CN110992351A - sMRI image classification method and device based on multi-input convolutional neural network - Google Patents

sMRI image classification method and device based on multi-input convolutional neural network Download PDF

Info

Publication number
CN110992351A
CN110992351A CN201911270414.5A CN201911270414A CN110992351A CN 110992351 A CN110992351 A CN 110992351A CN 201911270414 A CN201911270414 A CN 201911270414A CN 110992351 A CN110992351 A CN 110992351A
Authority
CN
China
Prior art keywords
convolutional neural
neural network
image
input
smoothing
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201911270414.5A
Other languages
Chinese (zh)
Other versions
CN110992351B (en
Inventor
李文梅
李壮壮
闫伟
张荣荣
袁媛
谢世平
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
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 Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN201911270414.5A priority Critical patent/CN110992351B/en
Publication of CN110992351A publication Critical patent/CN110992351A/en
Application granted granted Critical
Publication of CN110992351B publication Critical patent/CN110992351B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10088—Magnetic resonance imaging [MRI]
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a schizophrenia sMRI image classification method based on a multi-input convolutional neural network, which comprises the following steps: acquiring structural magnetic resonance imaging image data of schizophrenia and normal persons; preprocessing sMRI image data to respectively obtain gray matter density images which are not subjected to smoothing processing and space smoothing, and constructing an original data set; and (3) sending the original data set into a multi-input convolutional neural network model for training, finally inputting the sMRI image data into the multi-input convolutional neural network model, and finally outputting a classification result. The invention inputs the images without spatial smoothing and after spatial smoothing into the convolutional neural network at most, can make up the high-frequency information in the sMRI images lost in the spatial smoothing, and simultaneously can effectively solve the problem of low accuracy in the classification of the schizophrenia sMRI images because the noise contained in the images without the spatial smoothing can enhance the generalization capability of the convolutional neural network.

Description

sMRI image classification method and device based on multi-input convolutional neural network
Technical Field
The invention relates to a sMRI image classification method and device based on a multi-input convolutional neural network, and belongs to the technical field of image processing and artificial intelligence.
Background
Magnetic Resonance Imaging (MRI) is a relatively new medical Imaging technique that uses static and radio frequency Magnetic fields to image human tissue, and during the Imaging process, high-contrast sharp images can be obtained without using ionizing radiation or contrast agents. With the wide application of magnetic resonance imaging, higher requirements are put forward for establishing an accurate and detailed magnetic resonance image model library. The traditional establishment of a magnetic resonance image model library depends on manual retrieval and classification of professional technicians, but the retrieval efficiency is low, the precision is poor, and the workload is increased day by day.
In recent years, artificial intelligence has become the mainstream direction of science and application research, and deep learning has also made a great breakthrough as the most concerned field of artificial intelligence. Many domestic and foreign research institutes and research personnel introduce the innovative technology of deep learning into the field of intelligent medical treatment. Convolutional Neural Network (CNN) is an efficient recognition method developed in recent years, and has become a research hotspot in many scientific fields. However, the CNN is currently applied in the field of pattern classification, especially in the field of Magnetic resonance imaging (srmri) image classification of schizophrenia structures, and has the problems of less research, low recognition accuracy, slow model training speed, and the like.
Therefore, a method with higher classification precision is needed to provide reference for researchers to establish a pattern database with higher precision, so that manpower is saved and working efficiency is improved.
Disclosure of Invention
The invention provides a sMRI image classification method and device based on a multi-input convolutional neural network aiming at the problems of low accuracy, low model training speed and the like of the current sMRI image classification of schizophrenia, solves the problem of loss of high-frequency information of an sMRI image caused by space smoothness by using a multi-input CNN, and obviously improves the recognition rate of structural images for recognizing patients with first schizophrenia and normal persons.
The invention adopts the following technical scheme for solving the technical problems:
the invention discloses a schizophrenia sMRI image classification method based on a multi-input convolutional neural network, which specifically comprises the following steps:
acquiring magnetic resonance T1 image data by using a magnetic resonance scanner to obtain an sMRI (structured magnetic resonance Imaging) image;
carrying out data preprocessing on the obtained sMRI image data to obtain a three-dimensional gray matter density image which is not subjected to smoothing processing and a three-dimensional gray matter density image after space smoothing;
slicing the three-dimensional gray matter density image without smoothing treatment and the three-dimensional gray matter density image after spatial smoothing according to the Z axis, removing sliced slice images without characteristic information, and constructing an original data set without spatial smoothing treatment and spatial smoothing treatment, wherein the data set consists of data of schizophrenia and normal persons; (ii) a
Establishing a multi-input convolutional neural network model, wherein the convolutional neural network model comprises two input convolutional neural networks, and each input convolutional neural network comprises three convolutional layers and one pooling layer; fusing the features extracted by the two input convolutional neural networks through a feature fusion layer, connecting two full-connection layers behind the feature fusion layer, and adding a batch normalization layer behind each full-connection layer; initializing the number of hidden layer neurons, the size of a convolution kernel, iteration times and a learning rate;
inputting the data after the space smoothing into one input of the at most input convolutional neural network, inputting the data which is not subjected to the smoothing processing into the other input of the at most input convolutional neural network model, and training the multi-input convolutional neural network model to finally obtain an optimal multi-input convolutional neural network model; and sending the test set images into an optimal multi-input convolutional neural network model, and finally identifying images of the schizophrenia patient and normal persons.
The image without the spatial smoothing and after the spatial smoothing is input into the convolutional neural network at most, high-frequency information in the sMRI image lost in the spatial smoothing can be made up, meanwhile, the generalization capability of the convolutional neural network can be enhanced due to noise contained in the image without the spatial smoothing, and the problem of low accuracy in the classification of the schizophrenia sMRI image can be effectively solved.
Further, preprocessing the obtained srri image data to obtain a three-dimensional gray matter density image without smoothing and a three-dimensional gray matter density image after spatial smoothing, specifically comprising the following steps:
step 1, performing affine transformation on all original sMRI images to be analyzed, and then performing local nonlinear transformation correction to obtain a brain image with a standardized space;
step 2, comparing the brain structure by using the prior probability distribution of brain tissues and the gray value of the image, segmenting the brain image after space standardization into gray matter density images, and obtaining gray matter density images which are not subjected to smooth denoising;
step 3, carrying out space smoothing on the grey matter density image which is not subjected to smoothing denoising by utilizing a Gaussian smoothing kernel to obtain a grey matter density image subjected to smoothing denoising;
and 4, respectively registering the grey matter density image which is not subjected to smoothing and denoising and the grey matter density image which is subjected to smoothing and denoising by using a DARTEL algorithm to generate an optimal template, and then standardizing the MNI space to obtain a three-dimensional grey matter density image which is not subjected to smoothing and two three-dimensional NII format grey matter density images after space smoothing.
And further, utilizing a DARTEL algorithm to register and generate an optimal template, wherein the method comprises the following specific steps:
step 1, averaging all gray matter density images obtained by segmentation to obtain a gray matter average image;
step 2, taking the gray matter average image as an initial template, and respectively registering the templates of each gray matter density image by using a flow field theory;
step 3, averaging the registered images again to obtain a template of the next iteration;
and 4, repeating the process until the optimal template is obtained when the average value of two times obtains the same value.
Further, in the registration process by utilizing the DARTEL algorithm, the volume change information of each pixel point is obtainedJStored in the Jacobian determinant, which is as follows:
Figure 987681DEST_PATH_IMAGE001
wherein（x,y,z）And（x’,y’,z’）and respectively corresponding pixel points of the images before and after registration.
Further, the method further comprises: the integer label corresponding to the gray matter density image is converted to onehot label.
Further, performing Zeropadding edge filling on the sMRI data which is not subjected to the space smoothing processing and the space smoothing processing to obtain two pixel matrixes;
respectively sending the two pixel matrixes into a first convolutional layer of a multi-input convolutional neural network for extracting low-layer features, wherein the first convolutional layer comprises a hidden layer neurons, and the size of a convolutional core is m multiplied by m; inputting the low-level features into a second convolutional layer for feature extraction, wherein the second convolutional layer comprises b hidden layer neurons, and the size of a convolutional kernel is n multiplied by n; inputting the features extracted from the second layer into a third convolutional layer for extracting abstract features, wherein the third convolutional layer comprises c hidden layer neurons, and the size of a convolutional core is h multiplied by h; then, performing maximum pooling treatment on the features extracted from the convolutional layer to obtain global features, wherein the size of the posing is s multiplied by s; finally, performing feature fusion on the two input extracted global features through a feature fusion layer, inputting the fused features into two full-connection layers for classification, and adding a batch normalization layer, Dropout, behind each full-connection layer; the activation functions adopted in the whole process are all ReLU functions.
The training process of the convolutional neural network is further as follows: firstly, the convolutional neural network is propagated forwards, then the error is propagated backwards and updated, the partial derivative of the loss function to each weight is calculated, and then the weights are updated by using a gradient descent method, wherein the mathematical formula is as follows:
Figure 584010DEST_PATH_IMAGE002
wherein the error iszInput is asx i Weight parameterw i ，jFor the step size of the random gradient descent,vec()representing the vector labels where the tensor operation is converted to a vector operation.
And further, in the fourth step, the training data is sent to the convolutional neural network, the convolutional neural network is trained, a loss function value and an accuracy rate are output, and the performance of the multi-input convolutional neural network model is improved by adjusting parameters to reduce the loss function. Inputting an sMRI image to be classified into the input convolutional neural network;
calculating a classification result by utilizing a softmax function; the expression of the softmax function is as follows:
Figure 266795DEST_PATH_IMAGE003
whereinsoftmax(y) i To input firstiThe result of the classification of the individual data,y j for each of the categories there is a corresponding linear score,y i is the output of the neural network and is,y i ’ for the output after the softmax process,nis the total number of classifications.
Further, the conduction formula of the BN layer is as follows:
Figure 578959DEST_PATH_IMAGE004
in the formula (I), the compound is shown in the specification,
Figure 65304DEST_PATH_IMAGE005
in order to input the data of one batch,μ B the average value of the data input for a batch,
Figure 449012DEST_PATH_IMAGE006
is the variance of one input batch,
Figure 737036DEST_PATH_IMAGE007
training parameters for a batch normalized value
Figure 531686DEST_PATH_IMAGE008
Output datay i ，mIs the size of the batch,εin infinitesimal quantities.
Further, the adjusting parameters are specifically: increasing the number of hidden neurons; the parameter settings, convolution kernel size, and step size of the Dropout layer of each layer are adjusted.
In another aspect, the present invention provides a convolutional neural network-based mri image classification apparatus for schizophrenia, comprising: the device comprises an image acquisition module, an image preprocessing module, a multi-input convolutional neural network model establishing module, a parameter initialization setting module and a classification judging module;
the image acquisition module is a magnetic resonance scanner and is used for acquiring magnetic resonance T1 image data;
the image preprocessing module is used for preprocessing the obtained sMRI image data to obtain a three-dimensional gray matter density image which is not subjected to smoothing and a three-dimensional gray matter density image after space smoothing; slicing the three-dimensional gray matter density image which is not subjected to smoothing processing and the three-dimensional gray matter density image which is subjected to spatial smoothing and is obtained by the image preprocessing module according to the Z axis, removing the sliced slice image without characteristic information, and constructing a data set which comprises the original data set which is not subjected to spatial smoothing processing and is subjected to spatial smoothing processing, wherein the data set consists of the MRI data of schizophrenia and normal persons; (ii) a
The multi-input convolutional neural network model building module is used for building a convolutional neural network model, the convolutional neural network model comprises two input convolutional neural networks, and each input convolutional neural network comprises three convolutional layers and a pooling layer; fusing the features extracted by the two input convolutional neural networks through a feature fusion layer, connecting two full-connection layers behind the feature fusion layer, and adding a batch normalization layer behind each full-connection layer; the parameter initialization setting module is used for initializing the number of hidden layer neurons, the size of a convolution kernel, the number of iterations and the learning rate;
the classification judging module is used for respectively inputting the data after the space smoothing and the data without smoothing into two inputs of the multi-input convolutional neural network model, then training the multi-input convolutional neural network model, and finally obtaining the optimal multi-input convolutional neural network model; and sending the test set images into an optimal multi-input convolutional neural network model, and finally identifying images of the schizophrenia patient and normal persons.
By adopting the technical scheme, the invention has the following beneficial technical effects:
the method can effectively solve the problems of low accuracy in the classification of the sMRI images of schizophrenia and loss of high-frequency information of the sMRI images caused by spatial smoothness, and the accuracy of the machine learning algorithm on the classification of the sMRI images of the first schizophrenia is mostly lower than 90%, for example, the K neighbor acquisition accuracy is 85.58%, and the DT and RBF-SVM algorithm accuracy is lower than 80%. Therefore, the invention creatively introduces the CNN technology, sMRI images without space smoothing processing and space smoothing processing are input into the two convolutional neural networks of the multi-input convolutional neural network model, the characteristics extracted by the two convolutional neural networks are fused, and then the multi-input convolutional neural network model is trained, so that the images of schizophrenic patients and normal persons can be quickly, automatically and accurately identified, and the high-frequency information of the images lost in space smoothing is compensated;
under the premise of training of a large number of samples, the multi-input convolutional neural network model can achieve 99.85% of classification accuracy. Compared with the method that only the data after spatial smoothing is input to the convolutional neural network, the classification precision is improved to 99.85% from 99.41%, and the data (including noise) which is not subjected to spatial smoothing is input, so that the generalization capability of the whole multi-input convolutional neural network is enhanced;
according to the invention, a Batch Normalization (BN) layer is introduced to perform Normalization processing on the data of each Batch, so that the identification accuracy is improved;
the method of the invention provides a conduction formula of a BN layer, and the transformation reconstruction introduces learnable reconstruction parametersγ、βThe network can learn and restore the characteristic distribution to be learned by the original network, and the training speed is greatly improved by adopting the BN layer and is about 4 times of the original speed. The method of the invention adjusts parameters in the process of training the classification network structure and increases the number of hidden neurons, thereby realizing the aim of improving the accuracy.
Drawings
FIG. 1 is a flow chart of a method of an embodiment of the present invention;
FIG. 2 is a schematic diagram of a convolutional neural network model according to an embodiment of the present invention;
FIG. 3 is an original T1 image of an embodiment of the present invention;
FIG. 4 is a gray matter density image after smoothing according to an embodiment of the present invention;
FIG. 5 shows the gray matter density NII image after smoothing of the embodiment of the present invention is converted into a PNG image;
fig. 6 shows the conversion of an unsmoothed gray matter density NII image into a PNG image according to an embodiment of the present invention.
Detailed Description
The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
Example (b): the schizophrenia sMRI image classification method device based on the convolutional neural network (the flow diagram is shown in figure 1), which comprises the following steps:
acquiring magnetic resonance T1 image data by using a magnetic resonance scanner to obtain an sMRI image;
preprocessing the obtained sMRI image data to obtain a three-dimensional gray matter density image which is not subjected to smoothing and a three-dimensional gray matter density image after space smoothing; the embodiment specifically includes:
step 1, performing affine transformation on all original sMRI images to be analyzed, and then performing local nonlinear transformation correction to obtain a brain image with a standardized space;
step 2, comparing the brain structure by using the prior probability distribution of brain tissues and the gray value of the image, segmenting the brain image after space standardization into gray matter density images, and obtaining gray matter density images which are not subjected to smooth denoising;
step 3, carrying out smooth denoising on the grey matter density image which is not subjected to smooth denoising by utilizing a Gaussian smoothing kernel to obtain a grey matter density image subjected to smooth denoising;
and 4, respectively registering the grey matter density image which is not subjected to smooth denoising and the grey matter density image which is subjected to smooth denoising by using a DARTEL algorithm to generate an optimal template, and then standardizing the MNI space to obtain a three-dimensional grey matter density image which is not subjected to smooth processing and two three-dimensional NII format grey matter density images which are subjected to space smoothing (the grey matter density image which is subjected to smoothing is shown in figure 4).
The data used in this example were obtained from the brain hospital affiliated to the university of medical science of Nanjing, and there were 172 MRI images, 89 in the first schizophrenia group and 83 in the normal control group. All first schizophrenia patients were not dosed and they had no history of other neurological or serious medical conditions. The 172T 1 images (as shown in FIG. 3) were subjected to the above-described data preprocessing procedure to obtain 17253 images, which were used as training and testing sets of the convolutional neural network. Among them, 13802 images are used as a training set, and 3451 images are used as a testing set.
Slicing the three-dimensional gray matter density image without smoothing and the two three-dimensional NII or other format gray matter density images after spatial smoothing according to the Z axis, converting the sliced image into a two-dimensional PNG format (shown in figures 5 and 6 respectively) or other format (such as JPEG) image, removing slices without characteristic information in the gray image, and carrying out gray processing on the image (the smoothed gray matter density image is shown in figure 4); the image used in the experiment was 121 × 145 × 121, and the slices were taken along the Z-axis, resulting in 121 slices. 16 pictures without characteristic information in the gray matter image are removed, the residual images form an integral data set, 34506 images are totally formed, and an original large-scale data set of schizophrenia and normal persons is constructed.
Establishing a multi-input convolutional neural network model, wherein the multi-input convolutional neural network model comprises two input convolutional neural networks, each input convolutional neural network (a schematic diagram of the convolutional neural network model is shown in fig. 2) comprises three convolutional layers and a pooling layer, the features extracted by the two input convolutional neural networks are fused through a feature fusion layer, two full-connection layers are connected behind the feature fusion layer for classification, and a batch normalization layer is added behind each full-connection layer; initializing the number of hidden layer neurons and the size of a convolution kernel; sending training data into the convolutional neural network, training the convolutional neural network, and adjusting parameters until the best classification model is obtained;
and (4) sending the images of the test set into the best classification model, and finally identifying the images of the schizophrenic patients and the normal persons.
In a particular embodiment, a siemens scanner is used to acquire the srmri images.
The data preprocessing of the sMRI image specifically comprises the following steps:
step 2-1, performing affine transformation on an original image to a template image, and then performing local nonlinear transformation correction to eliminate local subtle differences to obtain a brain image with a standardized space;
2-2, segmenting a brain structure image compared with the brain image after space standardization into a gray matter density image, a white matter density image and a cerebrospinal fluid image by utilizing prior probability distribution and an image gray value of brain tissues in image segmentation;
and 2-3, smoothing and denoising the segmented gray matter density image by utilizing a Gaussian smoothing kernel.
And 2-4, registering by using a DARTEL algorithm to generate an optimal template, and then standardizing to MNI space.
Preferably, the specific steps of DARTEL for gray matter image registration are as follows:
step 1, segmenting each tested individual brain sMRI image to obtain a gray matter image;
step 2, averaging all grey matter images of the tested persons to obtain a grey matter average image;
step 3, taking the average image as an initial template, and respectively carrying out template registration on the gray matter image of each tested individual by utilizing a flow field theory;
step 4, averaging the registered images again to obtain a template of the next iteration;
and 5, repeating the process until the optimal template is obtained when the average value of two times obtains the same value.
By adopting the iterative algorithm, the influence of individuals with large brain structure difference on the analysis result can be avoided, and the image can be optimally registered.
In the registration process, the volume change information of each pixel point is storedJSpace normalization is performed in the Jacobian determinant as follows:
Figure 256059DEST_PATH_IMAGE001
wherein（x,y,z）And（x’,y’,z’）and respectively corresponding pixel points of the images before and after registration.
The image density value changes in the process of spatial standardization, which may cause non-brain components to be mistaken for gray matter, and the operations of image segmentation and registration can make the data used in the statistical analysis all come from gray matter or white matter.
On the basis of the above embodiment, in order to further improve the classification accuracy, in another embodiment, the specific steps of constructing and training the convolutional neural network are as follows:
step 1, performing graying processing on the obtained gray matter density image by utilizing a Python image processing library PIL-convert () to obtain a grayscale image;
and 2, converting the image after the gray processing into an array by using an img _ to _ array () function in the image.
Step 3, converting the two types of integer tags into onehot tags by utilizing np _ utilis.to _ category function in the keras;
and 4, independently designing and constructing an original multi-input CNN model, wherein each input of the original CNN model comprises three convolution layers, one pooling layer, two full-connection layers and a BN layer.
In order to ensure that the variable values of the non-partial order relation have no partial order and are equidistant to the origin, the method and the device convert two types of integer labels into onehot labels by adopting one-hot coding, expand the values of the discrete features to an Euclidean space, and ensure that a certain value of the discrete features corresponds to a certain point of the Euclidean space, thereby ensuring that the calculation between the features is more reasonable.
A diagram of a specific convolutional neural network model is shown in fig. 2.
Carrying out Zeropadding edge filling on MRI data which is not subjected to space smoothing processing and space smoothing processing to obtain two pixel matrixes, respectively sending the two pixel matrixes into a first convolutional layer of a multi-input convolutional neural network to extract low-layer characteristics, wherein the first convolutional layer comprises a (a is a positive integer) hidden layer neurons, and the size of a convolutional core is m multiplied by m (m is a positive odd number); inputting the low-level features into a second convolutional layer for feature extraction, wherein the second convolutional layer comprises b (b is a positive integer) hidden layer neurons, and the size of a convolutional kernel is n multiplied by n (n is a positive odd number); inputting the features extracted from the second layer into a third convolutional layer for extracting abstract features, wherein the third convolutional layer comprises c (c is a positive integer) hidden layer neurons, and the size of a convolutional kernel is h multiplied by h (h is a positive odd number); then, the maximum pooling processing is performed on the features extracted from the convolutional layer to obtain global features, and the size of the posing is s × s (s is a positive integer). And finally, inputting the global features extracted by the two inputs into two full-connection layers for classification, and adding a BN layer, namely Dropout, behind each full-connection layer. The activation functions adopted in the whole process are all ReLU functions, and then training, parameter adjustment and storage of the best classification network model are carried out. A Linear rectification function (called a modified Linear Unit, ReLU) is an activation function (activation function) commonly used in artificial neural networks, and generally refers to a nonlinear function represented by a ramp function and its variants, which is not introduced in the prior art.
The training process of the convolutional neural network is as follows:
firstly, forward propagation of the convolutional neural network is carried out, then, error backward propagation updating is carried out, the partial derivative of the loss function to each weight is calculated, then, the weight is updated by using a gradient descent method, and the mathematical formula is as follows:
Figure 663032DEST_PATH_IMAGE002
wherein the error iszInput is asx i Weight parameterw i ，jFor the step size of the random gradient descent,vec()for vectorization, tensor operations are converted into vector operations.
The adjusted parameters include the number of network layers, the size of a convolution kernel, the number of hidden layer neurons, the number of full connection layers, a Dropout value, and the like. Because the parameters are updated in the training process of the lower layer network, the distribution of the input data of the rear layer is changed, and the training precision is reduced, the BN layer is introduced to carry out normalization processing on the data of each batch so as to improve the accuracy. The BN layer solves the problem that the data distribution of the middle layer is changed in the deep learning training process, so that the gradient is prevented from disappearing or exploding, and the training speed is accelerated. The Dropout layer is to avoid overfitting of the network.
The conduction formula of the BN layer is as follows:
Figure 304098DEST_PATH_IMAGE004
in the formula (I), the compound is shown in the specification,
Figure 771114DEST_PATH_IMAGE005
in order to input the data of one batch,μ B the average value of the data input for a batch,
Figure 248362DEST_PATH_IMAGE006
is the variance of one input batch,
Figure 691982DEST_PATH_IMAGE009
training parameters for a batch normalized valueγAndβ，y i in order to output the data, the data is output,min order to be the size of the batch,εin infinitesimal quantities. The transformation reconstruction introduces this learnable reconstruction parameterγAndβthe network can learn and restore the characteristic distribution to be learned by the original network, and the training speed is greatly improved by adopting the BN layer and is about 4 times of the original speed. And finally, classifying by using a softmax classifier, wherein the softmax expression is as follows:
Figure 157861DEST_PATH_IMAGE003
whereiny i Is the output of the neural network and is,y i ’ for the output after the softmax process,nto be the total number of classifications,softmax(y) i to input firstiThe result of the classification of the individual data,y j a linear score for each category. The Softmax formula can calculate the normalized probability of the images of the schizophrenia patients and normal persons, and the classification result with higher probability is selected.
Training an initial model by adopting a training sample set, and performing parameter adjustment in the classification network structure, wherein the number of hidden neurons is increased on one hand, so that the accuracy is improved. And on the other hand, the hyper-parameters such as parameter setting, convolution kernel size, step length and the like of the Dropout layer of each layer are adjusted, initialization parameters, convolution kernel size and the like are set until the best classification model is obtained and stored.
TABLE 1 comparison of classification accuracy of the multi-input convolutional neural network model provided by the present invention with other algorithms
Figure 28734DEST_PATH_IMAGE010
The multi-input convolutional neural network model provided by the invention is compared with classification experiment results of other algorithms in a table 1, and the table 1 shows that the multi-input convolutional neural network model can effectively solve the problems of low accuracy in classifying schizophrenia sMRI images and loss of sMRI image high-frequency information caused by spatial smoothness, the accuracy of machine learning algorithm for classifying first-onset schizophrenia sMRI images is mostly lower than 90%, for example, the accuracy of K neighbor acquisition is 85.58%, and the accuracy of DT and RBF-SVM algorithms is lower than 80%, so that the creatively introduced convolutional neural network technology realizes rapid, accurate and automatic recognition of images of schizophrenia patients and normal persons, and makes up the problem of image high-frequency information loss caused by spatial smoothness; under the premise of training a large number of samples, the classification precision of the multi-input convolutional neural network can reach 99.85%. Compared with the method that only the data after spatial smoothing is input to the convolutional neural network, the classification precision is improved by 0.45%, and the data (including noise) which is not subjected to spatial smoothing is input, so that the generalization capability of the whole multi-input convolutional neural network is enhanced.
Example (b): schizophrenia sMRI image classification device based on multi-input convolutional neural network comprises: the device comprises an image acquisition module, an image preprocessing module, a multi-input convolutional neural network model establishing module, a parameter initialization setting module and a classification judging module;
the image acquisition module is a magnetic resonance scanner and is used for acquiring magnetic resonance T1 image data to obtain an sMRI image;
the image preprocessing module is used for preprocessing the obtained sMRI image data to obtain a three-dimensional gray matter density image which is not subjected to smoothing and a three-dimensional gray matter density image after space smoothing; slicing the three-dimensional gray matter density image which is not subjected to smoothing processing and the three-dimensional gray matter density image which is subjected to spatial smoothing and is obtained by the image preprocessing module according to the Z axis, removing the sliced slice image without characteristic information, and constructing a data set which comprises original data which are not subjected to spatial smoothing processing and spatial smoothing processing, wherein the data set consists of sMRI data of schizophrenia and normal persons;
the multi-input convolutional neural network model building module is used for building a convolutional neural network model, the convolutional neural network model comprises two input convolutional neural networks, and each input convolutional neural network comprises three convolutional layers and a pooling layer; fusing the features extracted by the two input convolutional neural networks through a feature fusion layer, connecting two full-connection layers behind the feature fusion layer, and adding a batch normalization layer behind each full-connection layer;
the parameter initialization setting module is used for initializing the number of hidden layer neurons, the size of a convolution kernel, the number of iterations and the learning rate;
the classification discrimination module inputs the data after the space smoothing into one input of the at most input convolutional neural network, inputs the data which is not subjected to the smoothing processing into the other input of the at most input convolutional neural network model, trains the multi-input convolutional neural network model, and finally obtains the optimal multi-input convolutional neural network model; and sending the test set images into an optimal multi-input convolutional neural network model, and finally identifying images of the schizophrenia patient and normal persons.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
While the present invention has been described with reference to the embodiments shown in the drawings, the present invention is not limited to the embodiments, which are illustrative and not restrictive, and it will be apparent to those skilled in the art that various changes and modifications can be made therein without departing from the spirit and scope of the invention as defined in the appended claims.

Claims (10)

1. The schizophrenia sMRI image classification method based on the multi-input convolutional neural network is characterized by comprising the following steps:
acquiring magnetic resonance T1 image data by using a magnetic resonance scanner to obtain an sMRI image;
carrying out data preprocessing on the obtained sMRI image data to obtain a three-dimensional gray matter density image which is not subjected to smoothing processing and a three-dimensional gray matter density image after space smoothing;
slicing the three-dimensional gray matter density image without smoothing treatment and the three-dimensional gray matter density image after spatial smoothing according to the Z axis respectively, removing slice images without characteristic information after slicing, and constructing an original data set without spatial smoothing treatment and spatial smoothing treatment, wherein the data set is composed of data of schizophrenia and normal persons;
establishing a multi-input convolutional neural network model, wherein the convolutional neural network model comprises two input convolutional neural networks, and each input convolutional neural network comprises three convolutional layers and one pooling layer; fusing the features extracted by the two input convolutional neural networks through a feature fusion layer, connecting two full-connection layers behind the feature fusion layer, and adding a batch normalization layer behind each full-connection layer; initializing the number of hidden layer neurons, the size of a convolution kernel, iteration times and a learning rate;
respectively inputting sMRI data after space smoothing and without smoothing into two input convolutional neural networks, training a multi-input convolutional neural network model, and finally obtaining an optimal multi-input convolutional neural network model; and sending the test set images into an optimal multi-input convolutional neural network model, and finally identifying images of the schizophrenia patient and normal persons.
2. The sMRI image classification method for schizophrenia based on a multi-input convolutional neural network as claimed in claim 1, characterized in that the obtained sMRI image data is preprocessed to obtain a three-dimensional gray matter density image without smoothing and a three-dimensional gray matter density image after spatial smoothing, and the method specifically comprises the following steps:
step 1, performing affine transformation on all original sMRI images to be analyzed, and then performing local nonlinear transformation correction to obtain a brain image with a standardized space;
step 2, comparing the brain structure by using the prior probability distribution of brain tissues and the gray value of the image, segmenting the brain image after space standardization into gray matter density images, and obtaining gray matter density images which are not subjected to smooth denoising;
step 3, carrying out space smoothing on the grey matter density image which is not subjected to smoothing denoising by utilizing a Gaussian smoothing kernel to obtain a grey matter density image subjected to smoothing denoising;
and 4, respectively registering the grey matter density image which is not subjected to smoothing and denoising and the grey matter density image which is subjected to smoothing and denoising by using a DARTEL algorithm to generate an optimal template, and then standardizing the MNI space to obtain a three-dimensional grey matter density image which is not subjected to smoothing and two three-dimensional NII format grey matter density images after space smoothing.
3. The method for classifying sMRI images based on multi-input convolutional neural network as claimed in claim 2, wherein the step of generating the optimal template by using DARTEL algorithm registration comprises the following steps:
step 1, averaging all gray matter density images obtained by segmentation to obtain a gray matter average image;
step 2, taking the gray matter average image as an initial template, and respectively registering the templates of each gray matter density image by using a flow field theory;
step 3, averaging the registered images again to obtain a template of the next iteration;
and 4, repeating the process until the optimal template is obtained when the average value of two times obtains the same value.
4. The method of classifying sMRI images for schizophrenia based on multi-input convolutional neural network as claimed in claim 2, wherein the volume change information of each pixel point is determined during the registration process using DARTEL algorithmJStored in the Jacobian determinant, which is as follows:
Figure 119994DEST_PATH_IMAGE001
wherein（x,y,z）And（x’,y’,z’）and respectively corresponding pixel points of the images before and after registration.
5. The method of classifying schizophrenia sMRI image based on multiple-input convolutional neural network according to claim 1, wherein the method further comprises: the integer label corresponding to the gray matter density image is converted to onehot label.
6. The method for classifying the sMRI image of schizophrenia based on the multi-input convolutional neural network as claimed in claim 1, wherein the spatially smoothed data is input to one input of the multi-input convolutional neural network, the data which is not smoothed is input to the other input of the multi-input convolutional neural network model, and then training is performed, comprising the following steps:
carrying out Zeropadding edge filling on the sMRI data which is not subjected to space smoothing processing and space smoothing processing to obtain two pixel matrixes;
respectively sending the two pixel matrixes into a first convolutional layer of two input convolutional neural networks for extracting low-layer features, wherein the first convolutional layer comprises a hidden layer neurons, and the size of a convolutional core is m multiplied by m; inputting the low-level features into a second convolutional layer for feature extraction, wherein the second convolutional layer comprises b hidden layer neurons, and the size of a convolutional kernel is n multiplied by n; inputting the features extracted from the second layer into a third convolutional layer for extracting abstract features, wherein the third convolutional layer comprises c hidden layer neurons, and the size of a convolutional core is h multiplied by h; then, performing maximum pooling treatment on the features extracted from the convolutional layer to obtain global features, wherein the size of the posing is s multiplied by s; finally, performing feature fusion on the global features extracted from the two input layers through a feature fusion layer, inputting the fused features into the two full-connection layers for classification, and adding a batch normalization layer Dropout behind each full-connection layer; the activation functions adopted in the whole process are all ReLU functions.
7. The method for classifying the sMRI images of schizophrenia based on the multi-input convolutional neural network as claimed in claim 1, wherein the training process of the convolutional neural network is as follows: firstly, the convolutional neural network is propagated forwards, then the error is propagated backwards and updated, the partial derivative of the loss function to each weight is calculated, and then the weight is updated by using a gradient descent method, wherein the partial derivative is updatediThe layer parameters, the mathematical formula of which is as follows:
Figure 828187DEST_PATH_IMAGE002
wherein the error iszInput is asx i Weight parameterw i ，jFor the step size of the random gradient descent,vec()representing the vector labels where the tensor operation is converted to a vector operation.
8. The method for classifying schizophrenia sMRI images based on a multi-input convolutional neural network as claimed in claim 1, wherein in the fourth step, training data is sent to the convolutional neural network, the convolutional neural network is trained, loss function values and accuracy are output, and the performance of a model of the multi-input convolutional neural network is improved by adjusting parameters to reduce the loss function values; calculating a classification result by utilizing a softmax function;
the softmax function expression is as follows:
Figure 814860DEST_PATH_IMAGE003
whereinsoftmax(y) i To input firstiThe result of the classification of the individual data,y i is the output of the neural network and is,y i ’ for the output after the softmax process,nto be the total number of classifications,y j a linear score for each category.
9. The method for classifying schizophrenia srmri image based on multiple-input convolutional neural network according to claim 1, wherein the conduction formula of the BN layer is as follows:
Figure 857771DEST_PATH_IMAGE004
，
in the formula (I), the compound is shown in the specification,x i in order to input the data of one batch,
Figure 650409DEST_PATH_IMAGE005
the average value of the data input for a batch,
Figure 213108DEST_PATH_IMAGE006
is the variance of one input batch,
Figure 931534DEST_PATH_IMAGE007
is a value normalized for one batch,βandγin order to train the parameters of the device,y i in order to output the data, the data is output,mis the size of the batch or batches,εis an infinitesimal quantity.
10. Schizophrenia sMRI image classification device based on multi-input convolutional neural network is characterized by comprising: the device comprises an image acquisition module, an image preprocessing module, a multi-input convolutional neural network model establishing module, a parameter initialization setting module and a classification judging module;
the image acquisition module is a magnetic resonance scanner and is used for acquiring magnetic resonance T1 image data to obtain an sMRI image;
the image preprocessing module is used for preprocessing the obtained sMRI image data to obtain a three-dimensional gray matter density image which is not subjected to smoothing processing and a three-dimensional gray matter density image after space smoothing; slicing the three-dimensional gray matter density image which is not subjected to smoothing processing and the three-dimensional gray matter density image which is subjected to spatial smoothing and is obtained by the image preprocessing module according to the Z axis, removing sliced slice images without characteristic information after slicing, and constructing a data set which comprises original data which are not subjected to spatial smoothing processing and spatial smoothing processing, wherein the data set consists of data of schizophrenia and normal persons;
the multi-input convolutional neural network model building module is used for building a convolutional neural network model, the convolutional neural network model comprises two input convolutional neural networks, and each input convolutional neural network comprises three convolutional layers and a pooling layer; fusing the features extracted by the two input convolutional neural networks through a feature fusion layer, connecting two full-connection layers behind the feature fusion layer, and adding a batch normalization layer behind each full-connection layer;
the parameter initialization setting module is used for initializing the number of hidden layer neurons, the size of a convolution kernel, the number of iterations and the learning rate;
the classification discrimination module is used for inputting the data after the space smoothing into one input of the multi-input convolutional neural network model, inputting the data which is not subjected to the smoothing into the other input of the multi-input convolutional neural network model, and then training to finally obtain the optimal multi-input convolutional neural network model; and sending the test set images into an optimal multi-input convolutional neural network model, and finally identifying images of the schizophrenia patient and normal persons.
CN201911270414.5A 2019-12-12 2019-12-12 sMRI image classification method and device based on multi-input convolution neural network Active CN110992351B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911270414.5A CN110992351B (en) 2019-12-12 2019-12-12 sMRI image classification method and device based on multi-input convolution neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911270414.5A CN110992351B (en) 2019-12-12 2019-12-12 sMRI image classification method and device based on multi-input convolution neural network

Publications (2)

Publication Number Publication Date
CN110992351A true CN110992351A (en) 2020-04-10
CN110992351B CN110992351B (en) 2022-08-16

Family

ID=70092606

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911270414.5A Active CN110992351B (en) 2019-12-12 2019-12-12 sMRI image classification method and device based on multi-input convolution neural network

Country Status (1)

Country Link
CN (1) CN110992351B (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111932523A (en) * 2020-08-14 2020-11-13 中国科学院心理研究所 Gender classifier based on brain imaging big data deep learning
CN112734726A (en) * 2020-09-29 2021-04-30 首都医科大学附属北京天坛医院 Typing method, device and equipment for angiography
CN112837274A (en) * 2021-01-13 2021-05-25 南京工业大学 Classification and identification method based on multi-mode multi-site data fusion
CN113197578A (en) * 2021-05-07 2021-08-03 天津医科大学 Schizophrenia classification method and system based on multi-center model
CN113206808A (en) * 2021-04-01 2021-08-03 中国电子科技集团公司第五十二研究所 Channel coding blind identification method based on one-dimensional multi-input convolutional neural network
CN114693964A (en) * 2022-03-09 2022-07-01 电子科技大学 MRI (magnetic resonance imaging) data feature extraction and classification identification method based on artificial neural network
CN114926396A (en) * 2022-04-13 2022-08-19 四川大学华西医院 Mental disorder magnetic resonance image preliminary screening model construction method
CN115409843A (en) * 2022-11-02 2022-11-29 长春理工大学 Brain nerve image feature extraction method based on scale equalization coupling convolution architecture
CN116342603A (en) * 2023-05-30 2023-06-27 杭州脉流科技有限公司 Method for obtaining arterial input function

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1815429A2 (en) * 2004-11-22 2007-08-08 Eastman Kodak Company Detecting and classifying lesions in ultrasound images
CN106096616A (en) * 2016-06-08 2016-11-09 四川大学华西医院 A kind of nuclear magnetic resonance image feature extraction based on degree of depth study and sorting technique
CN109222972A (en) * 2018-09-11 2019-01-18 华南理工大学 A kind of full brain data classification method of fMRI based on deep learning

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1815429A2 (en) * 2004-11-22 2007-08-08 Eastman Kodak Company Detecting and classifying lesions in ultrasound images
CN106096616A (en) * 2016-06-08 2016-11-09 四川大学华西医院 A kind of nuclear magnetic resonance image feature extraction based on degree of depth study and sorting technique
CN109222972A (en) * 2018-09-11 2019-01-18 华南理工大学 A kind of full brain data classification method of fMRI based on deep learning

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111932523B (en) * 2020-08-14 2023-02-10 中国科学院心理研究所 Gender classifier based on brain imaging big data deep learning
CN111932523A (en) * 2020-08-14 2020-11-13 中国科学院心理研究所 Gender classifier based on brain imaging big data deep learning
CN112734726A (en) * 2020-09-29 2021-04-30 首都医科大学附属北京天坛医院 Typing method, device and equipment for angiography
CN112734726B (en) * 2020-09-29 2024-02-02 首都医科大学附属北京天坛医院 Angiography typing method, angiography typing device and angiography typing equipment
CN112837274A (en) * 2021-01-13 2021-05-25 南京工业大学 Classification and identification method based on multi-mode multi-site data fusion
CN112837274B (en) * 2021-01-13 2023-07-07 南京工业大学 Classification recognition method based on multi-mode multi-site data fusion
CN113206808A (en) * 2021-04-01 2021-08-03 中国电子科技集团公司第五十二研究所 Channel coding blind identification method based on one-dimensional multi-input convolutional neural network
CN113206808B (en) * 2021-04-01 2022-06-14 中国电子科技集团公司第五十二研究所 Channel coding blind identification method based on one-dimensional multi-input convolutional neural network
CN113197578A (en) * 2021-05-07 2021-08-03 天津医科大学 Schizophrenia classification method and system based on multi-center model
CN114693964B (en) * 2022-03-09 2023-04-07 电子科技大学 MRI data feature extraction and classification identification method based on artificial neural network
CN114693964A (en) * 2022-03-09 2022-07-01 电子科技大学 MRI (magnetic resonance imaging) data feature extraction and classification identification method based on artificial neural network
CN114926396A (en) * 2022-04-13 2022-08-19 四川大学华西医院 Mental disorder magnetic resonance image preliminary screening model construction method
CN115409843A (en) * 2022-11-02 2022-11-29 长春理工大学 Brain nerve image feature extraction method based on scale equalization coupling convolution architecture
CN116342603A (en) * 2023-05-30 2023-06-27 杭州脉流科技有限公司 Method for obtaining arterial input function
CN116342603B (en) * 2023-05-30 2023-08-29 杭州脉流科技有限公司 Method for obtaining arterial input function

Also Published As

Publication number Publication date
CN110992351B (en) 2022-08-16

Similar Documents

Publication Publication Date Title
CN110992351B (en) sMRI image classification method and device based on multi-input convolution neural network
CN110110745A (en) Based on the semi-supervised x-ray image automatic marking for generating confrontation network
CN112070781B (en) Processing method and device of craniocerebral tomography image, storage medium and electronic equipment
CN107506761A (en) Brain image dividing method and system based on notable inquiry learning convolutional neural networks
CN110705555A (en) Abdomen multi-organ nuclear magnetic resonance image segmentation method, system and medium based on FCN
CN108537751B (en) Thyroid ultrasound image automatic segmentation method based on radial basis function neural network
CN112686898B (en) Automatic radiotherapy target area segmentation method based on self-supervision learning
WO2022127500A1 (en) Multiple neural networks-based mri image segmentation method and apparatus, and device
CN112052772A (en) Face shielding detection algorithm
CN111667027B (en) Multi-modal image segmentation model training method, image processing method and device
CN113781461A (en) Intelligent patient monitoring and sequencing method
CN117274599A (en) Brain magnetic resonance segmentation method and system based on combined double-task self-encoder
CN112836755B (en) Sample image generation method and system based on deep learning
CN111210398A (en) White blood cell recognition system based on multi-scale pooling
CN106709921B (en) Color image segmentation method based on space Dirichlet mixed model
CN115995040A (en) SAR image small sample target recognition method based on multi-scale network
CN108648187B (en) Depth feature bag based classification method
CN111428803A (en) Wasserstein distance-based depth domain adaptive image classification method
CN112927203A (en) Glioma patient postoperative life prediction method based on multi-sequence MRI global information
CN111428734A (en) Image feature extraction method and device based on residual countermeasure inference learning and computer readable storage medium
CN112581513A (en) Cone beam computed tomography image feature extraction and corresponding method
CN114693964B (en) MRI data feature extraction and classification identification method based on artificial neural network
Jia et al. Three-dimensional segmentation of hippocampus in brain MRI images based on 3CN-net
CN113361543B (en) CT image feature extraction method, device, electronic equipment and storage medium
CN109784356A (en) Matrix variables based on Fisher discriminant analysis are limited Boltzmann machine image classification method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant