特征数据的病变区域位置进行预测，输出病变区域预测掩膜；
参数更新模块，用于根据所述病变区域预测掩膜与所述图像训练数据对应的目标掩膜反向计算梯度，以更新所述编码器和所述解码器中级联的深度可分离卷积层以及所述特征相似性模块的参数。
第三方面，提供一种服务器，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第一方面中图像的病变区域分割方法。
本申请实施例提供的一种图像的病变区域分割方法、装置及服务器的有益效果在于：通过对磁共振扫描获得的图像进行特征提取，生成特征图像；通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。通过特征相似性模块对具有特征值的特征图像中任意两个位置的像素点计算关联性，得到包含任意两个位置的像素点之间的相似性信息的特征图像。实现在远距离上提取了特征图像中每一像素点之间的相似性信息，从而对图像中不同位置、大小的病变区域都可以进行有效的辨别。根据包含相似性信息的特征图像的特征值和相似性信息对图像中的病变区域进行预测，充分利用包含相似性信息的特征图像中所有像素点之间的上下文信息进行病变区域预测，输出准确的病变区域图像，提高分割图像中病变区域的准确率。
发明的有益效果
对附图的简要说明
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例或示范性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申请的一些实施例，对于本领域普通技术人员来讲，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它的附图。
图1是本申请实施例一提供的图像的病变区域分割方法的流程示意图；
图2是本申请实施例二提供的图像的病变区域分割方法的流程示意图；
图3是本申请实施例三提供的图像的病变区域分割装置的结构示意图；
图4是本申请实施例四提供的服务器的结构示意图。
发明实施例
本发明的实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例 ，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本申请，并不用于限定本申请。
需说明的是，本申请的全文及上述附图中的术语“包括”以及它们任何变形，意图在于覆盖不排他的包含。例如包含一系列步骤或单元的过程、方法或系统、产品或设备没有限定于已列出的步骤或单元，而是可选地还包括没有列出的步骤或单元，或可选地还包括对于这些过程、方法、产品或设备固有的其它步骤或单元。术语“上”、“下”、“左”、“右”等指示的方位或位置关系为基于附图所示的方位或位置关系，仅是为了便于描述，而不是指示或暗示所指的装置或元件必须具有特定的方位、以特定的方位构造和操作，因此不能理解为对本申请的限制，对于本领域的普通技术人员而言，可以根据具体情况理解上述术语的具体含义。术语“第一”、“第二”仅用于便于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明技术特征的数量。“多个”的含义是两个或两个以上，除非另有明确具体的限定。
为了说明本申请所述的技术方案，以下结合具体附图及实施例进行详细说明。
实施例一
如图1所示，是本申请实施例一提供的图像的病变区域分割方法的流程示意图。本实施例可适用于分割出磁共振扫描获得的图像中病变区域的应用场景，该方法可以由图像的病变区域分割装置执行，该装置可为服务器、智能终端、平板或PC等；在本申请实施例中以图像的病变区域分割装置作为执行主体进行说明，该方法具体包括如下步骤：
S110、对磁共振扫描获得的图像进行特征提取，生成特征图像；
在医学病理诊断上，常采用磁共振扫描病人身体获得身体各个部位，从而医生可通过确定磁共振获得的图像中病变区域判断病人身体内部的病变情况。由于人为判断图像中的病变区域具有主观性且极为耗时，可通过深度学习图像的病变区域分割方法实现对图像中病变区域的确定。具体地，为实现对图像中病变区域的识别，需对磁共振扫描获得的图像进行特征提取，得到具有特征值的特征图像。
在一个实施示例中，对磁共振扫描获得的图像进行特征提取，生成特征图像的 过程可为：通过编码器中级联的深度可分离卷积层对磁共振扫描获得的图像进行特征提取，生成特征图像。具体地，深度学习图像的病变区域分割可通过采用对称编码器-解码器架构的方法实现。其中，编码器可用于对磁共振扫描获得的图像进行特征提取。为获得包含更多信息的多尺度特征图像，编码器需对该图像进行高维特征提取，以完全提取图像中的特征信息。
若编码器采用普通神经网络卷积层如SegNet、U-Net和2D Dense-UNet等进行高维特征提取，对磁共振扫描得到的图像进行同时多通道卷积操作时需要大量的参数参与计算导致计算耗时过长。为减少编码器计算过程中的参数量，预先训练好的编码器中可使用级联的深度可分离卷积层对图像进行高维特征提取。具体地，通过编码器中的深度可分离卷积层对图像的每个通道在多尺度维度上进行卷积操作，然后再使用一个卷积核大小为1x1的普通卷积来实现通道维度上的卷积操作。根据上述规则将每一深度可分离卷积层的输出数据输入下一深度可分离卷积层针对下一通道在多尺度维度上进行卷积，经过N个级联的深度可分离卷积层卷积之后得到通道数目为N的特征图像，完成从磁共振扫描获得的图像中提取高维特征。通过使用编码器中级联的深度可分离卷积层对图像进行高维特征提取在不损失性能的同时大大降低卷积神经网络的参数量。
S120、通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；
从磁共振扫描获得的图像中提取高维特征生成特征图像后，还需获取特征图像中像素点之间的关联关系即上下文信息，以根据具有丰富的多尺度上下文信息的特征图像准确的分割图像中的病变区域。相关技术中通过基于长短期记忆的网络和基于空洞卷积的模型进行特征图像中像素点之间的关联信息提取。但基于长短期记忆的网络和基于空洞卷积的模型只是从几个周围的像素中收集信息，没有在一个较远的距离上提取关联信息，导致自动分割方法无法充分利用所有像素之间的上下文信息对图像中病变的不同大小和位置进行有效的处理。
为解决这一问题，可通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像。具体地，通过特征相似性模块计算编码器输出的特征图像上所有位置上的像素点之间的关联信息，不局限于从 每一像素点的周围获取关系信息而是获取特征图像中每一像素点与特征图像中除该像素点外所有像素点之间的关联信息，实现在远距离上提取特征图像中像素点之间的关联信息。并将计算得到的关联信息与特征图像中的特征值组合，从而得到包含相似性信息的特征图像，从而对图像中不同位置、大小的病变区域都可以进行有效的辨别。
在一个实施示例中，通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像的过程可为：对于所述特征图像每一像素点，计算所述像素点与所述特征图像上其它任一像素点之间的相似性得到关联信息；将每一所述像素点的关联信息与所述特征图像相乘得到包含相似性信息的特征图像。具体地，特征图像输入特征相似性模块后通过卷积计算得到每一像素点与特征图像上除该像素点以外的任一像素点之间的相似性，得到包含特征图像上任意两个位置的像素点之间的相似性的关联信息。通过将该关联信息与原先输入的特征图像相乘得到包含相似性信息的特征图像。
可选地，考虑图像滤波中的非局部均值算法，包含相似性信息的特征图像可为：v(x)＝∑(y∈I)w(x，y)·u(y)；其中，u(y)为特征图像中任一像素点y对应的特征值；v(x)为包含相似性信息的特征图像数据；w(x，y)表示被加权的像素点y与当前像素点x的相似性。其中，w(x，y)定义为θ(x)·ψ(y)，且θ和ψ都是卷积操作，通过对θ(x)和ψ(y)的结果进行矩阵相乘得到特征图像上任意两个位置的像素点之间的相似性。在得到特征图像上任意两个位置的像素点之间的相似性w(x，y)后，w(x，y)与u(y)原输入的特征图像的特征值进行对应位置相乘，得到包含相似性信息的特征图像v(x)。且为避免特征相似性模型中的梯度消失还引入残差，将包含相似性信息的特征图像v(x)与原输入的特征图像相加得到最终输出的包含相似性信息的特征图像。
由于特征相似性模块计算所述特征图像中每一像素点之间的关联性得到包含相似性信息的特征图像的过程中并未改变特征图像的尺寸，且对特征图像的尺寸也没有要求，因此特征相似性模块可插入编码器或解码器的任意位置进行计算，而无需编码器和解码器做任何调整。
S130、预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像 。
在通过特征相似性模块计算得到包含相似性信息的特征图像后，可根据该特征图像包含的全面丰富的多尺度上下文信息进行图像中病变区域的预测，从而输出病变区域图像。
在一个实施示例中，预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像的具体过程可为：通过解码器中级联的深度可分离卷积层对所述包含相似性信息的特征图像的病变区域位置进行预测，输出病变区域掩膜。由于编码器-解码器为对称结构，为减少参数量通过编码器中级联的深度可分离卷积层对图像的多通道在多尺度维度上进行卷积操作提取高维特征得到特征图像时，需通过预先训练好的解码器中相同的级联的深度可分离卷积层对包含相似性信息的特征图像进行解码和病变区域预测，从而得到病变区域掩膜。具体地，解码器根据包含相似性信息的特征图像中所有像素点之间的上下文信息和多尺度特征信息进行病变区域预测得到病变区域掩膜。
若需具体图像，可根据解码器输出的病变区域掩膜对磁共振扫描获得的图像进行分割得到病变区域图像。通过包含在远距离上提取特征图像中像素点之间的关联信息的特征图像实现准确预测图像中的病变区域，提高分割图像中病变区域的准确率。
本申请实施例提供的一种图像的病变区域分割方法，通过对磁共振扫描获得的图像进行特征提取，生成特征图像；通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。通过特征相似性模块对具有特征值的特征图像中任意两个位置的像素点计算关联性，得到包含任意两个位置的像素点之间的相似性信息的特征图像。实现在远距离上提取了特征图像中每一像素点之间的相似性信息，从而对图像中不同位置、大小的病变区域都可以进行有效的辨别。根据包含相似性信息的特征图像的特征值和相似性信息对图像中的病变区域进行预测，充分利用包含相似性信息的特征图像中所有像素点之间的上下文信息进行病变区域预测，输出准确的病变区域图像，提高分割图像中病变区域的准确率。
实施例二
如图2所示的是本申请实施例二提供的图像的病变区域分割方法的流程示意图。在实施例一的基础上，本实施例还提供了优化图像的病变区域分割方法中的参数的过程，从而进一步分割图像中病变区域的准确率。该方法具体包括：
S210、将预设图像训练数据输入所述编码器中级联的深度可分离卷积层进行特征提取，生成第一特征数据；
为实现通过深度学习图像的病变区域分割方法实现对图像中病变区域的确定，需将预设图像训练数据输入所述编码器中级联的深度可分离卷积层进行特征提取生成第一特征数据，实现对编码器的训练。通过编码器中的深度可分离卷积层对图像训练数据的每个通道在多尺度维度上进行卷积操作，然后再使用一个卷积核大小为1x1的普通卷积来实现通道维度上的卷积操作。根据上述规则将每一深度可分离卷积层的输出数据输入下一深度可分离卷积层针对下一通道在多尺度维度上进行卷积，经过N个级联的深度可分离卷积层卷积之后得到通道数目为N的第一特征数据，完成从磁共振扫描获得的图像训练数据中提取高维特征。
S220、通过所述特征相似性模块计算所述第一特征数据中每一像素点之间的关联性，得到包含相似性信息的第二特征数据；
将通过编码器对预设图像训练数据进行特征提取得到的特征图像输入特征相似性模块计算所述第一特征数据中每一像素点之间的关联性，得到包含相似性信息的第二特征数据，完成对特征相似性模块进行关联性计算的训练过程。具体地，第一特征数据输入特征相似性模块后通过卷积计算得到每一像素点与第一特征数据上除该像素点以外的任一像素点之间的相似性，得到包含第一特征数据上任意两个位置的像素点之间的相似性的关联信息。通过将该关联信息与原先输入的特征图像相乘得到包含相似性信息的第二特征数据。
S230、通过所述解码器中级联的深度可分离卷积层对所述第二特征数据的病变区域位置进行预测，输出病变区域预测掩膜；
在通过特征相似性模块计算得到包含相似性信息的第二特征数据后，可根据该第二特征数据包含的全面丰富的多尺度上下文信息进行图像中病变区域的预测训练，从而输出病变区域预测掩膜。
S240、根据所述病变区域预测掩膜与所述图像训练数据对应的目标掩膜反向计算梯度，以更新所述编码器和所述解码器中级联的深度可分离卷积层以及所述特征相似性模块的参数。
由于预设的图像训练数据具有确定的病变区域掩膜，将确定的病变区域掩膜作为目标掩膜。可根据病变区域预测掩膜与图像训练数据对应的目标掩膜反向计算梯度，从而根据计算得到的梯度更新编码器和解码器中级联的深度可分离卷积层以及特征相似性模块的参数。并采用参数更新后的编码器和解码器中级联的深度可分离卷积层以及特征相似性模块进行图像的病变区域分割或进行下一次迭代训练。实现对编码器和解码器中级联的深度可分离卷积层以及特征相似性模块的优化，进一步提高分割图像中病变区域的准确率。并且还可通过对图像的病变区域分割方法进行多次迭代训练，以计算梯度优化编码器和解码器中级联的深度可分离卷积层以及特征相似性模块中的参数，使得图像的病变区域分割方法中模型收敛。
实施例三
如图3所示的是本申请实施例三提供的图像的病变区域分割装置。在实施例一或二的基础上，本申请实施例还提供了一种图像的病变区域分割装置3，该装置包括：
特征提取模块301，用于对磁共振扫描获得的图像进行特征提取，生成特征图像；
相似性计算模块302，用于通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；
病变区域分割模块303，用于预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。
在一个实施示例中，所述特征相似性模块包括：
关联信息计算单元，用于对于所述特征图像每一像素点，计算所述像素点与所述特征图像上其它任一像素点之间的相似性得到关联信息；
信息组合单元，用于将每一所述像素点的关联信息与所述特征图像相乘得到包含相似性信息的特征图像。
在一个实施示例中，所述特征提取模块301包括：
特征提取单元，用于通过编码器中级联的深度可分离卷积层对磁共振扫描获得的图像进行高维特征提取，生成特征图像。
在一个实施示例中，所述病变区域分割模块303包括：
病变区域预测单元，用于通过解码器中级联的深度可分离卷积层对所述包含相似性信息的特征图像的病变区域位置进行预测，输出病变区域掩膜。
在一个实施示例中，该装置还包括：
编码器训练模块，用于将预设图像训练数据输入所述编码器中级联的深度可分离卷积层进行特征提取，生成第一特征数据；
特征相似性训练模块，用于通过所述特征相似性模块计算所述第一特征数据中每一像素点之间的关联性，得到包含相似性信息的第二特征数据；
解码器训练模块，用于通过所述解码器中级联的深度可分离卷积层对所述第二特征数据的病变区域位置进行预测，输出病变区域预测掩膜；
参数更新模块，用于根据所述病变区域预测掩膜与所述图像训练数据对应的目标掩膜反向计算梯度，以更新所述编码器和所述解码器中级联的深度可分离卷积层以及所述特征相似性模块的参数。
本申请实施例提供的一种图像的病变区域分割装置，通过对磁共振扫描获得的图像进行特征提取，生成特征图像；通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。通过特征相似性模块对具有特征值的特征图像中任意两个位置的像素点计算关联性，得到包含任意两个位置的像素点之间的相似性信息的特征图像。实现在远距离上提取了特征图像中每一像素点之间的相似性信息，从而对图像中不同位置、大小的病变区域都可以进行有效的辨别。根据包含相似性信息的特征图像的特征值和相似性信息对图像中的病变区域进行预测，充分利用包含相似性信息的特征图像中所有像素点之间的上下文信息进行病变区域预测，输出准确的病变区域图像，提高分割图像中病变区域的准确率。
实施例四
图4是本申请实施例四提供的服务器的结构示意图。该服务器包括：处理器41、存储器42以及存储在所述存储器42中并可在所述处理器41上运行的计算机程序43，例如用于图像的病变区域分割方法的程序。所述处理器41执行所述计算机程序43时实现上述图像的病变区域分割方法实施例中的步骤，例如图1所示的步骤S110至S130。
示例性的，所述计算机程序43可以被分割成一个或多个模块，所述一个或者多个模块被存储在所述存储器42中，并由所述处理器41执行，以完成本申请。所述一个或多个模块可以是能够完成特定功能的一系列计算机程序指令段，该指令段用于描述所述计算机程序43在所述服务器中的执行过程。例如，所述计算机程序43可以被分割成特征提取模块、相似性计算模块和病变区域分割模块，各模块具体功能如下：
特征提取模块，用于对磁共振扫描获得的图像进行特征提取，生成特征图像；
相似性计算模块，用于通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；
病变区域分割模块，用于预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。
所述服务器可包括，但不仅限于，处理器41、存储器42以及存储在所述存储器42中的计算机程序43。本领域技术人员可以理解，图4仅仅是服务器的示例，并不构成对服务器的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如所述服务器还可以包括输入输出设备、网络接入设备、总线等。
所述处理器41可以是中央处理单元(Central Processing Unit，CPU)，还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器42可以是所述服务器的内部存储单元，例如服务器的硬盘或内存。 所述存储器42也可以是外部存储设备，例如服务器上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。进一步地，所述存储器42还可以既包括服务器的内部存储单元也包括外部存储设备。所述存储器42用于存储所述计算机程序以及图像的病变区域分割方法所需的其他程序和数据。所述存储器42还可以用于暂时地存储已经输出或者将要输出的数据。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。实施例中的各功能单元、模块可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中，上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。另外，各功能单元、模块的具体名称也只是为了便于相互区分，并不用于限制本申请的保护范围。上述系统中单元、模块的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
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或记载的部分，可以参见其它实施例的相关描述。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
在本申请所提供的实施例中，应该理解到，所揭露的装置/终端设备和方法，可以通过其它的方式实现。例如，以上所描述的装置/终端设备实施例仅仅是示意性的，例如，所述模块或单元的划分，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，例如多个单元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讯连接可以是通过一些接口，装置或单元的间接耦合或通讯连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
所述集成的模块/单元如果以软件功能单元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，本申请实现上述实施例方法中的全部或部分流程，也可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一计算机可读存储介质中，该计算机程序在被处理器执行时，可实现上述各个方法实施例的步骤。其中，所述计算机程序包括计算机程序代码，所述计算机程序代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述计算机可读介质可以包括：能够携带所述计算机程序代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、电载波信号、电信信号以及软件分发介质等。需要说明的是，所述计算机可读介质包含的内容可以根据司法管辖区内立法和专利实践的要求进行适当的增减，例如在某些司法管辖区，根据立法和专利实践，计算机可读介质不包括电载波信号和电信信号。
以上仅为本申请的可选实施例而已，并不用于限制本申请。对于本领域的技术人员来说，本申请可以有各种更改和变化。凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的权利要求范围之内。

Claims (11)

  1. 一种图像的病变区域分割方法，其特征在于，包括：
    对磁共振扫描获得的图像进行特征提取，生成特征图像；
    通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；
    预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。
  2. 根据权利要求1所述的图像的病变区域分割方法，其特征在于，所述通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像，包括：
    对于所述特征图像每一像素点，计算所述像素点与所述特征图像上其它任一像素点之间的相似性得到关联信息；
    将每一所述像素点的关联信息与所述特征图像相乘得到包含相似性信息的特征图像。
  3. 根据权利要求1或2所述的图像的病变区域分割方法，其特征在于，所述对磁共振扫描获得的图像进行高维特征提取，生成特征图像，包括：
    通过编码器中级联的深度可分离卷积层对磁共振扫描获得的图像进行特征提取，生成特征图像。
  4. 根据权利要求3所述的图像的病变区域分割方法，其特征在于，所述预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像，包括：
    通过解码器中级联的深度可分离卷积层对所述包含相似性信息的特征图像的病变区域位置进行预测，输出病变区域掩膜。
  5. 根据权利要求4所述的图像的病变区域分割方法，其特征在于，在对磁共振扫描获得的图像进行高维特征提取，生成特征图像之前，包括：
    将预设图像训练数据输入所述编码器中级联的深度可分离卷积层 进行特征提取，生成第一特征数据；
    通过所述特征相似性模块计算所述第一特征数据中每一像素点之间的关联性，得到包含相似性信息的第二特征数据；
    通过所述解码器中级联的深度可分离卷积层对所述第二特征数据的病变区域位置进行预测，输出病变区域预测掩膜；
    根据所述病变区域预测掩膜与所述图像训练数据对应的目标掩膜反向计算梯度，以更新所述编码器和所述解码器中级联的深度可分离卷积层以及所述特征相似性模块的参数。
  6. 一种图像的病变区域分割装置，其特征在于，包括：
    特征提取模块，用于对磁共振扫描获得的图像进行特征提取，生成特征图像；
    相似性计算模块，用于通过特征相似性模块计算所述特征图像中每一像素点之间的关联性，得到包含相似性信息的特征图像；
    病变区域分割模块，用于预测所述包含相似性信息的特征图像中的病变区域，输出病变区域图像。
  7. 根据权利要求6所述的图像的病变区域分割装置，其特征在于，所述特征相似性模块包括：
    关联信息计算单元，用于对于所述特征图像每一像素点，计算所述像素点与所述特征图像上其它任一像素点之间的相似性得到关联信息；
    信息组合单元，用于将每一所述像素点的关联信息与所述特征图像相乘得到包含相似性信息的特征图像。
  8. 根据权利要求6或7所述的图像的病变区域分割装置，其特征在于，所述特征提取模块包括：
    特征提取单元，用于通过编码器中级联的深度可分离卷积层对磁共振扫描获得的图像进行高维特征提取，生成特征图像。
  9. 根据权利要求8所述的图像的病变区域分割装置，其特征在于，所述病变区域分割模块包括：
    病变区域预测单元，用于通过解码器中级联的深度可分离卷积层对所述包含相似性信息的特征图像的病变区域位置进行预测，输出病变区域掩膜。
  10. 根据权利要求9所述的图像的病变区域分割装置，其特征在于，所述装置还包括：
    编码器训练模块，用于将预设图像训练数据输入所述编码器中级联的深度可分离卷积层进行特征提取，生成第一特征数据；
    特征相似性训练模块，用于通过所述特征相似性模块计算所述第一特征数据中每一像素点之间的关联性，得到包含相似性信息的第二特征数据；
    解码器训练模块，用于通过所述解码器中级联的深度可分离卷积层对所述第二特征数据的病变区域位置进行预测，输出病变区域预测掩膜；
    参数更新模块，用于根据所述病变区域预测掩膜与所述图像训练数据对应的目标掩膜反向计算梯度，以更新所述编码器和所述解码器中级联的深度可分离卷积层以及所述特征相似性模块的参数。
  11. 一种服务器，其特征在于，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5任一项所述图像的病变区域分割方法的步骤。
PCT/CN2019/119099 2019-11-18 2019-11-18 图像的病变区域分割方法、装置及服务器 WO2021097595A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/119099 WO2021097595A1 (zh) 2019-11-18 2019-11-18 图像的病变区域分割方法、装置及服务器

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/119099 WO2021097595A1 (zh) 2019-11-18 2019-11-18 图像的病变区域分割方法、装置及服务器

Publications (1)

Publication Number Publication Date
WO2021097595A1 true WO2021097595A1 (zh) 2021-05-27

Family

ID=75980288

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/119099 WO2021097595A1 (zh) 2019-11-18 2019-11-18 图像的病变区域分割方法、装置及服务器

Country Status (1)

Country Link
WO (1) WO2021097595A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114092744A (zh) * 2021-11-26 2022-02-25 山东大学 一种颈动脉超声图像斑块分类检测方法及系统
WO2024078512A1 (en) * 2022-10-10 2024-04-18 Alibaba Damo (Hangzhou) Technology Co., Ltd. Pre-analysis based image compression methods

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104517116A (zh) * 2013-09-30 2015-04-15 北京三星通信技术研究有限公司 在图像中确定目标区域的装置和方法
CN107506761A (zh) * 2017-08-30 2017-12-22 山东大学 基于显著性学习卷积神经网络的脑部图像分割方法及系统
US20180158186A1 (en) * 2016-12-07 2018-06-07 Fujitsu Limited Method and device for determining image similarity
CN109086809A (zh) * 2018-07-18 2018-12-25 安徽安大笃北信息科技有限责任公司 基于改进的监督局部线性嵌入的早期阿尔茨海默病分类方法
CN109410194A (zh) * 2018-10-19 2019-03-01 山东省计算中心（国家超级计算济南中心） 一种基于深度学习的食管癌病理图像处理方法
CN109885712A (zh) * 2019-02-12 2019-06-14 山东中医药大学 基于内容的肺结节图像检索方法及系统
CN110276777A (zh) * 2019-06-26 2019-09-24 山东浪潮人工智能研究院有限公司 一种基于深度图学习的图像分割方法及装置

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104517116A (zh) * 2013-09-30 2015-04-15 北京三星通信技术研究有限公司 在图像中确定目标区域的装置和方法
US20180158186A1 (en) * 2016-12-07 2018-06-07 Fujitsu Limited Method and device for determining image similarity
CN107506761A (zh) * 2017-08-30 2017-12-22 山东大学 基于显著性学习卷积神经网络的脑部图像分割方法及系统
CN109086809A (zh) * 2018-07-18 2018-12-25 安徽安大笃北信息科技有限责任公司 基于改进的监督局部线性嵌入的早期阿尔茨海默病分类方法
CN109410194A (zh) * 2018-10-19 2019-03-01 山东省计算中心（国家超级计算济南中心） 一种基于深度学习的食管癌病理图像处理方法
CN109885712A (zh) * 2019-02-12 2019-06-14 山东中医药大学 基于内容的肺结节图像检索方法及系统
CN110276777A (zh) * 2019-06-26 2019-09-24 山东浪潮人工智能研究院有限公司 一种基于深度图学习的图像分割方法及装置

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114092744A (zh) * 2021-11-26 2022-02-25 山东大学 一种颈动脉超声图像斑块分类检测方法及系统
CN114092744B (zh) * 2021-11-26 2024-05-17 山东大学 一种颈动脉超声图像斑块分类检测方法及系统
WO2024078512A1 (en) * 2022-10-10 2024-04-18 Alibaba Damo (Hangzhou) Technology Co., Ltd. Pre-analysis based image compression methods

Similar Documents

Publication Publication Date Title
US11170482B2 (en) Image processing method and device
US10810735B2 (en) Method and apparatus for analyzing medical image
CN110874594B (zh) 基于语义分割网络的人体外表损伤检测方法及相关设备
CN109741346B (zh) 感兴趣区域提取方法、装置、设备及存储介质
CN110599528A (zh) 一种基于神经网络的无监督三维医学图像配准方法及系统
CN111080654B (zh) 图像的病变区域分割方法、装置及服务器
WO2020125498A1 (zh) 心脏磁共振图像分割方法、装置、终端设备及存储介质
CN110276408B (zh) 3d图像的分类方法、装置、设备及存储介质
CN112132265B (zh) 模型训练方法、杯盘比确定方法、装置、设备及存储介质
CN111667459B (zh) 一种基于3d可变卷积和时序特征融合的医学征象检测方法、系统、终端及存储介质
CN114820584B (zh) 肺部病灶定位装置
WO2021097595A1 (zh) 图像的病变区域分割方法、装置及服务器
CN116228787A (zh) 图像勾画方法、装置、计算机设备、存储介质
CN114332132A (zh) 图像分割方法、装置和计算机设备
CN111968130A (zh) 脑部造影图像处理方法、装置、介质及电子设备
CN112634231A (zh) 一种图像分类方法、装置、终端设备和存储介质
CN113888566A (zh) 目标轮廓曲线确定方法、装置、电子设备以及存储介质
CN111383207A (zh) 一种肌骨超声图像处理方法、系统、装置及可读存储介质
CN111489318B (zh) 医学图像增强方法和计算机可读存储介质
CN113362345A (zh) 图像分割方法、装置、计算机设备和存储介质
Mihaylova et al. Multistage approach for automatic spleen segmentation in MRI sequences
CN115035133A (zh) 一种模型训练方法、图像分割方法及相关装置
CN112102235A (zh) 人体部位识别方法、计算机设备和存储介质
CN111986213A (zh) 一种切片图像的处理方法、训练方法、装置及存储介质
CN111192252A (zh) 一种图像分割结果优化方法、装置、智能终端及存储介质

Legal Events

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

Ref document number: 19953465

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: 19953465

Country of ref document: EP

Kind code of ref document: A1