WO2019200747A1 - 分割股骨近端的方法、装置、计算机设备和存储介质 - Google Patents

分割股骨近端的方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2019200747A1
WO2019200747A1 PCT/CN2018/095496 CN2018095496W WO2019200747A1 WO 2019200747 A1 WO2019200747 A1 WO 2019200747A1 CN 2018095496 W CN2018095496 W CN 2018095496W WO 2019200747 A1 WO2019200747 A1 WO 2019200747A1
Authority
WO
WIPO (PCT)
Prior art keywords
net
femur
segmentation model
segmentation
proximal end
Prior art date
Application number
PCT/CN2018/095496
Other languages
English (en)
French (fr)
Inventor
王健宗
吴天博
刘新卉
刘莉红
马进
肖京
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2019200747A1 publication Critical patent/WO2019200747A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
    • 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
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V2201/00—Indexing scheme relating to image or video recognition or understanding
    • G06V2201/03—Recognition of patterns in medical or anatomical images
    • G06V2201/033—Recognition of patterns in medical or anatomical images of skeletal patterns

Definitions

  • the present application relates to the field of computer technology, and more particularly to a method, apparatus, computer device and storage medium for segmenting the proximal end of the femur.
  • Deep neural network technology has developed by leaps and bounds. Deep neural network technology has achieved great results in image, text or audio processing, but it has not been widely used in the medical field.
  • Femoral acetabular impingement is one of the causes of pain in the buttocks of adults, and it is considered to be a key factor causing cartilage damage and a precursor to osteoarthritis.
  • MRI diagnosis is now used as a standard tool for the diagnosis of femoral acetabular impingement because it does not pose a radiological hazard.
  • doctors generally perform manual diagnosis based on medical experience by analyzing 2DMRI slice images, and the diagnosis efficiency is low and the probability of diagnosis error is large. Therefore, how to more accurately and effectively separate the proximal end of the femur has become an urgent problem to be solved.
  • the main purpose of the present application is to provide a method for segmenting the proximal femur, which aims to solve the technical problem of accurately separating the proximal femur in a 3D MRI image.
  • the present application proposes a method of segmenting the proximal end of the femur, comprising:
  • the proximal end of the femur in the 3D MRI image is segmented according to the segmentation boundary.
  • the application also provides a device for segmenting the proximal end of the femur, comprising:
  • a first input module for inputting a 3D MRI image of the femur into a segmentation model obtained by pre-training through the 3D U-net;
  • An identification module configured to identify, by the segmentation model, a segmentation boundary of a proximal femur in the 3D MRI image
  • a segmentation module configured to segment the proximal end of the femur in the 3D MRI image according to the segmentation boundary.
  • the application also provides a computer device comprising a memory and a processor, the memory storing computer readable instructions, the processor implementing the steps of the method when the computer readable instructions are executed.
  • the present application also provides a computer non-transitory readable storage medium having stored thereon computer readable instructions that, when executed by a processor, implement the steps of the methods described above.
  • the present invention has the beneficial technical effects: the present application automatically separates the proximal femur from the 3D MRI image through the segmentation model, and reduces the diagnostic interference information by separating the proximal end of the femur from the original image, thereby greatly improving the diagnosis efficiency of the doctor;
  • This application proposes a 3D MRI proximal femoral segmentation technique based on 3DU-net. Through a 3DU-net network with deep supervised learning effects, a small segmentation sample training is used to obtain an accurate segmentation model to achieve accurate 3D MRI proximal femur. Segmentation makes up for the lack of 3D MRI image data of existing annotations, and it is difficult to obtain the technical problem of accurate segmentation.
  • By assembling the diagnostic data of lesions to form a prior database it can help improve the diagnostic accuracy of doctors to diagnose the disease, and make up for the diagnosis caused by lack of experience. Defects with low accuracy have practical application value.
  • FIG. 1 is a schematic flow chart of a method for segmenting a proximal femur according to an embodiment of the present application
  • FIG. 2 is a schematic structural view of a device for dividing a proximal femur according to an embodiment of the present application
  • FIG. 3 is a schematic diagram showing an optimized structure of a device for segmenting a proximal femur according to an embodiment of the present application
  • FIG. 4 is a schematic structural diagram of a second input module according to an embodiment of the present application.
  • FIG. 5 is a schematic structural diagram of a classification unit according to an embodiment of the present application.
  • FIG. 6 is a schematic structural view of a device for dividing a proximal end of a femur according to still another embodiment of the present application;
  • FIG. 7 is a schematic structural view of a device for splitting a proximal end of a femur according to still another embodiment of the present application.
  • FIG. 8 is a schematic structural view of a device for splitting a proximal end of a femur according to still another embodiment of the present application.
  • FIG. 9 is a schematic structural view of a device for splitting a proximal end of a femur according to still another embodiment of the present application.
  • FIG. 10 is a schematic diagram showing the internal structure of a computer device according to an embodiment of the present application.
  • a method for segmenting a proximal femur includes:
  • S1 The 3D MRI image of the femur is input into a segmentation model obtained by pre-training through 3D U-net.
  • the 3D MRI (Magnetic Resonance Imaging) of the femur of the present embodiment is a "digital image" spatially encoded by a nuclear magnetic resonance signal.
  • the magnetic resonance signal is directly from the object itself, and the magnetic resonance imaging can obtain the object in any direction.
  • the tomographic image and the three-dimensional volume image can reconstruct an object image, such as an anatomical and pathological cross-sectional image showing different human femur structures in different gray scales.
  • the 3DU-net of this embodiment is a split network model, and the entire network is in a "U" shape, which is also a source of the network name U-net.
  • the downlink area of the "U" character belongs to the editor, and the uplink area belongs to the decoder.
  • the 3D U-net network is a deep supervisory learning network.
  • the so-called deep supervision learning refers to the process of adjusting the parameters of the classifier to achieve the required performance by using a set of samples with known correct answers.
  • the network is learned by using tagged data.
  • the initialized network continually modifies the parameters in the network according to the difference between the predicted value and the tag, so that the predicted value of the network is closer to the tag to achieve the purpose of learning.
  • the precise segmentation model can be obtained by using a small number of labeled sample training to achieve accurate segmentation of the proximal 3D MRI femur of the femur, which makes up for the lack of existing 3D MRI image data, difficult to obtain precise segmentation technical problems, and training time. Short, reducing model costs.
  • S2 Identify, by the segmentation model, a segmentation boundary of a proximal femur in the 3D MRI image.
  • Each layer of the 3D U-net editor convolves and pools the input 3D MRI data for feature extraction.
  • Each layer of the decoder uses deconvolution to decode the extracted features to obtain a mapping layer and output the mapping layer.
  • the size is the same as the size of the input image.
  • the mapping layer indicates the meaning of each part of the original 3D MRI image, that is, the segmentation result.
  • the segmentation model of the 3D U-net training identifies which part of the original 3D MRI image is the femur. end.
  • the 3D U-net training is used to obtain the segmentation model to segment the proximal femur of the 3D MRI, so that the accuracy of the segmentation data is higher and has higher practical value.
  • the proximal end of the femur is automatically separated from the 3D MRI image by the segmentation model, and the proximal end of the femur is separated from the original image to reduce the diagnostic interference information, thereby greatly improving the diagnostic efficiency of the doctor.
  • step S1 above the method includes:
  • the initialization parameters of this step include random initialization parameters or optimized initialization parameters.
  • the weight matrix is used to judge whether the training process converges.
  • the training convergence is considered, and the training converges.
  • the parameters of the 3D U-net encoder and decoder are optimized to determine the network weight value matrix in the optimized state, so as to output a segmentation model with high accuracy.
  • the training set data consisting of the 3D MRI image with detailed annotation information and the segmentation boundary of the proximal femur corresponding to each 3D MRI image is input to the 3D U-net for training to obtain the proximal edge of the femur.
  • the feature is extracted in order to train a segmentation model that can identify the segmentation boundary of the proximal femur from the 3D MRI image.
  • S13 Determine whether the accuracy of the segmentation model reaches a preset threshold by inputting preset test set data into the segmentation model.
  • the test set data and the training set data of this embodiment are both from the same sample set and undergo the same data processing, but the test set data and the training set have no intersection.
  • the data set is divided into a training set and a test set, the model is trained on the training set, and the test set is used to test the effect of the model.
  • the accuracy of the segmentation model in this embodiment is measured by the Dice coefficient, and the Dice coefficient is a set similarity measure function.
  • the Dice coefficient calculation formula of this embodiment is: 2*
  • step S11 the method includes:
  • S10 migrating layer parameters of the C3D convolutional layer network having the same number of feature maps as the 3D U-net compiler to the 3D U-net compiler as initialization parameters of the 3D U-net .
  • this embodiment uses C3D (3D convolution) through migration learning.
  • the network parameters are used as the parameters of the 3D U-net compiler.
  • the C3D network structure has 8 convolution layers and 4 pooling layers.
  • the size of the convolution kernel is 3*3*3, and the size of the pooled core is 2*2*2.
  • the number of feature maps of the first six convolutional layers (64, 128, 256, 256, 512, 512) is the same as the number of feature maps of the corresponding editor layers in the 3DU-net network (64, 128, 256, 256, 512, 512), so the number of parameters of the convolution layer is the same, and the training has been completed.
  • the C3D convolutional layer parameter initializes the convolutional layer of the 3D U-net network, that is, the layer parameters of the encoder, so this embodiment obtains the parameters of the first six convolutional layers of the C3D model for initializing the 3D U- The various layers of the net network compiler.
  • the process of initializing the convolution layer of the 3D U-net network by using the C3D convolution layer parameters that have been trained is a migration learning process, and the data used in the training process of the C3D network and the 3D U-net network in this embodiment is different.
  • the training time can be reduced while the training effect is ensured by the migration learning, and the training effect is better optimized by migrating the learning training network model than the initialization of the entire network by the random initialization method.
  • step S12 includes:
  • S120 Input the training set data into the 3D U-net, reduce the spatial dimension of the pooling layer by using the 3D U-net encoder, and locate each pixel level of the training set data.
  • the encoder of this embodiment is a full convolutional neural network structure in which the feature size gradually shrinks and the number of channels gradually increases, and can receive an input image of any size, and the convolution layer and the pooling layer work alternately to capture upper and lower layers. Information, and gradually downsample the feature map to restore the image size.
  • the encoder gradually shrinks to reduce the spatial dimension of the pooling layer, expands the receptive field through the pooling layer, and locates each pixel level of the training set data.
  • the convolutional layer of this embodiment preferably uses a 3*3*3 convolution core, and the largest pooled layer uses a 2*2*2 pooled core.
  • the size of the convolution kernel and the pooled core are adjustable, and may be different in different networks, and different convolution layers and pooling layers in the same network may also be different.
  • S121 Stepwise repairing details and spatial dimensions of the training set data by a decoder connected to the encoder, classifying each of the pixel levels to form the segmentation model identifying a segmentation boundary of the proximal end of the femur .
  • the decoder of the embodiment is a symmetrical structure with the encoder, the feature image size is gradually expanded, the number of channels is gradually reduced, the detail and spatial dimension of the object are gradually restored, and the high resolution is gradually restored by upsampling the feature image. Rate picture details.
  • the pooling layer of the encoder of this embodiment expands the receptive field but causes the loss of position information, the pixel level classification requires that the position information be retained.
  • a large number of feature channels can transmit upper and lower layer information to higher resolution layers, resulting in a prediction for each pixel while upsampling the feature image of the last convolution layer. Revert to the same size of the input image, while retaining the spatial information in the original input image, and finally classify the pixels on the upsampled feature map to identify the segmentation boundary of the proximal femur.
  • step S121 includes:
  • the decoder gradually repairs the details of the training data and the output information after the spatial dimension, and inputs the information to the branch network connected to each decoder for training.
  • the image input into the training set data in the convolution training process by the 3D U-net passes through the convolution kernel and the pooled kernel, and the image size becomes smaller, and needs to be upsampled by deconvolution to obtain the same size as the original image. Mapping layer, but the larger the upsampling, the more details are lost. Therefore, the results of different upsampling of different layers need to be combined to obtain the final result.
  • This embodiment passes the lower layer output layer L, the middle layer output layer M and the upper layer.
  • the three outputs of output layer U represent the results at different upsampling multiples, respectively.
  • the input image size is 64*64*64
  • the size before the lower output layer L is 16*16*16, which requires 4 times upsampling to obtain the original image size
  • the size before the middle output layer M is 32. *32*32, you need to perform 2 times upsampling to get the original image size.
  • the outputs of the decoder layers of this embodiment are respectively input as a branch network, and the branch network finally obtains three outputs of the lower layer output layer L, the middle layer output layer M and the upper layer output layer U, respectively.
  • the weight of the entire 3D U-net network is further optimized, so that the abstract expression is more accurate and closer to the global optimum, and a better segmentation effect is obtained.
  • S1211 Obtain the segmentation model of the optimized weight parameter by analyzing an output result of the branch network.
  • the loss function of the 3D U-net of this embodiment is defined as the sum of the three output cross entropy loss functions of the branch network, and the specific form is as follows:
  • W is the weight of the main network of the 3D U-net
  • ⁇ l , ⁇ m , ⁇ u are the lower output layer L of the branch network, the weights of the middle layer output layer M and the upper layer output layer U, respectively, and represent the training data. It is a regular term that further avoids overfitting.
  • the weight of the primary network and the weight of the branch network are combined to the loss function. When the loss function converges, the optimized weight parameter of the segmentation model is obtained, that is, the weight of the corresponding primary network and the weight of the branch network when the loss function converges.
  • the loss function of the 3D U-net in this embodiment is defined as the sum of the three output cross entropy loss functions of the branch network, so that the weight of the main network and the weight of the branch network affect the size of the loss function, which affects the prediction of the segmentation model.
  • c ⁇ l,m,u ⁇ that is, c is an index, which refers to one of ⁇ l,mu ⁇ , written under the summation symbol ,
  • c is taken as ⁇ l
  • the values of m, u ⁇ are summed.
  • the method before step S12, includes:
  • a small number of labeled samples are used for segmentation model training, and the original data is enhanced to prevent over-fitting.
  • data enhancement is performed by rotating the original data by a specified angle. For example, the original data is rotated (90 degrees, 180 degrees, 270 degrees), because 90 degrees, 180 degrees, and 270 degrees are evenly distributed at 0-360 degrees, and by rotating the three angles, the enhancement data is uniformly changed.
  • this step by randomly cutting the sub-blocks from each picture of the enhanced data set as the training set data, in the embodiment, a plurality of 64*64*64 sub-blocks are randomly cut out in each picture, and 10 pieces are selected as training. Set data.
  • S124 All the sub-blocks are randomly divided into two groups, one group is used as training set data, and the other group is used as test set data.
  • the original image has a total of 10, and after three angles of data rotation enhancement, it becomes 30 images, and each image is cut out as 10 training set data, and the training set data has a total of 300.
  • Other embodiments of the present application normalize the above training set data to a mean of 0 and a variance of one.
  • data normalization is also referred to as normalization, and data of different dimensions and orders of magnitude are converted into data that can be mathematically calculated and comparable to each other, and normalization makes the distribution of data in various dimensions relatively close. It is possible to appropriately expand the data difference and accelerate the convergence of the model training process.
  • the calculation process of the mean value of the present embodiment with a variance of 0 is as follows: For example, a set of data is:
  • step S3 the method includes:
  • S30 Obtain position information of the lesion region at the proximal end of the segmented femur and distribution information of the lesion region.
  • the position information of the lesion area and the distribution information of the lesion area of the present embodiment are obtained by identifying the edge information of the lesion area by the pre-trained lesion area segmentation model.
  • the positional information of the lesion area can be obtained by the coordinate information located at the proximal end of the femur.
  • the distribution information of the lesion area includes the number of lesion areas, the relative relationship of the position information of each lesion area, and the like, for example, the lesion position is the acetabulum, and the lesion area is The distribution information is a lesion area, specifically the acetabular over-coverage and the acetabular fossa deepening; for example, the lesion location is the femoral head-neck intersection, and the distribution information of the lesion area is a lesion area, specifically the presence of protrusions in the femoral head-neck intersection.
  • S31 Outputting the disease information of the femoral acetabular impact by analyzing the position information of the lesion area and the distribution information of the lesion area.
  • the disease information of the segmented femoral acetabular impact can be obtained through a pre-established analysis model, and the network structure of the analysis model includes, but is not limited to, a convolutional neural network, a full convolutional neural network or U-net.
  • step S31 the method includes:
  • S32 Combine the disease information with the identification information of the case and the diagnosis feedback information into a priori database for diagnosing the impact of the femoral acetabulum.
  • the identity information of this step includes, but is not limited to, gender, age, exercise preference, etc.
  • the diagnosis feedback information includes, but is not limited to, acetabular abnormality, femoral head abnormality, and the like
  • the condition information includes a feature matrix corresponding to each disease information.
  • step S32 the method includes:
  • the similarity of the characteristic matrix of the new output disorder information outputted by the analysis model and the feature matrix of each disorder information in the prior database is compared to realize the diagnosis of the similar disorder.
  • the prior database is combined with the image segmentation model to be extended to the actual automated disease diagnosis, thereby improving the diagnosis efficiency and accuracy of the doctor.
  • the proximal end of the femur is automatically separated from the 3D MRI image by the segmentation model, and the proximal end of the femur is separated from the original image, thereby reducing diagnostic interference information and greatly improving the diagnostic efficiency of the doctor;
  • 3DU-net's 3D MRI proximal femoral segmentation technology uses a small number of labeled sample training to obtain a precise segmentation model, achieving accurate segmentation of the proximal end of the 3D MRI femur, making up for the present
  • the marked 3D MRI image data is scarce, and it is difficult to obtain the technical problem of accurate segmentation.
  • an apparatus for segmenting a proximal femur includes:
  • the first input module 1 is configured to input a 3D MRI image of the femur into a segmentation model obtained by pre-training through the 3D U-net.
  • the 3D MRI (Magnetic Resonance Imaging) of the present embodiment is a "digital image" spatially encoded by a nuclear magnetic resonance signal.
  • the magnetic resonance signal is directly from the object itself, and the magnetic resonance imaging can obtain a tomographic image of the object in any direction.
  • the three-dimensional image can reconstruct the image of the object, such as the anatomical and pathological cross-sectional images of different human femur structures in different gray levels.
  • the 3D U-net of this embodiment is a split network model, and the entire network is in a "U" shape, which is also a source of the network name U-net.
  • the downlink area of the "U" character belongs to the editor, and the uplink area belongs to the decoder.
  • the 3DU-net network is a deep supervisory learning network.
  • the so-called deep supervisory learning refers to the process of adjusting the parameters of a classifier to achieve the required performance by using a set of samples with known correct answers.
  • the network is learned by using tagged data.
  • the initialized network continually modifies the parameters in the network according to the difference between the predicted value and the tag, so that the predicted value of the network is closer to the tag to achieve the purpose of learning.
  • the precise segmentation model can be obtained by using a small number of labeled sample training to achieve accurate segmentation of the proximal 3D MRI femur of the femur, which makes up for the lack of existing 3D MRI image data, difficult to obtain precise segmentation technical problems, and training time. Short, reducing model costs.
  • the identification module 2 is configured to identify, by the segmentation model, a segmentation boundary of a proximal femur in the 3D MRI image.
  • Each layer of the 3D U-net editor convolves and pools the input 3D MRI data for feature extraction.
  • Each layer of the decoder uses deconvolution to decode the extracted features to obtain a mapping layer and output the mapping layer.
  • the size is the same as the size of the input image.
  • the mapping layer indicates the meaning of each part of the original 3D MRI image, that is, the segmentation result.
  • the segmentation model of the 3D U-net training identifies which part of the original 3D MRI image is the femur. end.
  • the segmentation module 3 is configured to segment the proximal end of the femur in the 3D MRI image according to the segmentation boundary.
  • the proximal end of the femur is segmented by 3D U-net training, which makes the segmentation data more accurate and has higher practical value.
  • the proximal end of the femur is automatically separated from the 3D MRI image by the segmentation model, and the proximal end of the femur is separated from the original image to reduce the diagnostic interference information, thereby greatly improving the diagnostic efficiency of the doctor.
  • the device for segmenting the proximal end of the femur of the embodiment includes:
  • the first obtaining module 11 is configured to obtain, by using supervised learning, the optimization parameters corresponding to the encoder and the decoder of the 3D U-net respectively under the initialization parameters of the 3D U-net.
  • the initialization parameters of this embodiment include random initialization parameters or optimized initialization parameters.
  • the weight matrix is used to determine whether the training process converges.
  • the training convergence is considered, and the training convergence is considered.
  • the parameters of the 3D U-net encoder and decoder are optimized to determine the network weight value matrix in the optimized state, so as to output a segmentation model with high accuracy.
  • the second input module 12 is configured to input the preset training set data into the 3D U-net to train the segmentation model under the optimization parameter.
  • the training set data consisting of the 3D MRI image with detailed annotation information and the segmentation boundary of the proximal femur corresponding to each 3D MRI image is input to the 3D U-net for training to obtain the proximal edge of the femur.
  • the feature is extracted in order to train a segmentation model that can identify the segmentation boundary of the proximal femur from the 3D MRI image.
  • the determining module 13 is configured to determine whether the accuracy of the segmentation model reaches a preset threshold by inputting preset test set data into the segmentation model under the optimal parameter.
  • the test set data and the training set data of this embodiment are both from the same sample set and undergo the same data processing, but the test set data and the training set have no intersection.
  • the data set is divided into a training set and a test set, the model is trained on the training set, and the test set is used to test the effect of the model.
  • the accuracy of the segmentation model in this embodiment is measured by the Dice coefficient, and the Dice coefficient is a set similarity measure function.
  • the Dice coefficient calculation formula of this embodiment is: 2*
  • the determining module 14 is configured to determine that the segmentation model meets an application requirement if an accuracy rate of the segmentation model reaches a preset threshold.
  • the apparatus for segmenting the proximal end of the femur of the embodiment includes:
  • the initialization module 10 is configured to migrate, in the C3D convolutional layer network, each layer parameter having the same number of feature maps as the 3D U-net compiler into the 3D U-net compiler as the 3D U- Net initialization parameters.
  • this embodiment uses C3D (3D convolution) through migration learning.
  • the network parameters are used as the parameters of the 3D U-net compiler.
  • the C3D network structure has 8 convolution layers and 4 pooling layers.
  • the size of the convolution kernel is 3*3*3, and the size of the pooled core is 2*2*2.
  • the number of feature maps of the first six convolutional layers (64, 128, 256, 256, 512, 512) is the same as the number of feature maps of the corresponding editor layers in the 3DU-net network (64, 128, 256, 256, 512, 512), so the number of parameters of the convolution layer is the same, and the training has been completed.
  • the C3D convolutional layer parameter initializes the convolutional layer of the 3D U-net network, that is, the layer parameters of the encoder, so this embodiment obtains the parameters of the first six convolutional layers of the C3D model for initializing the 3D U- The various layers of the net network compiler.
  • the process of initializing the convolution layer of the 3D U-net network by using the C3D convolution layer parameters that have been trained is a migration learning process, and the data used in the training process of the C3D network and the 3D U-net network in this embodiment is different.
  • the training time can be reduced while the training effect is ensured by the migration learning, and the training effect is better optimized by migrating the learning training network model than the initialization of the entire network by the random initialization method.
  • the second input module 12 of this embodiment includes:
  • the locating unit 120 is configured to input the training set data into the 3D U-net, reduce the spatial dimension of the pooling layer by the 3D U-net encoder, and locate each pixel level of the training set data.
  • the encoder of this embodiment is a full convolutional neural network structure in which the feature size gradually shrinks and the number of channels gradually increases, and can receive an input image of any size, and the convolution layer and the pooling layer work alternately to capture upper and lower layers. Information, and gradually downsample the feature map to restore the image size.
  • the encoder gradually shrinks to reduce the spatial dimension of the pooling layer, expands the receptive field through the pooling layer, and locates each pixel level of the training set data.
  • the convolutional layer of this embodiment preferably uses a 3*3*3 convolution core, and the largest pooled layer uses a 2*2*2 pooled core.
  • the size of the convolution kernel and the pooled core are adjustable, and may be different in different networks, and different convolution layers and pooling layers in the same network may also be different.
  • a classifying unit 121 configured to gradually repair the details and spatial dimensions of the training set data by a decoder connected to the encoder, and classify each of the pixel levels to form a segmentation boundary identifying the proximal end of the femur Split the model.
  • the decoder of the embodiment is a symmetrical structure with the encoder, the feature image size is gradually expanded, the number of channels is gradually reduced, the detail and spatial dimension of the object are gradually restored, and the high resolution is gradually restored by upsampling the feature image. Rate picture details.
  • the pooling layer of the encoder of this embodiment expands the receptive field but causes the loss of position information, the pixel level classification requires that the position information be retained.
  • a large number of feature channels can transmit upper and lower layer information to higher resolution layers, resulting in a prediction for each pixel while upsampling the feature image of the last convolution layer. Revert to the same size of the input image, while retaining the spatial information in the original input image, and finally classify the pixels on the upsampled feature map to identify the segmentation boundary of the proximal femur.
  • the classification unit 121 of this embodiment includes:
  • the training sub-unit 1210 is configured to gradually repair the output information of the training data and the output information after the spatial dimension, and input the information to the branch network connected to each decoder for training.
  • the image input into the training set data in the convolution training process by the 3D U-net passes through the convolution kernel and the pooled kernel, and the image size becomes smaller, and needs to be upsampled by deconvolution to obtain the same size as the original image. Mapping layer, but the larger the upsampling, the more details are lost. Therefore, the results of different upsampling of different layers need to be combined to obtain the final result.
  • This embodiment passes the lower layer output layer L, the middle layer output layer M and the upper layer.
  • the three outputs of output layer U represent the results at different upsampling multiples, respectively.
  • the input image size is 64*64*64
  • the size before the lower output layer L is 16*16*16, which requires 4 times upsampling to obtain the original image size
  • the size before the middle output layer M is 32. *32*32, you need to perform 2 times upsampling to get the original image size.
  • the outputs of the decoder layers of this embodiment are respectively input as a branch network, and the branch network finally obtains three outputs of the lower layer output layer L, the middle layer output layer M and the upper layer output layer U, respectively.
  • the weight of the entire 3D U-net network is further optimized, so that the abstract expression is more accurate and closer to the global optimum, and a better segmentation effect is obtained.
  • the obtaining subunit 1211 is configured to obtain the segmentation model of the optimized weight parameter by analyzing an output result of the branch network.
  • the loss function of the 3D U-net of this embodiment is defined as the sum of the three output cross entropy loss functions of the branch network, and the specific form is as follows:
  • W is the weight of the main network of the 3D U-net
  • ⁇ l , ⁇ m , ⁇ u are the lower output layer L of the branch network, the weights of the middle layer output layer M and the upper layer output layer U, respectively, and represent the training data. It is a regular term that further avoids overfitting.
  • the weight of the primary network and the weight of the branch network are combined to the loss function. When the loss function converges, the optimized weight parameter of the segmentation model is obtained, that is, the weight of the corresponding primary network and the weight of the branch network when the loss function converges.
  • the loss function of the 3D U-net in this embodiment is defined as the sum of the three output cross entropy loss functions of the branch network, so that the weight of the main network and the weight of the branch network affect the size of the loss function, which affects the prediction of the segmentation model.
  • c ⁇ l,m,u ⁇ that is, c is an index, which refers to one of ⁇ l,mu ⁇ , written under the summation symbol ,
  • c is taken as ⁇ l
  • the values of m, u ⁇ are summed.
  • ⁇ c (x; W, ⁇ l , ⁇ m , ⁇ u ) ⁇ l ⁇ l (x; W, ⁇ l ) + ⁇ ( ⁇ (W) + ⁇ ( ⁇ l ))
  • a device for segmenting a proximal femur includes:
  • the component module 122 is configured to compose the original 3D MRI image data and the enhanced data after rotating the original 3D MRI image data by a specified angle into a data set.
  • a small number of labeled samples are used for segmentation model training, and the original data is enhanced to prevent over-fitting.
  • data enhancement is performed by rotating the original data by a specified angle. For example, the original data is rotated (90 degrees, 180 degrees, 270 degrees), because 90 degrees, 180 degrees, and 270 degrees are evenly distributed at 0-360 degrees, and by rotating the three angles, the enhancement data is uniformly changed.
  • the cutting module 123 is configured to cut each 3D MRI image data in the data set into a specified number and a sub-block of a specified size.
  • the distinguishing module 124 is configured to randomly divide all the sub-blocks into two groups, one group as the training set data and the other group as the test set data.
  • the original image has a total of 10, and after three angles of data rotation enhancement, it becomes 30 images, and each image is cut out as 10 training set data, and the training set data has a total of 300.
  • Other embodiments of the present application normalize the above training set data to a mean of 0 and a variance of one.
  • data normalization is also referred to as normalization, and data of different dimensions and orders of magnitude are converted into data that can be mathematically calculated and comparable to each other, and normalization makes the distribution of data in various dimensions relatively close. It is possible to appropriately expand the data difference and accelerate the convergence of the model training process.
  • the calculation process of the mean value of the present embodiment with a variance of 0 is as follows: For example, a set of data is:
  • a device for segmenting a proximal femur includes:
  • the second obtaining module 30 is configured to acquire position information of the lesion region at the proximal end of the segment and the distribution information of the lesion region.
  • the position information of the lesion area and the distribution information of the lesion area of the present embodiment are obtained by identifying the edge information of the lesion area by the pre-trained lesion area segmentation model.
  • the positional information of the lesion area can be obtained by the coordinate information located at the proximal end of the femur.
  • the distribution information of the lesion area includes the number of lesion areas, the relative relationship of the position information of each lesion area, and the like, for example, the lesion position is the acetabulum, and the lesion area is The distribution information is a lesion area, specifically the acetabular over-coverage and the acetabular fossa deepening; for example, the lesion location is the femoral head-neck intersection, and the distribution information of the lesion area is a lesion area, specifically the presence of protrusions in the femoral head-neck intersection.
  • the first output module 31 is configured to output the disease information of the femoral acetabular impact by analyzing the position information of the lesion area and the distribution information of the lesion area.
  • the disease information of the segmented femoral acetabular impact can be obtained through a pre-established analysis model, and the network structure of the analysis model includes, but is not limited to, a convolutional neural network, a full convolutional neural network, or U-net.
  • a device for segmenting a proximal femur includes:
  • the aggregating module 32 is configured to collect the disease information and the identity information and the diagnosis feedback information of the case into a priori database for diagnosing the impact of the femoral acetabulum.
  • the identity information of this embodiment includes, but is not limited to, gender, age, sports preference, etc.
  • the diagnosis feedback information includes, but is not limited to, an acetabular abnormality, an abnormality of the femoral head, and the like
  • the symptom information includes a feature matrix corresponding to each disease information.
  • a device for segmenting a proximal femur includes:
  • the searching module 33 is configured to search, in the a priori database, historical condition information that is most similar to the new illness information.
  • the diagnosis of a similar condition is achieved by comparing the similarity between the feature matrix of the new output symptom information outputted by the analysis model and the feature matrix of each disease information in the a priori database.
  • the second output module 34 is configured to output the diagnosis feedback information corresponding to the historical condition information.
  • the prior database is combined with the image segmentation model to be extended to the actual automated disease diagnosis, thereby improving the diagnosis efficiency and accuracy of the doctor.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 9.
  • the computer device includes a processor, memory, network interface, and database connected by a system bus. Among them, the computer designed processor is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium, an internal memory.
  • the non-volatile storage medium stores an operating system, computer readable instructions, and a database.
  • the memory provides an environment for the operation of operating systems and computer readable instructions in a non-volatile storage medium.
  • the database of the computer device is used to store data such as splitting the proximal end of the femur.
  • the network interface of the computer device is used to communicate with an external terminal via a network connection.
  • the computer readable instructions when executed, perform the flow of an embodiment of the methods described above. It will be understood by those skilled in the art that the structure shown in FIG. 9 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation of the computer device to which the present application is applied.
  • An embodiment of the present application also provides a computer non-volatile readable storage medium having stored thereon computer readable instructions that, when executed, perform the processes of the embodiments of the methods described above.
  • the above description is only the preferred embodiment of the present application, and is not intended to limit the scope of the patent application, and the equivalent structure or equivalent process transformations made by the specification and the drawings of the present application, or directly or indirectly applied to other related The technical field is equally included in the scope of patent protection of the present application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

一种分割股骨近端的方法、装置、计算机设备和存储介质，方法包括：将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中（S1）；通过分割模型识别3D MRI图像中股骨近端的分割边界（S2）；根据分割边界对3D MRI图像中的股骨近端进行分割（S3）。该方法从3D MRI图像中通过分割模型分离出来股骨近端，减少诊断干扰信息，提高诊断效率。

Description

分割股骨近端的方法、装置、计算机设备和存储介质
本申请要求于2018年4月20日提交中国专利局、申请号为2018103621986，发明名称为“分割股骨近端的方法、装置、计算机设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及到计算机技术领域，特别是涉及到分割股骨近端的方法、装置、计算机设备和存储介质。
背景技术
近年来，由于计算机硬件水平的提高，深度神经网络技术得到了突飞猛进的发展，深度神经网络技术在图像、文字或音频处理等方面取得了巨大的成果，但在医学领域的还未得到广泛应用。股骨髋臼撞击症是引起成年人臀部疼痛的原因之一，更是被认为是引起软骨损伤的关键因素以及骨关节炎的先兆。由于不会造成放射性危害，MRI诊断现在被作为进行股骨髋臼撞击症诊断的标准工具。但是，现有技术中医生一般会通过分析2DMRI切片图像凭医疗经验进行人工诊断，诊断效率低且诊断误差几率大。因此，如何更精确有效地分离股骨近端成为亟待解决的问题。
技术问题
本申请的主要目的为提供分割股骨近端的方法，旨在解决3D MRI图像中精准分离股骨近端的技术问题。
技术解决方案
本申请提出一种分割股骨近端的方法，包括：
将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中；
通过所述分割模型识别所述3D MRI图像中股骨近端的分割边界；
根据所述分割边界对所述3D MRI图像中的股骨近端进行分割。
本申请还提供了一种分割股骨近端的装置，包括：
第一输入模块，用于将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中；
识别模块，用于通过所述分割模型识别所述3D MRI图像中股骨近端的分割边界；
分割模块，用于根据所述分割边界对所述3D MRI图像中的股骨近端进行分割。
本申请还提供了一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现上述方法的步骤。
本申请还提供了一种计算机非易失性可读存储介质，其上存储有计算机可读指令，所述计算机可读指令被处理器执行时实现上述的方法的步骤。
有益效果
本申请有益技术效果：本申请通过从3D MRI图像中通过分割模型自动分离出来股骨近端，通过将股 骨近端从原图中分离出来，减少诊断干扰信息，极大的提高医生的诊断效率；本申请提出了基于3DU-net的3D MRI股骨近端分割技术，通过具有深度监督学习效果的3DU-net网络，利用少量标注样本训练获得了精准的分割模型，实现对3D MRI股骨近端的精准分割，弥补了现有标注的3D MRI图像数据困乏，难以获得精准分割的技术难题；通过汇集病变诊断数据形成先验数据库，辅助提高医生诊断病症的诊断准确度，弥补医生因经验不足而导致诊断准确度不高的缺陷，具有实际应用价值。
附图说明
图1本申请一实施例的分割股骨近端的方法流程示意图；
图2本申请一实施例的分割股骨近端的装置结构示意图；
图3本申请一实施例的分割股骨近端的装置优化结构示意图；
图4本申请一实施例的第二输入模块的结构示意图；
图5本申请一实施例的分类单元的结构示意图；
图6本申请再一实施例的分割股骨近端的装置结构示意图；
图7本申请又一实施例的分割股骨近端的装置结构示意图；
图8本申请又一实施例的分割股骨近端的装置结构示意图；
图9本申请又一实施例的分割股骨近端的装置结构示意图；
图10本申请一实施例的计算机设备内部结构示意图。
本发明的最佳实施方式
参照图1，本申请一实施例的分割股骨近端的方法，包括：
S1：将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中。
本实施例的股骨的3D MRI(Magnetic Resonance Imaging，图像磁共振成像)为通过核磁共振信号进行空间编码的“数字图像”，磁共振信号直接来自于物体本身，磁共振成像可以得到物体任何方向的断层图像、三维体图像，可以重建出物体图像，比如以不同灰度显示不同人体股骨结构的解剖和病理的断面图像。本实施例的3DU-net是分割型网络模型，整个网络呈“U”形，也是网络名为U-net的来源。在“U”字的下行区域属于编辑器，上行区域属于解码器。3D U-net网络是深度监督学习网络，所谓深度监督学习是指：利用一组已知正确答案的样本调整分类器的参数，使其达到所要求性能的过程。对于深度监督学习网络就是利用有标签的数据来学习的网络，被初始化的网络根据预测值与标签的差别不断修改网络中的参数，使网络的预测值越来越接近标签，以达到学习的目的，可利用少量标注样本训练获得了精准的分割模型，实现对股骨的3D MRI股骨近端的精准分割，弥补了现有标注的3D MRI图像数据困乏，难以获得精准分割的技术难题，且训练时间短，降低了模型成本。
S2：通过所述分割模型识别所述3D MRI图像中股骨近端的分割边界。
3D U-net的编辑器各层对输入的3D MRI数据进行卷积和池化，以进行特征提取，解码器各层使用 反卷积对提取的特征进行解码得到映射层并输出，映射层的大小与输入图像大小相同，映射层指示出原始3D MRI图的每一部分代表的含义，即分割结果，本实施例通过3D U-net训练的分割模型识别出原始3D MRI图的哪一部分是股骨近端。
S3：根据所述分割边界对所述3D MRI图像中的股骨近端进行分割。
本实施例通过3D U-net训练得到分割模型分割3D MRI的股骨近端，使得分割数据的准确率更高，具有较高的实用价值。本实施例通过从3D MRI图像中通过分割模型自动分离出来股骨近端，通过将股骨近端从原图中分离出来，减少诊断干扰信息，极大的提高医生的诊断效率。
进一步地，上述步骤S1之前，包括：
S11：在3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化参数。
本步骤的初始化参数包括随机的初始化参量或优化的初始化参量。本步骤通过监督学习训练3D U-net时，通过权值矩阵判断训练过程是否收敛，当两次迭代之间的权值变化很小，比如变化值小于1％，则认为训练收敛，训练收敛时3D U-net编码器和解码器的参数达到最大优化，以确定优化状态下的网络权重值矩阵，以便输出准确率高的分割模型。
S12：在所述优化参数下，将预设训练集数据输入3D U-net训练所述分割模型。
本实施例通过由具有详细标注信息的3D MRI图像以及每个3D MRI图像对应的股骨近端的分割边界组成的训练集数据，输入到3D U-net进行训练，以获取到股骨近端边缘的特征的提取方式，以便训练出可以从3D MRI图像中识别股骨近端的分割边界的分割模型。
S13：通过将预设的测试集数据输入所述分割模型，判断所述分割模型的准确率是否达到预设阈值。
本实施例的测试集数据与训练集数据均来自同一样本集且经过同样的数据处理，但测试集数据与训练集无交集。本实施例通过把数据集分为训练集与测试集，在训练集上训练模型，通过测试集来测试模型效果。本实施例中分割模型的准确率通过Dice系数来衡量，Dice系数是集合相似度度量函数，本实施例的Dice系数计算公式为：2*|X∩Y|/(|X|+|Y|)，X是预测区域，Y是图层表面实况，即标注区域(值域范围：0-1)。本实施例的分割模型在测试集上的最终Dice系数达到0.98时，表示测试集与原始标注的训练集重合度非常高。
S14：若达到，则确定所述分割模型满足应用需求。
进一步地，步骤S11之前，包括：
S10：将C3D的卷积层网络中与所述3D U-net编译器具有相同特征图个数的各层参数迁移至所述3D U-net编译器中作为所述3D U-net的初始化参数。
为避免过拟合，训练集数据的集合应该足够大，但训练耗时太长，为了减小训练时间并防止过拟合，本实施例通过迁移学习把C3D(3D convolution，三维卷积网络)的网络参数作为3D U-net编译器的各 层参数。C3D网络结构共有8个卷积层，4次池化层。其中卷积核的大小均为3*3*3，池化核的大小为2*2*2。且前6个卷积层的特征图个数(64,128,256,256,512,512)与3DU-net网络中对应编辑器各层的特征图个数(64,128,256,256,512,512)相同，所以卷积层的参数个数相同，通过已经训练好的C3D的卷积层参数初始化3D U-net网络的卷积层，即编码器的各层参数，所以本实施例通过获取C3D模型的前6个卷积层的参数，用于初始化3D U-net网络的编译器的各层参数。本实施例通过已经训练好的C3D的卷积层参数初始化3D U-net网络的卷积层的过程为迁移学习过程，本实施例C3D网络与3D U-net网络的训练过程中使用的数据不同。本实施例通过迁移学习在保证训练效果的同时可以减少训练时间，通过迁移学习训练的网络模型相比于通过随机初始化的方式初始化整个网络的训练效果更优化。
进一步地，上述步骤S12，包括：
S120：将所述训练集数据输入3D U-net，通过3D U-net的编码器减少池化层的空间维度，定位训练集数据的各像素级。
本实施例的编码器是一个特征图尺寸逐渐收缩、通道数逐渐增加的全卷积神经网络结构，可以接收任意尺寸的输入图像，卷积层和池化层交替工作，用来捕捉上下图层的信息，并逐步对特征图进行下采样，以恢复图像大小。在网络训练向前传播的过程中，编码器逐渐收缩，以减少池化层的空间维度，通过池化层扩大了感受野，定位训练集数据的各像素级。本实施例的卷积层均优选使用3*3*3卷积核，最大池化层均使用2*2*2池化核。本申请其他实施例中卷积核与池化核的大小可调，在不同的网络中可以不同，同一网络中的不同卷积层和池化层也可以不同。
S121：通过与所述编码器相连的解码器逐步修复所述训练集数据的细节和空间维度，对各所述像素级进行分类，以形成识别所述股骨近端的分割边界的所述分割模型。
本实施例的解码器是一个与编码器具有对称的结构，其特征图尺寸逐渐扩张，通道数逐渐减少，逐步修复物体的细节和空间维度，并通过对特征图进行上采样，逐步恢复高分辨率的图片细节。本实施例的编码器的池化层虽然扩大了感受野但造成了位置信息的丢失，像素级分类要求保留位置信息。本实施例的编码器和解码器之间通常存在跳跃连接，跳跃连接能够将低层次的特征图和高层次的特征图结合起来，能帮助解码器更好地修复目标的细节，有利于像素级的分类。在上采样部分，大量的特征通道能向更高分辨率的层传送上下图层信息，从而对每个像素都产生了一个预测，同时对最后一个卷积层的特征图进行上采样,使它恢复到输入图像相同的尺寸，同时保留了原始输入图像中的空间信息,最后在上采样的特征图上进行像素分类，进而识别股骨近端的分割边界。
进一步地，上述步骤S121，包括：
S1210：将所述解码器逐步修复所述训练数据的细节和空间维度之后的输出信息，输入到与各解码器相连的分支网络进行训练。
本实施例通过3D U-net在卷积训练过程中输入训练集数据的图像经过卷积核、池化核之后会图像 尺寸变小，需要通过反卷积进行上采样，得到与原图一样大小的映射层，但上采样倍数越大会丢失越多的细节，所以需要将不同层的不同上采样的结果进行综合来得到最后的结果，本实施例通过下层输出层L，中层输出层M和上层输出层U的三个输出分别代表在不同上采样倍数下的结果。比如输入图像大小是64*64*64，在下层输出层L之前的大小是16*16*16，需要进行4倍的上采样才能获取原图大小，而在中层输出层M之前的大小是32*32*32，需要进行2倍的上采样才能获取原图大小。本实施例的解码器各层的输出，都分别作为一个分支网络的输入，分支网络最终分别得到下层输出层L，中层输出层M和上层输出层U的三个输出。以便通过三个输出的交叉熵损失函数来表示3D U-net的损失函数，以便进一步优化整个3D U-net网络权重，使抽象表述更准确、更接近全局最优，取得更好的分割效果。
S1211：通过分析所述分支网络的输出结果，获取优化权重参量的所述分割模型的。
本实施例的3D U-net的损失函数定义为分支网络的三个输出交叉熵损失函数的加和，具体形式如下：
Figure PCTCN2018095496-appb-000001
Figure PCTCN2018095496-appb-000002
其中，W是3D U-net的主网络的权重，ω l,ω m,ω u分别是分支网络的下层输出层L，中层输出层M和上层输出层U的权重，χ代表训练数据，
Figure PCTCN2018095496-appb-000003
是进一步避免过拟合的正则项。本实施例的损失函数越大说明差别越大，损失函数越小差别越小。本实施例通过主网络的权重与分支网络的权重共同作用于损失函数，上述损失函数收敛时获取分割模型的优化权重参量，即上述损失函数收敛时对应的主网络的权重与分支网络的权重。本实施例的3D U-net的损失函数定义为分支网络的三个输出交叉熵损失函数的加和，使主网络的权重与分支网络的权重共同影响着损失函数的大小，影响着分割模型预测得到的股骨近端的分割边界与训练集中输入的已知股骨近端的分割边界的差距。上述公式中c∈{l,m,u}，即c是一个指代符，指代{l,m.u}中的一个，写在求和符号Σ下面表示，把c为取值为{l,m,u}三个的值进行求和。即下述公式
Figure PCTCN2018095496-appb-000004
等价于：
Figure PCTCN2018095496-appb-000005
进一步地，本申请再一实施例中，步骤S12之前，包括：
S122：将原始3D MRI图像数据以及对原始3D MRI图像数据旋转指定角度后的增强数据组成数据集。
本实施例使用少量的标注样本进行分割模型训练，为防止过拟合的发生，对原始数据进行了增强操作。本实施例通过对原始数据进行旋转指定角度进行数据增强。举例地，对原始数据进行(90度、180 度、270度)旋转，因为90度、180度、270度均匀分布在0-360度，通过旋转此三个角度，使增强数据均匀变化。
S123：将所述数据集中的每个3D MRI图像数据切割成指定数量和指定大小的子块。
本步骤通过从增强数据集的每个图片中随机切割子块作为训练集数据，本实施例中每个图片中随机切割出多个64*64*64的子块，并选出10个作为训练集数据。
S124：将所有的所述子块随机分为两组，一组作为训练集数据，另一组作为测试集数据。
举例地，原始图像共10个，做三个角度数据旋转增强之后变为30个图像，每个图像切出10个作为训练集数据，则训练集数据总共有300个。本申请其他实施例将上述训练集数据标准化到均值为0、方差为1。本实施例通过数据标准化也称为归一化，将不同量纲和不同数量级大小的数据变成可以相互进行数学计算和具有可比性的数据，归一化使得数据在各个维度的分布比较接近，能够适当扩大数据差异，加速模型训练过程的收敛。本实施例的均值为0方差为1的计算过程如下：比如一组数据为：
X:{x1,x2,x3,.....,xn}，设其平均值为：E(X)＝μ，设其方差为：D(X)＝σ 2，那么经过变换：y＝(x-μ)/σ，得到的新变量：Y:{y1,y2,y3,......yn}的均值为0、方差为1。
进一步地，本申请又一实施例中，步骤S3之后，包括：
S30：获取分割后的股骨近端的病变区域的位置信息以及病变区域的分布信息。
本实施例的病变区域的位置信息以及病变区域的分布信息，通过预训练的病变区域分割模型识别病变区域的边缘信息获得。病变区域的位置信息可通过位于股骨近端的坐标信息获得，病变区域的分布信息包括病变区域的数量、各病变区域的位置信息的相对关系等，比如，病变位置为髋臼处，病变区域的分布信息为一个病变区域，具体为髋臼过度覆盖且髋臼窝加深；再比如，病变位置为股骨头颈交区，病变区域的分布信息为一个病变区域，具体为股骨头颈交区存在突起等。
S31：通过分析所述病变区域的位置信息以及病变区域的分布信息，输出股骨髋臼撞击的病症信息。
比如：髋臼过度覆盖且髋臼窝加深，则表示盂唇撕裂、软骨损伤易发生于前上、后下部等病症信息。本步骤可以通过预先建立的分析模型来获取分割后的股骨髋臼撞击的病症信息，分析模型的网络结构包括但不限于卷积神经网络、全卷积神经网络或U-net。
进一步地，本申请又一实施例中，步骤S31之后，包括：
S32：将所述病症信息与所述病例的身份信息、确诊反馈信息汇集成诊断股骨髋臼撞击的先验数据库。
本步骤的身份信息包括但不限于性别、年龄、运动喜好等，所述的确诊反馈信息包括但不限于髋臼异常、股骨头异常等，所述病症信息包括各病症信息分别对应的特征矩阵。本实施例通过形成先验数据库，以提高类似病症的再次诊断效率和诊断精准度。
进一步地，本申请又一实施例中，步骤S32之后，包括：
S33：在所述先验数据库搜寻与新病症信息相似度最高的历史病症信息。
本步骤通过比较分析模型输出的新输出病症信息的特征矩阵和先验数据库中各病症信息的特征矩阵的相似度，实现相似病症的诊断。
S34：输出所述历史病症信息对应的确诊反馈信息。
本实施例通过将先验数据库结合影像分割模型，推广到现实的自动化病症诊断中，提高医生的诊断效率和准确性。
本实施例通过从3D MRI图像中通过分割模型自动分离出来股骨近端，通过将股骨近端从原图中分离出来，减少诊断干扰信息，极大的提高医生的诊断效率；本申请提出了基于3DU-net的3D MRI股骨近端分割技术，通过具有深度监督学习效果的3DU-net网络，利用少量标注样本训练获得了精准的分割模型，实现对3D MRI股骨近端的精准分割，弥补了现有标注的3D MRI图像数据困乏，难以获得精准分割的技术难题；通过汇集病变诊断数据形成先验数据库，辅助提高医生诊断病症的诊断准确度，弥补医生因经验不足而导致诊断准确度不高的缺陷，具有实际应用价值。
参照图2，本申请一实施例的分割股骨近端的装置，包括：
第一输入模块1，用于将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中。
本实施例的3D MRI(Magnetic Resonance Imaging，图像磁共振成像)为通过核磁共振信号进行空间编码的“数字图像”，磁共振信号直接来自于物体本身，磁共振成像可以得到物体任何方向的断层图像、三维体图像，可以重建出物体图像，比如以不同灰度显示不同人体股骨结构的解剖和病理的断面图像。本实施例的3D U-net是分割型网络模型，整个网络呈“U”形，也是网络名为U-net的来源。在“U”字的下行区域属于编辑器，上行区域属于解码器。3DU-net网络是深度监督学习网络，所谓深度监督学习是指：利用一组已知正确答案的样本调整分类器的参数，使其达到所要求性能的过程。对于深度监督学习网络就是利用有标签的数据来学习的网络，被初始化的网络根据预测值与标签的差别不断修改网络中的参数，使网络的预测值越来越接近标签，以达到学习的目的，可利用少量标注样本训练获得了精准的分割模型，实现对股骨的3D MRI股骨近端的精准分割，弥补了现有标注的3D MRI图像数据困乏，难以获得精准分割的技术难题，且训练时间短，降低了模型成本。
识别模块2，用于通过所述分割模型识别所述3D MRI图像中股骨近端的分割边界。
3D U-net的编辑器各层对输入的3D MRI数据进行卷积和池化，以进行特征提取，解码器各层使用反卷积对提取的特征进行解码得到映射层并输出，映射层的大小与输入图像大小相同，映射层指示出原始3D MRI图的每一部分代表的含义，即分割结果，本实施例通过3D U-net训练的分割模型识别出原始3D MRI图的哪一部分是股骨近端。
分割模块3，用于根据所述分割边界对所述3D MRI图像中的股骨近端进行分割。
本实施例通过3D U-net训练得到分割模型分割3D MRI的股骨近端，使得分割数据的准确率更高， 具有较高的实用价值。本实施例通过从3D MRI图像中通过分割模型自动分离出来股骨近端，通过将股骨近端从原图中分离出来，减少诊断干扰信息，极大的提高医生的诊断效率。
参照图3，本实施例的分割股骨近端的装置，包括：
第一获取模块11，用于在所述3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化参数。
本实施例的初始化参数包括随机的初始化参量或优化的初始化参量。本实施例通过监督学习训练3D U-net时，通过权值矩阵判断训练过程是否收敛，当两次迭代之间的权值变化很小，比如变化值小于1％，则认为训练收敛，训练收敛时3D U-net编码器和解码器的参数达到最大优化，以确定优化状态下的网络权重值矩阵，以便输出准确率高的分割模型。
第二输入模块12，用于在所述优化参数下，将预设训练集数据输入3D U-net训练所述分割模型。
本实施例通过由具有详细标注信息的3D MRI图像以及每个3D MRI图像对应的股骨近端的分割边界组成的训练集数据，输入到3D U-net进行训练，以获取到股骨近端边缘的特征的提取方式，以便训练出可以从3D MRI图像中识别股骨近端的分割边界的分割模型。
判断模块13，用于通过将预设的测试集数据输入最佳参量下的分割模型，判断所述分割模型的准确率是否达到预设阈值。
本实施例的测试集数据与训练集数据均来自同一样本集且经过同样的数据处理，但测试集数据与训练集无交集。本实施例通过把数据集分为训练集与测试集，在训练集上训练模型，通过测试集来测试模型效果。本实施例中分割模型的准确率通过Dice系数来衡量，Dice系数是集合相似度度量函数，本实施例的Dice系数计算公式为：2*|X∩Y|/(|X|+|Y|)，X是预测区域，Y是图层表面实况，即标注区域(值域范围：0-1)。本实施例的分割模型在测试集上的最终Dice系数达到0.98时，表示测试集与原始标注的训练集重合度非常高。
确定模块14，用于若分割模型的准确率达到预设阈值，则确定所述分割模型满足应用需求。
进一步地，本实施例的分割股骨近端的装置，包括：
初始化模块10，用于将C3D的卷积层网络中与所述3D U-net编译器具有相同特征图个数的各层参数迁移至所述3D U-net编译器中作为所述3D U-net的初始化参数。
为避免过拟合，训练集数据的集合应该足够大，但训练耗时太长，为了减小训练时间并防止过拟合，本实施例通过迁移学习把C3D(3D convolution，三维卷积网络)的网络参数作为3D U-net编译器的各层参数。C3D网络结构共有8个卷积层，4次池化层。其中卷积核的大小均为3*3*3，池化核的大小为2*2*2。且前6个卷积层的特征图个数(64,128,256,256,512,512)与3DU-net网络中对应编辑器各层的特征图个数(64,128,256,256,512,512)相同，所以卷积层的参数个数相同，通过已经训练好的C3D的卷积层参数初始化3D U-net网络的卷积层，即编码器的各层参数，所以本实施例通过获取C3D模型 的前6个卷积层的参数，用于初始化3D U-net网络的编译器的各层参数。本实施例通过已经训练好的C3D的卷积层参数初始化3D U-net网络的卷积层的过程为迁移学习过程，本实施例C3D网络与3D U-net网络的训练过程中使用的数据不同。本实施例通过迁移学习在保证训练效果的同时可以减少训练时间，通过迁移学习训练的网络模型相比于通过随机初始化的方式初始化整个网络的训练效果更优化。
参照图4，本实施例的第二输入模块12，包括：
定位单元120，用于将所述训练集数据输入3D U-net，通过3D U-net的编码器减少池化层的空间维度，定位训练集数据的各像素级。
本实施例的编码器是一个特征图尺寸逐渐收缩、通道数逐渐增加的全卷积神经网络结构，可以接收任意尺寸的输入图像，卷积层和池化层交替工作，用来捕捉上下图层的信息，并逐步对特征图进行下采样，以恢复图像大小。在网络训练向前传播的过程中，编码器逐渐收缩，以减少池化层的空间维度，通过池化层扩大了感受野，定位训练集数据的各像素级。本实施例的卷积层均优选使用3*3*3卷积核，最大池化层均使用2*2*2池化核。本申请其他实施例中卷积核与池化核的大小可调，在不同的网络中可以不同，同一网络中的不同卷积层和池化层也可以不同。
分类单元121，用于通过与所述编码器相连的解码器逐步修复所述训练集数据的细节和空间维度，对各所述像素级进行分类，以形成识别所述股骨近端的分割边界的分割模型。
本实施例的解码器是一个与编码器具有对称的结构，其特征图尺寸逐渐扩张，通道数逐渐减少，逐步修复物体的细节和空间维度，并通过对特征图进行上采样，逐步恢复高分辨率的图片细节。本实施例的编码器的池化层虽然扩大了感受野但造成了位置信息的丢失，像素级分类要求保留位置信息。本实施例的编码器和解码器之间通常存在跳跃连接，跳跃连接能够将低层次的特征图和高层次的特征图结合起来，能帮助解码器更好地修复目标的细节，有利于像素级的分类。在上采样部分，大量的特征通道能向更高分辨率的层传送上下图层信息，从而对每个像素都产生了一个预测，同时对最后一个卷积层的特征图进行上采样,使它恢复到输入图像相同的尺寸，同时保留了原始输入图像中的空间信息,最后在上采样的特征图上进行像素分类，进而识别股骨近端的分割边界。
参照图5，本实施例的分类单元121，包括：
训练子单元1210，用于将所述解码器逐步修复所述训练数据的细节和空间维度之后的输出信息，输入到与各解码器相连的分支网络进行训练。
本实施例通过3D U-net在卷积训练过程中输入训练集数据的图像经过卷积核、池化核之后会图像尺寸变小，需要通过反卷积进行上采样，得到与原图一样大小的映射层，但上采样倍数越大会丢失越多的细节，所以需要将不同层的不同上采样的结果进行综合来得到最后的结果，本实施例通过下层输出层L，中层输出层M和上层输出层U的三个输出分别代表在不同上采样倍数下的结果。比如输入图像大小是64*64*64，在下层输出层L之前的大小是16*16*16，需要进行4倍的上采样才能获取原图大小，而 在中层输出层M之前的大小是32*32*32，需要进行2倍的上采样才能获取原图大小。本实施例的解码器各层的输出，都分别作为一个分支网络的输入，分支网络最终分别得到下层输出层L，中层输出层M和上层输出层U的三个输出。以便通过三个输出的交叉熵损失函数来表示3D U-net的损失函数，以便进一步优化整个3D U-net网络权重，使抽象表述更准确、更接近全局最优，取得更好的分割效果。
获取子单元1211，用于通过分析所述分支网络的输出结果，获取优化权重参量的所述分割模型。
本实施例的3D U-net的损失函数定义为分支网络的三个输出交叉熵损失函数的加和，具体形式如下：
Figure PCTCN2018095496-appb-000006
Figure PCTCN2018095496-appb-000007
其中，W是3D U-net的主网络的权重，ω l,ω m,ω u分别是分支网络的下层输出层L，中层输出层M和上层输出层U的权重，χ代表训练数据，
Figure PCTCN2018095496-appb-000008
是进一步避免过拟合的正则项。本实施例的损失函数越大说明差别越大，损失函数越小差别越小。本实施例通过主网络的权重与分支网络的权重共同作用于损失函数，上述损失函数收敛时获取分割模型的优化权重参量，即上述损失函数收敛时对应的主网络的权重与分支网络的权重。本实施例的3D U-net的损失函数定义为分支网络的三个输出交叉熵损失函数的加和，使主网络的权重与分支网络的权重共同影响着损失函数的大小，影响着分割模型预测得到的股骨近端的分割边界与训练集中输入的已知股骨近端的分割边界的差距。上述公式中c∈{l,m,u}，即c是一个指代符，指代{l,m.u}中的一个，写在求和符号Σ下面表示，把c为取值为{l,m,u}三个的值进行求和。即下述公式
Figure PCTCN2018095496-appb-000009
等价于：
ζ c(x；W,ω l,ω m,ω u)＝α lζ l(x；W,ω l)+λ(ψ(W)+ψ(ω l))
+α mζ m(x；W,ω m)+λ(ψ(W)+ψ(ω m))
+α uζ u(x；W,ω l)+λ(ψ(W)+ψ(ω l))
参照图6，本申请再一实施例的分割股骨近端的装置，包括：
组成模块122，用于将原始3D MRI图像数据以及对原始3D MRI图像数据旋转指定角度后的增强数据组成数据集。
本实施例使用少量的标注样本进行分割模型训练，为防止过拟合的发生，对原始数据进行了增强操作。本实施例通过对原始数据进行旋转指定角度进行数据增强。举例地，对原始数据进行(90度、180度、270度)旋转，因为90度、180度、270度均匀分布在0-360度，通过旋转此三个角度，使增强数据均匀变化。
切割模块123，用于将所述数据集中的每个3D MRI图像数据切割成指定数量和指定大小的子块。
本实施例通过从增强数据集的每个图片中随机切割子块作为训练集数据，本实施例中每个图片中随机切割出多个64*64*64的子块，并选出10个作为训练集数据。
区分模块124，用于将所有的所述子块随机分为两组，一组作为训练集数据，另一组作为测试集数据。
举例地，原始图像共10个，做三个角度数据旋转增强之后变为30个图像，每个图像切出10个作为训练集数据，则训练集数据总共有300个。本申请其他实施例将上述训练集数据标准化到均值为0、方差为1。本实施例通过数据标准化也称为归一化，将不同量纲和不同数量级大小的数据变成可以相互进行数学计算和具有可比性的数据，归一化使得数据在各个维度的分布比较接近，能够适当扩大数据差异，加速模型训练过程的收敛。本实施例的均值为0方差为1的计算过程如下：比如一组数据为：
X:{x1,x2,x3,.....,xn}，设其平均值为：E(X)＝μ，设其方差为：D(X)＝σ 2，那么经过变换：y＝(x-μ)/σ，得到的新变量：Y:{y1,y2,y3,......yn}的均值为0、方差为1。
参照图7，本申请又一实施例的分割股骨近端的装置，包括：
第二获取模块30，用于获取分割后的股骨近端的病变区域的位置信息以及病变区域的分布信息。
本实施例的病变区域的位置信息以及病变区域的分布信息，通过预训练的病变区域分割模型识别病变区域的边缘信息获得。病变区域的位置信息可通过位于股骨近端的坐标信息获得，病变区域的分布信息包括病变区域的数量、各病变区域的位置信息的相对关系等，比如，病变位置为髋臼处，病变区域的分布信息为一个病变区域，具体为髋臼过度覆盖且髋臼窝加深；再比如，病变位置为股骨头颈交区，病变区域的分布信息为一个病变区域，具体为股骨头颈交区存在突起等。
第一输出模块31，用于通过分析所述病变区域的位置信息以及病变区域的分布信息，输出股骨髋臼撞击的病症信息。
比如：髋臼过度覆盖且髋臼窝加深，则表示盂唇撕裂、软骨损伤易发生于前上、后下部等病症信息。本实施例可以通过预先建立的分析模型来获取分割后的股骨髋臼撞击的病症信息，分析模型的网络结构包括但不限于卷积神经网络、全卷积神经网络或U-net。
参照图8，本申请又一实施例的分割股骨近端的装置，包括：
汇集模块32，用于将所述病症信息与所述病例的身份信息、确诊反馈信息汇集成诊断股骨髋臼撞击的先验数据库。
本实施例的身份信息包括但不限于性别、年龄、运动喜好等，所述的确诊反馈信息包括但不限于髋臼异常、股骨头异常等，所述病症信息包括各病症信息分别对应的特征矩阵。本实施例通过形成先验数据库，以提高类似病症的再次诊断效率和诊断精准度。
参照图9，本申请又一实施例的分割股骨近端的装置，包括：
搜寻模块33，用于在所述先验数据库搜寻与新病症信息相似度最高的历史病症信息。
本实施例通过比较分析模型输出的新输出病症信息的特征矩阵和先验数据库中各病症信息的特征矩阵的相似度，实现相似病症的诊断。
第二输出模块34，用于输出所述历史病症信息对应的确诊反馈信息。
本实施例通过将先验数据库结合影像分割模型，推广到现实的自动化病症诊断中，提高医生的诊断效率和准确性。
参照图10，本申请实施例中还提供一种计算机设备，该计算机设备可以是服务器，其内部结构可以如图9所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设计的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储分割股骨近端等数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令在执行时，执行如上述各方法的实施例的流程。本领域技术人员可以理解，图9中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定。
本申请一实施例还提供一种计算机非易失性可读存储介质，其上存储有计算机可读指令，该计算机可读指令在执行时，执行如上述各方法的实施例的流程。以上所述仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种分割股骨近端的方法，其特征在于，包括：
    将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中；
    通过所述分割模型识别所述3D MRI图像中股骨近端的分割边界；
    根据所述分割边界对所述3D MRI图像中的股骨近端进行分割。
  2. 根据权利要求1所述的分割股骨近端的方法，其特征在于，所述将3D MRI图像输入到通过3D U-net预先训练得到的分割模型中的步骤之前，包括：
    在所述3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化参数；
    在所述优化参数下，将预设训练集数据输入3D U-net训练所述分割模型；
    通过将预设的测试集数据输入所述分割模型，判断所述分割模型的准确率是否达到预设阈值；
    若达到，则确定所述分割模型满足应用需求。
  3. 根据权利要求2所述的分割股骨近端的方法，其特征在于，所述在所述3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化参数的步骤之前，包括：
    将C3D的卷积层网络中与所述3D U-net编译器具有相同特征图个数的各层参数迁移至所述3D U-net编译器中作为所述3D U-net的初始化参数。
  4. 根据权利要求2所述的分割股骨近端的方法，其特征在于，所述在所述优化参数下，将预设训练集数据输入3D U-net训练所述分割模型的步骤，包括：
    将所述训练集数据输入3D U-net，通过3D U-net的编码器减少池化层的空间维度，定位训练集数据的各像素级；
    通过与所述编码器相连的解码器逐步修复所述训练集数据的细节和空间维度，对各所述像素级进行分类，以形成识别所述股骨近端的分割边界的所述分割模型。
  5. 根据权利要求4所述的分割股骨近端的方法，其特征在于，所述通过与所述编码器相连的解码器逐步修复所述训练数据的细节和空间维度，对各所述像素级进行分类，以形成识别所述股骨近端分割边界的所述分割模型的步骤，包括：
    将所述解码器逐步修复所述训练数据的细节和空间维度之后的输出信息，输入到与各解码器相连的分支网络进行训练；
    通过分析所述分支网络的输出结果，获取优化权重参量的所述分割模型。
  6. 根据权利要求2所述的分割股骨近端的方法，其特征在于，所述在所述优化参数下，将预设训练集数据输入3D U-net训练所述分割模型的步骤之前，包括：
    将原始3D MRI图像数据以及对原始3D MRI图像数据旋转指定角度后的增强数据组成数据集；
    将所述数据集中的每个3D MRI图像数据切割成指定数量和指定大小的子块；
    将所有的所述子块随机分为两组，一组作为训练集数据，另一组作为测试集数据。
  7. 根据权利要求1所述的分割股骨近端的方法，其特征在于，所述根据所述分割边界对所述3D MRI图像中的股骨近端进行分割的步骤之后，包括：
    获取分割后的股骨近端的病变区域的位置信息以及病变区域的分布信息；
    通过分析所述病变区域的位置信息以及病变区域的分布信息，输出股骨髋臼撞击的病症信息。
  8. 一种分割股骨近端的装置，其特征在于，包括：
    第一输入模块，用于将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中；
    识别模块，用于通过所述分割模型识别所述3D MRI图像中股骨近端的分割边界；
    分割模块，用于根据所述分割边界对所述3D MRI图像中的股骨近端进行分割。
  9. 根据权利要求8所述的分割股骨近端的装置，其特征在于，包括：
    第一获取模块，用于在所述3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化参数；
    第二输入模块，用于在所述优化参数下，将预设训练集数据输入3D U-net训练所述分割模型；
    判断模块，用于通过将预设的测试集数据输入所述分割模型，判断所述分割模型的准确率是否达到预设阈值；
    确定模块，用于若分割模型的准确率达到预设阈值，则确定所述分割模型满足应用需求。
  10. 根据权利要求9所述的分割股骨近端的装置，其特征在于，包括：
    初始化模块，用于将C3D的卷积层网络中与所述3D U-net编译器具有相同特征图个数的各层参数迁移至所述3D U-net编译器中作为所述3D U-net的初始化参数。
  11. 根据权利要求9所述的分割股骨近端的装置，其特征在于，所述第二输入模块，包括：
    定位单元，用于将所述训练集数据输入3D U-net，通过3D U-net的编码器减少池化层的空间维度，定位训练集数据的各像素级；
    分类单元，用于通过与所述编码器相连的解码器逐步修复所述训练集数据的细节和空间维度，对各所述像素级进行分类，以形成识别所述股骨近端的分割边界的所述分割模型。
  12. 根据权利要求11所述的分割股骨近端的装置，其特征在于，所述分类单元，包括：
    训练子单元，用于将所述解码器逐步修复所述训练数据的细节和空间维度之后的输出信息，输入到与各解码器相连的分支网络进行训练；
    获取子单元，用于通过分析所述分支网络的输出结果，获取优化权重参量的所述分割模型。
  13. 根据权利要求9所述的分割股骨近端的装置，其特征在于，包括：
    组成模块，用于将原始3D MRI图像数据以及对原始3D MRI图像数据旋转指定角度后的增强数据组 成数据集；
    切割模块，用于将所述数据集中的每个3D MRI图像数据切割成指定数量和指定大小的子块；
    区分模块，用于将所有的所述子块随机分为两组，一组作为训练集数据，另一组作为测试集数据。
  14. 根据权利要求8所述的分割股骨近端的装置，其特征在于，包括：
    第二获取模块，用于获取分割后的股骨近端的病变区域的位置信息以及病变区域的分布信息；
    第一输出模块，用于通过分析所述病变区域的位置信息以及病变区域的分布信息，输出股骨髋臼撞击的病症信息。
  15. 一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现分割股骨近端的方法，方法包括：
    将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中；
    通过所述分割模型识别所述3D MRI图像中股骨近端的分割边界；
    根据所述分割边界对所述3D MRI图像中的股骨近端进行分割。
  16. 根据权利要求15所述的计算机设备，其特征在于，所述将3D MRI图像输入到通过3D U-net预先训练得到的分割模型中的步骤之前，包括：
    在所述3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化参数；
    在所述优化参数下，将预设训练集数据输入3D U-net训练所述分割模型；
    通过将预设的测试集数据输入所述分割模型，判断所述分割模型的准确率是否达到预设阈值；
    若达到，则确定所述分割模型满足应用需求。
  17. 根据权利要求16所述的计算机设备，其特征在于，所述在所述3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化参数的步骤之前，包括：
    将C3D的卷积层网络中与所述3D U-net编译器具有相同特征图个数的各层参数迁移至所述3D U-net编译器中作为所述3D U-net的初始化参数。
  18. 一种计算机非易失性可读存储介质，其上存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现分割股骨近端的方法，方法包括：
    将股骨的3D MRI图像输入到通过3D U-net预先训练得到的分割模型中；
    通过所述分割模型识别所述3D MRI图像中股骨近端的分割边界；
    根据所述分割边界对所述3D MRI图像中的股骨近端进行分割。
  19. 根据权利要求18所述的计算机非易失性可读存储介质，其特征在于，所述将3D MRI图像输入到通过3D U-net预先训练得到的分割模型中的步骤之前，包括：
    在所述3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化 参数；
    在所述优化参数下，将预设训练集数据输入3D U-net训练所述分割模型；
    通过将预设的测试集数据输入所述分割模型，判断所述分割模型的准确率是否达到预设阈值；
    若达到，则确定所述分割模型满足应用需求。
  20. 根据权利要求19所述的计算机非易失性可读存储介质，其特征在于，所述在所述3D U-net的初始化参数下通过监督学习获取到3D U-net的编码器和解码器分别对应的优化参数的步骤之前，包括：
    将C3D的卷积层网络中与所述3D U-net编译器具有相同特征图个数的各层参数迁移至所述3D U-net编译器中作为所述3D U-net的初始化参数。
PCT/CN2018/095496 2018-04-20 2018-07-12 分割股骨近端的方法、装置、计算机设备和存储介质 WO2019200747A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810362198.6A CN108764241A (zh) 2018-04-20 2018-04-20 分割股骨近端的方法、装置、计算机设备和存储介质
CN201810362198.6 2018-04-20

Publications (1)

Publication Number Publication Date
WO2019200747A1 true WO2019200747A1 (zh) 2019-10-24

Family

ID=64011191

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/095496 WO2019200747A1 (zh) 2018-04-20 2018-07-12 分割股骨近端的方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN108764241A (zh)
WO (1) WO2019200747A1 (zh)

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111008643A (zh) * 2019-10-29 2020-04-14 平安科技（深圳）有限公司 基于半监督学习的图片分类方法、装置和计算机设备
CN111145188A (zh) * 2019-12-25 2020-05-12 西安电子科技大学 一种基于ResNet与UNet模型的图像分割方法
CN111179229A (zh) * 2019-12-17 2020-05-19 中信重工机械股份有限公司 一种基于深度学习的工业ct缺陷检测方法
CN111369565A (zh) * 2020-03-09 2020-07-03 麦克奥迪（厦门）医疗诊断系统有限公司 一种基于图卷积网络的数字病理图像的分割与分类方法
CN111429474A (zh) * 2020-02-27 2020-07-17 西北大学 基于混合卷积的乳腺dce-mri图像病灶分割模型建立及分割方法
CN111461053A (zh) * 2020-04-13 2020-07-28 安徽大学 基于迁移学习的多个生育期小麦倒伏区域识别系统
CN111507993A (zh) * 2020-03-18 2020-08-07 南方电网科学研究院有限责任公司 一种基于生成对抗网络的图像分割方法、装置及存储介质
CN111714145A (zh) * 2020-05-27 2020-09-29 浙江飞图影像科技有限公司 基于弱监督分割的股骨颈骨折检测方法及系统
CN111784653A (zh) * 2020-06-28 2020-10-16 西安电子科技大学 基于形状约束的多尺度网络mri胰腺轮廓定位方法
CN111784704A (zh) * 2020-06-24 2020-10-16 中国人民解放军空军军医大学 Mri髋关节炎症分割与分类自动定量分级序贯方法
CN112419267A (zh) * 2020-11-23 2021-02-26 齐鲁工业大学 基于深度学习的脑胶质瘤分割模型及分割方法
CN112613374A (zh) * 2020-12-16 2021-04-06 厦门美图之家科技有限公司 人脸可见区域解析与分割方法、人脸上妆方法及移动终端
CN112734723A (zh) * 2021-01-08 2021-04-30 温州医科大学 一种面向多源数据的乳腺肿瘤图像分类预测方法及装置
CN113496494A (zh) * 2021-06-17 2021-10-12 北京理工大学 基于drr模拟数据生成的二维骨骼分割方法及装置
CN117668497A (zh) * 2024-01-31 2024-03-08 山西卓昇环保科技有限公司 基于深度学习实现环境保护下的碳排放分析方法及系统

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109859233B (zh) * 2018-12-28 2020-12-11 上海联影智能医疗科技有限公司 图像处理、图像处理模型的训练方法及系统
CN109754389B (zh) * 2018-12-07 2021-08-24 北京市商汤科技开发有限公司 一种图像处理方法、装置及设备
CN111489357A (zh) * 2019-01-29 2020-08-04 广州市百果园信息技术有限公司 一种图像分割方法、装置、设备及存储介质
CN109919932A (zh) * 2019-03-08 2019-06-21 广州视源电子科技股份有限公司 目标对象的识别方法和装置
CN110148114A (zh) * 2019-04-02 2019-08-20 成都真实维度科技有限公司 一种基于2d断层扫描图数据集的深度学习模型训练方法
CN110298844B (zh) * 2019-06-17 2021-06-29 艾瑞迈迪科技石家庄有限公司 X光造影图像血管分割与识别方法及装置
JP7226199B2 (ja) * 2019-09-04 2023-02-21 株式会社島津製作所 画像解析方法、画像処理装置および骨密度測定装置
CN110717060B (zh) * 2019-09-04 2023-08-18 平安科技（深圳）有限公司 图像mask的过滤方法、装置及存储介质
CN110738107A (zh) * 2019-09-06 2020-01-31 上海衡道医学病理诊断中心有限公司 一种具备模型迁移功能的显微图像识别分割方法
CN111193254B (zh) * 2019-12-06 2021-10-29 北京国电通网络技术有限公司 一种住宅日用电负荷预测方法和设备
CN111260055B (zh) * 2020-01-13 2023-09-01 腾讯科技（深圳）有限公司 基于三维图像识别的模型训练方法、存储介质和设备
CN111553882A (zh) * 2020-03-31 2020-08-18 江苏广宇淘璞网络科技有限公司 Mr图像股骨头缺血型坏死识别方法及装置
CN113744214B (zh) * 2021-08-24 2022-05-13 北京长木谷医疗科技有限公司 基于深度强化学习的股骨柄放置装置及电子设备
CN113920128B (zh) * 2021-09-01 2023-02-21 北京长木谷医疗科技有限公司 膝关节股骨胫骨分割方法和装置
CN113822231A (zh) * 2021-11-08 2021-12-21 中国人民解放军陆军特色医学中心 一种基于深度学习图像识别的转子间骨折手术辅助系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107358608A (zh) * 2017-08-23 2017-11-17 西安邮电大学 基于图像处理技术的骨组织几何形态学参数自动测量装置及方法
CN107680088A (zh) * 2017-09-30 2018-02-09 百度在线网络技术（北京）有限公司 用于分析医学影像的方法和装置
US20180061059A1 (en) * 2016-08-26 2018-03-01 Elekta, Inc. System and methods for image segmentation using convolutional neural network
CN107909581A (zh) * 2017-11-03 2018-04-13 杭州依图医疗技术有限公司 Ct影像的肺叶段分割方法、装置、系统、存储介质及设备

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107886510A (zh) * 2017-11-27 2018-04-06 杭州电子科技大学 一种基于三维全卷积神经网络的前列腺mri分割方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180061059A1 (en) * 2016-08-26 2018-03-01 Elekta, Inc. System and methods for image segmentation using convolutional neural network
CN107358608A (zh) * 2017-08-23 2017-11-17 西安邮电大学 基于图像处理技术的骨组织几何形态学参数自动测量装置及方法
CN107680088A (zh) * 2017-09-30 2018-02-09 百度在线网络技术（北京）有限公司 用于分析医学影像的方法和装置
CN107909581A (zh) * 2017-11-03 2018-04-13 杭州依图医疗技术有限公司 Ct影像的肺叶段分割方法、装置、系统、存储介质及设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ABDULKADIR, A. ET AL.: "3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation", MEDICAL IMAGE COMPUTING AND COMPUTER-ASSISTED INTERVENTION - MICCAI 2016, 2 October 2016 (2016-10-02), pages 424 - 432, XP047392527, DOI: 10.1007/978-3-319-46723-8_49 *

Cited By (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111008643B (zh) * 2019-10-29 2024-03-19 平安科技（深圳）有限公司 基于半监督学习的图片分类方法、装置和计算机设备
CN111008643A (zh) * 2019-10-29 2020-04-14 平安科技（深圳）有限公司 基于半监督学习的图片分类方法、装置和计算机设备
CN111179229A (zh) * 2019-12-17 2020-05-19 中信重工机械股份有限公司 一种基于深度学习的工业ct缺陷检测方法
CN111179229B (zh) * 2019-12-17 2023-10-20 中信重工机械股份有限公司 一种基于深度学习的工业ct缺陷检测方法
CN111145188A (zh) * 2019-12-25 2020-05-12 西安电子科技大学 一种基于ResNet与UNet模型的图像分割方法
CN111145188B (zh) * 2019-12-25 2023-06-06 西安电子科技大学 一种基于ResNet与UNet模型的图像分割方法
CN111429474A (zh) * 2020-02-27 2020-07-17 西北大学 基于混合卷积的乳腺dce-mri图像病灶分割模型建立及分割方法
CN111429474B (zh) * 2020-02-27 2023-04-07 西北大学 基于混合卷积的乳腺dce-mri图像病灶分割模型建立及分割方法
CN111369565B (zh) * 2020-03-09 2023-09-15 麦克奥迪（厦门）医疗诊断系统有限公司 一种基于图卷积网络的数字病理图像的分割与分类方法
CN111369565A (zh) * 2020-03-09 2020-07-03 麦克奥迪（厦门）医疗诊断系统有限公司 一种基于图卷积网络的数字病理图像的分割与分类方法
CN111507993A (zh) * 2020-03-18 2020-08-07 南方电网科学研究院有限责任公司 一种基于生成对抗网络的图像分割方法、装置及存储介质
CN111461053A (zh) * 2020-04-13 2020-07-28 安徽大学 基于迁移学习的多个生育期小麦倒伏区域识别系统
CN111461053B (zh) * 2020-04-13 2023-05-23 安徽大学 基于迁移学习的多个生育期小麦倒伏区域识别系统
CN111714145A (zh) * 2020-05-27 2020-09-29 浙江飞图影像科技有限公司 基于弱监督分割的股骨颈骨折检测方法及系统
CN111784704A (zh) * 2020-06-24 2020-10-16 中国人民解放军空军军医大学 Mri髋关节炎症分割与分类自动定量分级序贯方法
CN111784704B (zh) * 2020-06-24 2023-11-24 中国人民解放军空军军医大学 Mri髋关节炎症分割与分类自动定量分级序贯方法
CN111784653B (zh) * 2020-06-28 2023-08-01 西安电子科技大学 基于形状约束的多尺度网络mri胰腺轮廓定位方法
CN111784653A (zh) * 2020-06-28 2020-10-16 西安电子科技大学 基于形状约束的多尺度网络mri胰腺轮廓定位方法
CN112419267A (zh) * 2020-11-23 2021-02-26 齐鲁工业大学 基于深度学习的脑胶质瘤分割模型及分割方法
CN112613374A (zh) * 2020-12-16 2021-04-06 厦门美图之家科技有限公司 人脸可见区域解析与分割方法、人脸上妆方法及移动终端
CN112734723B (zh) * 2021-01-08 2023-06-30 温州医科大学 一种面向多源数据的乳腺肿瘤图像分类预测方法及装置
CN112734723A (zh) * 2021-01-08 2021-04-30 温州医科大学 一种面向多源数据的乳腺肿瘤图像分类预测方法及装置
CN113496494A (zh) * 2021-06-17 2021-10-12 北京理工大学 基于drr模拟数据生成的二维骨骼分割方法及装置
CN117668497A (zh) * 2024-01-31 2024-03-08 山西卓昇环保科技有限公司 基于深度学习实现环境保护下的碳排放分析方法及系统
CN117668497B (zh) * 2024-01-31 2024-05-07 山西卓昇环保科技有限公司 基于深度学习实现环境保护下的碳排放分析方法及系统

Also Published As

Publication number Publication date
CN108764241A (zh) 2018-11-06

Similar Documents

Publication Publication Date Title
WO2019200747A1 (zh) 分割股骨近端的方法、装置、计算机设备和存储介质
CN111476292B (zh) 医学图像分类处理人工智能的小样本元学习训练方法
WO2020215984A1 (zh) 基于深度学习的医学图像检测方法及相关设备
CN107506761B (zh) 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
CN108062753B (zh) 基于深度对抗学习的无监督域自适应脑肿瘤语义分割方法
US20240062369A1 (en) Detection model training method and apparatus, computer device and storage medium
EP3665703B1 (en) Computer-aided diagnostics using deep neural networks
Bi et al. Automatic liver lesion detection using cascaded deep residual networks
CN106203432B (zh) 一种基于卷积神经网显著性图谱的感兴趣区域的定位系统
EP3252671A1 (en) Method of training a deep neural network
CN111461232A (zh) 一种基于多策略批量式主动学习的核磁共振图像分类方法
US20230043026A1 (en) Learning-based active surface model for medical image segmentation
EP2756456A1 (en) Method and system for the automatic analysis of an image of a biological sample
CN112862830B (zh) 一种多模态图像分割方法、系统、终端及可读存储介质
CN104484886B (zh) 一种mr图像的分割方法及装置
CN105760874A (zh) 面向尘肺的ct图像处理系统及其ct图像处理方法
CN108549912A (zh) 一种基于机器学习的医学图像肺结节检测方法
Abdullah et al. Multi-sectional views textural based SVM for MS lesion segmentation in multi-channels MRIs
CN112163599A (zh) 一种基于多尺度多层次融合的图像分类方法
Feng et al. Deep learning for chest radiology: a review
CN111798424A (zh) 一种基于医学图像的结节检测方法、装置及电子设备
Chen et al. The effect of kernel size of CNNs for lung nodule classification
Debnath et al. Brain tumour segmentation using memory based learning method
Fazilov et al. Patch-based lesion detection using deep learning method on small mammography dataset
CN115018780B (zh) 一种融合全局推理和mlp架构的甲状腺结节分割方法

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: 18915031

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: 18915031

Country of ref document: EP

Kind code of ref document: A1