A brain intracranial cavity segmentation method and device based on a U_Net model

Abstract

The application discloses a brain intracranial cavity segmentation method and device based on a U_Net model, and relates to the technical field of picture semantic segmentation.The application constructs a multistage down-sampling encoder and a corresponding up-sampling decoder in a U-shaped structure, adds an efficient attention mechanism to improve segmentation accuracy, and utilizes a multistage feature fusion module to fuse features; and the optimal segmentation model is trained on a training set.The down-sampling encoder fused with a residual block avoids the gradient vanishing phenomenon caused by excessively deep network, and the attention mechanism is introduced in the middle layer between the down-sampling and the up-sampling to further improve the sensitivity to the target region.Three sizes of convolution kernels are used to segment the image to improve the receptive field of the target region.The application has a deeper network and thus has stronger feature extraction capability, and can mine the target region in a more complex image.Multiple calculation modes make the model structure more flexible and variable, better adapt to business scenarios, and improve the generalization capability of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of picture semantic segmentation, and particularly relates to a brain intracranial cavity segmentation method and device based on a U_Net model. BACKGROUND

[0002] Brain dysplasia can cause low intelligence, movement disorders and language disorders in children after birth. The brain development of a fetus can only be judged by a doctor artificially through a magnetic resonance picture. A pediatrician in China serves more than 1,000 children, and it takes a lot of effort to look at the magnetic resonance picture. The fetal brain picture has a complex organizational structure, and there are not only brain tissues but also brain stems, cerebella, and mixed tissue fluids. The pictures taken by medical instruments are uneven, and some pictures have low resolution. These all cause great difficulty in diagnosing the fetal brain picture.

[0003] The brain development diagnosis method is often closely related to the experience of an imaging doctor. However, the same brain magnetic resonance picture has a high possibility of inconsistent results at different time points or by different doctors, and has a great observer difference. Therefore, in the face of a large number of patients and a shortage of imaging specialists, how to enable doctors without fetal brain development diagnosis experience to quickly master the imaging features becomes an urgent matter.

[0004] Picture semantic segmentation technology is one of the key technologies for researching computer picture segmentation algorithms. This technology is used to obtain a target region as a foreground from a picture, and the rest is a background, which is a binary classification algorithm. Under the assistance of deep learning, different from traditional segmentation methods, a neural network can effectively extract a target in a complex scene. In recent years, due to the development of deep learning, the field of medical picture segmentation has also become a hot spot, and the research on computer picture segmentation algorithms continues to break through. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a brain intracranial cavity segmentation method and device based on a U_Net model, which is used for assisting segmentation of a maternal magnetic resonance picture by a doctor without fetal brain diagnosis experience or improving the work efficiency of an experienced doctor.

[0006] In a first aspect, the present application provides a brain intracranial cavity segmentation method based on a U_Net model, comprising:

[0007] Step 10, acquiring a magnetic resonance image of a fetus in a mother and making a label picture according to a medical diagnosis sentence, and then performing preprocessing to obtain a training data set;

[0008] Step 20, construct a brain cranial magnetic resonance image lumen segmentation model, the model comprises a down-sampling module, an improved ECA attention module, an up-sampling module and a spatial pyramid fusion module connected in turn; wherein the down-sampling module first uses a depth separable convolution as a model convolution layer and a pooling layer, for reducing the picture size and increasing the picture dimension, and obtaining a specified dimension feature picture through multiple convolutions; the improved ECA attention module facilitates hierarchical attention by adding a log operation with high sensitivity when the value is low; the up-sampling module recovers the picture size and reduces the feature dimension layer by layer through deconvolution and adds the same size feature picture of the same dimension in the down-sampling process; the spatial pyramid fusion module is used for feature fusion of the picture obtained by the up-sampling part processing;

[0009] Step 30, input the training data set into the model for training to obtain a trained brain cranial magnetic resonance image lumen segmentation model;

[0010] Step 40, obtain the magnetic resonance image of the fetus to be diagnosed and input the trained brain cranial magnetic resonance image lumen segmentation model to obtain a brain cranial lumen region segmentation map.

[0011] Further, in the step 10, the magnetic resonance image of the fetus in the mother is obtained by magnetic resonance, and a group of transverse, sagittal and coronal photos of the same fetus are obtained at the same time.

[0012] Further, in the step 20, the improved ECA attention module specifically adds a log operation before the dot multiplication operation.

[0013] Further, in the step 20, the up-sampling module first recovers the picture size and reduces the feature dimension layer by layer through deconvolution, and the upper three layers use element-wise add to fuse the features; the outputs of the two features of the same layer of the down-sampling and the up-sampling are combined to generate a fusion feature set: the information of the down-sampling and the same layer information of the up-sampling , different weights are given to the information of each layer of the down-sampling , and the add fusion formula is as follows:

[0014] .

[0015] Further, in the step 20, the spatial pyramid fusion module uses the Adam optimizer to accelerate the convergence speed of the model, and the dice loss is used as a loss function to evaluate the difference between the segmentation sample and the label sample.

[0016] Secondly, the application provides a brain cranial lumen segmentation device based on a U_Net model, comprising a data set processing module, a model construction module, a model training module and a segmentation module.

[0017] The data set processing module is configured to acquire magnetic resonance images of fetuses in uteruses, make label pictures according to medical diagnosis sentences, and then perform preprocessing to obtain a training data set;

[0018] The model construction module is configured to construct a brain skull magnetic resonance image inner cavity segmentation model, the model comprising a down-sampling module, an improved ECA attention module, an up-sampling module, and a spatial pyramid fusion module connected in sequence; wherein the down-sampling module first uses a depth separable convolution as a model convolution layer and a pooling layer to reduce the size of a picture and increase the dimension of the picture, and obtains a feature picture of a specified dimension through multiple convolutions; the improved ECA attention module facilitates hierarchical attention by adding a log operation with high sensitivity when the value is low; the up-sampling module reduces the feature dimension by recovering the picture size layer by layer through deconvolution and adds the same dimension and size feature pictures in the down-sampling process; and the spatial pyramid fusion module is configured to fuse the features of the pictures obtained after the up-sampling part is processed;

[0019] The model training module is configured to input the training data set into the model for training to obtain a trained brain skull magnetic resonance image inner cavity segmentation model;

[0020] The segmentation module is configured to acquire magnetic resonance images of fetuses to be diagnosed and input the trained brain skull magnetic resonance image inner cavity segmentation model to obtain a brain skull inner cavity region segmentation map.

[0021] Further, in the data set processing module, the magnetic resonance images of fetuses in uteruses are acquired by using a magnetic resonance method, and a set of transverse, sagittal, and coronal pictures of the same fetus are obtained at the same time.

[0022] Further, in the model construction module, the improved ECA attention module specifically adds a log operation before the point multiplication operation.

[0023] Further, in the model construction module, the up-sampling module first recovers the picture size layer by layer through deconvolution to reduce the feature dimension, and the top three layers use element-wise add to fuse the features; the outputs of the two features of the same layer of the down-sampling and the up-sampling are combined to generate a fusion feature set: the information of the down-sampling and the information of the same layer of the up-sampling For each layer of the down-sampling information, different weights are given , and the add fusion formula is as follows:

[0024] .

[0025] Further, in the model construction module, the feature fusion of the spatial pyramid fusion module uses the Adam optimizer to accelerate the model convergence speed, and the dice loss is used as the loss function to evaluate the difference between the segmentation sample and the label sample.

[0026] The technical scheme provided in the embodiment of the application has at least the following technical effects or advantages:

[0027] By using the improved U_Net neural network, the neural network is trained by using the data set of the brain and skull of the fetus to be diagnosed, and through supervised training, a stable segmentation accuracy is obtained.

[0028] The above description is only a summary of the technical scheme of the application, in order to more clearly understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0029] The application will be further described below with reference to the embodiments and the accompanying drawings.

[0030] Figure 1 A flowchart of a brain and skull cavity segmentation method based on a U_Net model according to an embodiment of the application;

[0031] Figure 2 A schematic diagram of the overall structure of the model according to an embodiment of the application;

[0032] Figure 3 A schematic diagram of the residual block structure according to an embodiment of the application;

[0033] Figure 4 A schematic diagram of the attention module structure according to an embodiment of the application;

[0034] Figure 5 A schematic diagram of the improved ECA attention module structure according to an embodiment of the application;

[0035] Figure 6 A schematic diagram of the spatial pyramid fusion module according to the application;

[0036] Figure 7 A schematic diagram of the structure of the device according to the second embodiment of the application. DETAILED DESCRIPTION

[0037] The embodiment of the application provides a brain and skull cavity segmentation method and device based on a U_Net model, which is used for cultivating doctors without fetal brain diagnosis experience to perform maternal magnetic resonance picture auxiliary segmentation or improving the work efficiency of experienced doctors.

[0038] The technical solutions in the embodiments of the present application have the following general ideas:

[0039] By constructing a multi-level down-sampling encoder with a U-shaped structure and a corresponding up-sampling decoder, an efficient attention mechanism is added to improve the segmentation accuracy, and a multi-level feature fusion module is used to fuse features; the optimal segmentation model is trained on the training set; the trained segmentation model is used to segment the fetal intracranial cavity of the test set with unknown segmentation labels. The down-sampling encoder with a residual block avoids the gradient vanishing phenomenon caused by too deep network, and the attention mechanism is introduced in the middle layer of down-sampling and up-sampling to further improve the sensitivity to the target area. Using 3x3, 5x5, and 7x7 convolution kernels to segment the image improves the receptive field of the target area. Compared with the traditional U_Net model, the network is deeper and has stronger feature extraction capability, which can extract the target area from more complex images. Various calculation methods make the model structure more flexible and adaptable to business scenarios, improving the generalization ability of the model. Experiments show that our model can accurately segment the brain region in the fetal MRI image, providing technical support and reference for judging the fetal brain development and calculating the fetal brain volume for subsequent work.

[0040] The embodiments of the present application aim to assist doctors in judging the development of the fetal brain, so the 3D imaging method can better assist doctors to make better judgments, and therefore the magnetic resonance method is used to obtain medical images, a set of transverse, sagittal and coronal photos of the same fetus are obtained at the same time, and the picture format is png picture. The pulse Fourier transform nuclear magnetic resonance instrument is considered to be used on the collection instrument, and the advantage of this kind of instrument is that the single pulse time is short, the imaging is fast, and the time interval of each pulse is only a few seconds. For fetal brain medical pictures, considering the health problem of the fetus, it is not appropriate to use an instrument with too long collection time, and the fetus is easy to displace in the mother, so it is required to shorten the collection time as much as possible to obtain a better effect data set.

[0041] The obtained medical pictures do not have corresponding target region label pictures. Medical data is difficult to obtain, which is also a factor limiting the use of artificial intelligence in medical image processing. Artificial intelligence researchers are proficient in various network structures and hyperparameter settings, but they are not good at basic medical data production. Medical personnel are on the contrary, so mutual cooperation is often needed in data set production. The magnetic resonance image of the fetal brain is complex, containing the brain tissue of the fetus and the organ tissue of the mother. It is necessary to obtain the corresponding medical diagnosis statements of the hospital for the brain region, such as brain structure characteristics, imaging characteristics, and edge texture with other brain structures. Effectively separate the brain, midbrain and cerebellum, and segment the corresponding tissue structure.

[0042] In order to make up for the defects of insufficient data, the medical image pictures are preprocessed by scaling, rotating, enhancing contrast and Gaussian blurring to expand the image data, which can also help to solve the problem of overfitting of deep model. The calculation formula of the processing method is as follows:

[0043] ①The calculation formula of image rotation is as follows:

[0044] (1)

[0045] wherein, is the coordinate of the original image , is the rotation angle, θ is the coordinate of the pixel point after rotation corresponding to the pixel point ;

[0046] ②The calculation formula of image scaling is as follows:

[0047] (2)

[0048] wherein, is the coordinate of the original image , is the horizontal scaling ratio, is the vertical scaling ratio, is the coordinate of the pixel point after scaling corresponding to the pixel point ; ③The calculation formula of enhancing contrast is as follows:

[0049]

[0050] (3) wherein, y is the pixel value of a pixel point of the original image, ct is the contrast value, bt is the brightness value, y * is the pixel value of the pixel point after processing the corresponding pixel value y; for the calculation of the pixel value exceeding 255, take 255, and for the calculation of the pixel value less than 0, take 0;

[0051] ④The calculation formula of Gaussian blurring is as follows:

[0052]

[0053]

[0054] wherein, is the horizontal and vertical coordinates of a pixel point of the original image, a new Gaussian kernel with kernel k (k is generally an odd number) is created, the weight of the middle node of the Gaussian kernel is taken as 1, and the Gaussian distribution of the surrounding nodes of the middle node is calculated according to the above formula. Then, the convolution operation is performed on the picture, and it can be found that the picture becomes blurred after the convolution operation, which effectively prevents overfitting in the training process.

[0055] ​​The medical picture is processed to obtain a label picture of only a target region, including:

[0056] The medical picture is first processed by a canny operator to obtain edges of each tissue in the maternal magnetic resonance picture:

[0057] ① The picture is smoothed by a Gaussian filter

[0058] ② The first-order partial derivative operator is used to find the partial derivatives of the picture gray scale along the horizontal direction Gx and the vertical direction Gy, and the gradient amplitude and direction are calculated according to the following formula.

[0059] The amplitude formula is as follows:

[0060]

[0061] The direction formula is as follows:

[0062]

[0063] ③ The gradient amplitude is subjected to non-maximum suppression, that is, the local gradient maximum is found.

[0064] ④ The double-threshold algorithm is used to detect and link the edges.

[0065] The target edge is outlined by using picture software, and the label picture with the best effect is sent to the hospital for confirmation. The png picture with only the target region is saved as a label picture as supervision data in the training process to improve the segmentation accuracy.

[0066] In order to obtain the segmentation result under different receptive fields, the model uses a spatial pyramid fusioner to combine the outputs of three neural networks, and the convolution kernels of each neural network are 3, 5 and 7 respectively. The obtained feature pictures are stacked and then compressed to 1 dimension by using a dimension compressor to obtain the output segmentation picture.

[0067] As shown in Figure 2 , it is a single-layer neural network flowchart of the embodiment of the application, including:

[0068] Step 1, input the medical picture to be trained;

[0069] Since the medical picture requires accurate segmentation accuracy, the picture enhancement method is not used to expand the data set. In the obtaining of the label picture, the edges of each tissue in the medical picture are obtained by the canny operator, and then the brain region is extracted by using the ITK-SNAP software.

[0070] Step 2, obtain a feature picture by a depth separable convolution layer;

[0071] The feature picture is obtained through a deep separable convolution layer, the feature picture is obtained through a deep separable convolution layer, a deep separable convolution with a 3*3 convolution kernel is used to extract a low-level feature picture with a dimension of 32, and the deep separable convolution can effectively reduce the convolution operation parameter quantity and calculation quantity. In the network, all the convolutions are deep separable convolutions. Meanwhile, the Batch normalization operation and the relu activation function are used after the convolution to avoid the problem of gradient disappearance and give the network the ability of nonlinearity.

[0072] Step 3, the feature picture is transmitted into an attention unit;

[0073] The feature picture is transmitted into an attention unit, and the attention unit is divided into a backbone network and a branch network.

[0074] The backbone network is composed of two residual blocks. The input data of each residual block is added after being processed by two deep separable convolutions. The residual structure can effectively avoid the problem of gradient disappearance when the network depth is deepened.

[0075] The branch network first performs a convolution operation with a step of 2 on the input picture to reduce the picture size to half of the original. Then, the residual operation is performed again, and then the deconvolution is performed to enlarge the picture size by one time, and then two convolution operations are performed, and then the sigmoid activation function is used to obtain the output result of the branch network.

[0076] The branch network result is multiplied by the backbone network, and then added to obtain the attention-added parameters, and then a residual operation is performed to output the attention result.

[0077] Step 4, the discard pooling layer uses a convolution layer with a step of 2 to reduce the picture size;

[0078] The discard pooling layer uses a convolution layer with a step of 2 to reduce the picture size, and the down-sampling process is a process of continuously reducing the picture size and increasing the picture dimension. The advantage of using a pooling layer to reduce the picture size is that the calculation is simple, and the disadvantage is that a large amount of non-recoverable picture details are lost. The advantage of using a convolution layer to reduce the picture size is that the local information is fully considered while reducing the size.

[0079] Step 5, repeat steps 3 and 4, the feature picture size becomes 1 / 32 of the original, and the dimension is 1024;

[0080] Step 6, the 1024-dimensional feature picture is sent into an improved ECA attention module;

[0081] Step 7, the deconvolution picture is processed twice, and the picture becomes 1 / 8 of the original picture, and the dimension is 512;

[0082] Upsampling operation increases the feature picture pixels while reducing the dimension. Each time the length and width are doubled, and the dimension is halved.

[0083] Step 8, the picture is deconvolved and added to the feature picture of the same size and dimension in the downsampling process;

[0084] Since the information carried by the upper three layers of downsampling is relatively rich, the upsampling is restored to the same feature picture size and dimension after the deconvolution operation in the same layer, and then added to the parameters of the same layer in the downsampling process, so that the loss caused by the picture shrinking in the downsampling process can be recovered to a certain extent. Since the information carrying capacity of different layers is different, different weights k are given to different layers for fusion operation. The information in the top layer has redundant elements, so the weight can be smaller, and the lower the layer, the higher the weight, and more information needs to be retained.

[0085] Step 9, repeat step 8 three times, restore the feature picture to the original picture size, and the dimension is 32, output the 32-dimensional feature picture to the pyramid structure for fusion; this process is similar to the horizontal link of Unet, and the 32-dimensional picture is output to the pyramid fusion device to synthesize the final picture.

[0086] The input image increases from 3 dimensions to 32 dimensions after the first convolution, and the picture length and width are reduced by 1 / 2 after the pooling operation. The dimension of each layer is 64, 128, 256, and 512 respectively. Before entering the improved ECA attention module, the picture is reduced from 256x256 to 8x8, and the dimension is increased from 3 to 512. After deconvolution, the length and width are doubled, and the dimension is reduced by 1 / 2. In order to recover the precision loss caused by the pooling layer, the top three layers are added with a jump link operation, and the top three layer residual block results are added to the feature picture after deconvolution.

[0087] Please refer to Figure 3 , in order to solve the problem that the deep learning network is easy to fall into gradient disappearance in the training process after deepening the picture depth, the invention uses a residual structure in the downsampling process.

[0088] Due to the influence of the chain rule, when one unit gradient in the chain is 0, the entire chain gradient is 0, and the parameter cannot be updated in the back propagation, as shown in the following formula:

[0089]

[0090] A residual structure is added to the network, and the formula is as follows:

[0091]

[0092] When updating the parameters in the back propagation, the formula is as follows:

[0093]

[0094] It is found that the gradient of the high layer can be passed to the bottom layer due to the presence of 1 in the derivation process, effectively preventing the case of gradient disappearance. And when updating the gradient, some multiplication calculations are changed into addition, improving the calculation efficiency.

[0095] Please refer to Figure 4 In order to make the network pay more attention to the target area, an attention module is added to divide the backbone network and the branch network. The backbone network mainly extracts deep features, and the attention is more reflected in the branch network. After downsampling the incoming feature picture, the residual module is added, and then the size is restored before upsampling. After 2 3x3 convolutions, multiply the output of the backbone network first and then add it. The formula is as follows:

[0096]

[0097] Among them, B(x) is the backbone network, and b(x) is the branch network.

[0098] Please refer to Figure 5 ECA is the abbreviation of Efficient Channel Attention, which is a kind of efficient channel attention module. After global average pooling of the incoming feature picture, the result elements are tiled and expanded. Convolution operation is performed on every K (5 in this invention) elements, and sliding is performed until all elements are traversed. After sigmoid operation, a C-dimensional list is obtained, each dimension has only one element. Then multiply the list with the initial feature picture of the same C dimension to obtain the feature picture with attention weight added.

[0099] In this invention, a log operation is added after the attention module. The reason is that the numerical value of the parameter is small at this stage, and the log is sensitive to the low value. After adding the log operation, it is more beneficial to layer the attention.

[0100] Please refer to Figure 6 By increasing the receptive field of the model through the pyramid structure, the segmentation of the target area under different receptive fields is obtained and stacked and combined. After dimension compression, the brain area is more highlighted in training, and better segmentation results are obtained.

[0101] Embodiment one

[0102] The embodiment provides a brain intracranial cavity segmentation method based on a U_Net model, as shown in Figure 1 , comprising:

[0103] Step 10, acquiring the magnetic resonance image of the fetus in the mother and making a label picture according to the medical diagnosis sentence, and then pre-processing to obtain a training data set;

[0104] Step 20, construct a brain magnetic resonance image lumen segmentation model, the model comprises a down-sampling module, an improved ECA attention module, an up-sampling module and a spatial pyramid fusion module connected in sequence; wherein the down-sampling module first uses a depth separable convolution as a model convolution layer and a pooling layer, for reducing the picture size and increasing the picture dimension, and obtaining a specified dimension feature picture through multiple convolutions; the improved ECA attention module adds a log operation with high sensitivity at low values to facilitate hierarchical attention; the up-sampling module recovers the picture size and reduces the feature dimension layer by layer through deconvolution, and adds the same dimension and size feature pictures in the down-sampling process; the spatial pyramid fusion module is used for feature fusion of the pictures obtained by the up-sampling part;

[0105] Step 30, input the training data set into the model for training to obtain a trained brain magnetic resonance image lumen segmentation model;

[0106] Step 40, obtain the magnetic resonance image of the fetus to be diagnosed and input the trained brain magnetic resonance image lumen segmentation model to obtain a brain intracranial lumen region segmentation map.

[0107] The deep neural network adopts an encoder-decoder framework based on U_Net. The down-sampling part of the encoder first uses a depth separable convolution (Depthwise separable convolution) as a model convolution layer and a pooling layer. Compared with the conventional convolution operation, the number of parameters and the operation cost of this convolution are lower, and in the case of the same computing power of GPU, the number of layers is effectively deepened. In order to prevent the problem of gradient disappearance caused by too many layers, a residual block (residual block) is used to alleviate the problem. An improved ECA attention module is added between the encoder and the decoder to improve the sensitivity to the target area.

[0108] In one possible implementation, in the step 10, the magnetic resonance image of the fetus in the mother is obtained by magnetic resonance, and a group of transverse, sagittal and coronal pictures of the same fetus are obtained at the same time.

[0109] In one possible implementation, in the step 20, the improved ECA attention module specifically adds a log operation before the point multiplication operation.

[0110] In one possible implementation, in the step 20, the up-sampling module first recovers the picture size and reduces the feature dimension layer by layer through deconvolution, and the upper three layers use element-wise add to fuse the features; the outputs of the two features of the same layer of the down-sampling and the up-sampling are combined to generate a fusion feature set: the information of the down-sampling and the same layer information of the up-sampling Different weights are given to the down-sampled information of each layer The add fusion formula is as follows:

[0111] .

[0112] The information of the encoder and the same-layer information of the decoder are added, and the dimension of the feature picture itself is not changed, but the amount of information in each dimension is increased, which is obviously beneficial to picture segmentation and avoids the loss caused by information loss in picture down-sampling. Different weights are given to each layer, and the information in the upper layer is more abundant than that in the lower layer, so it also contains a lot of redundant information, and therefore different weights are given.

[0113] In a possible implementation, in the step 20, the feature fusion of the spatial pyramid fusion module uses an Adam optimizer to accelerate the convergence speed of the model, and a dice loss is used as a loss function to evaluate the difference between the segmented sample and the label sample.

[0114] Based on the same inventive concept, the application also provides a device corresponding to the method in Embodiment 1, which is described in detail in Embodiment 2.

[0115] Embodiment 2

[0116] In this embodiment, a brain intracranial cavity segmentation device based on a U_Net model is provided, as shown in Figure 7 which includes a data set processing module, a model construction module, a model training module, and a segmentation module.

[0117] The data set processing module is configured to obtain magnetic resonance images of a fetus in a mother and make label pictures according to medical diagnosis sentences, and then perform preprocessing to obtain a training data set.

[0118] The model construction module is configured to construct a brain intracranial cavity segmentation model of a magnetic resonance image, and the model includes a down-sampling module, an improved ECA attention module, an up-sampling module, and a spatial pyramid fusion module connected in sequence; wherein the down-sampling module first uses a depth separable convolution as a model convolution layer and a pooling layer, for reducing the picture size and increasing the picture dimension, and obtaining a feature picture of a specified dimension through multiple convolutions; the improved ECA attention module adds a log operation with high sensitivity at low values to facilitate hierarchical attention; the up-sampling module reduces the picture size and the feature dimension through deconvolution layer by layer and adds the feature pictures of the same dimension and size in the down-sampling process; and the spatial pyramid fusion module is configured to perform feature fusion on the pictures obtained after the up-sampling process is completed.

[0119] The model training module is configured to input a training data set into the model for training to obtain a trained brain magnetic resonance image inner cavity segmentation model.

[0120] The segmentation module is configured to obtain a magnetic resonance image of a fetus to be diagnosed and input the trained brain magnetic resonance image inner cavity segmentation model to obtain a brain inner cavity region segmentation map.

[0121] In a possible implementation, in the data set processing module, the magnetic resonance image of the fetus in the mother is obtained by using magnetic resonance, and a group of transverse, sagittal and coronal images of the same fetus are obtained at the same time.

[0122] In a possible implementation, in the model construction module, the improved ECA attention module specifically adds a log operation before point multiplication operation.

[0123] In a possible implementation, in the model construction module, the up-sampling module first restores the picture size and reduces the feature dimension by deconvolution layer by layer, and the upper three layers use element-wise add to fuse the features; the output of the two kinds of features of the same layer of the down-sampling and the up-sampling is combined to generate a fused feature set: the information of the down-sampling and the information of the same layer of the up-sampling For each layer of the down-sampling information, different weights are given , and the add fusion formula is as follows:

[0124] .

[0125] In a possible implementation, in the model construction module, the feature fusion of the spatial pyramid fusion module uses the Adam optimizer to accelerate the model convergence speed, and the dice loss is used as a loss function to evaluate the difference between the segmentation sample and the label sample.

[0126] Since the device introduced in the second embodiment of the present application is the device used to implement the method of the first embodiment of the present application, the specific structure and deformation of the device can be understood by those skilled in the art based on the method introduced in the first embodiment of the present application, and therefore will not be described here. Any device used in the method of the first embodiment of the present application belongs to the scope of the present application.

[0127] The present application trains the neural network by using the data set of the brain of the fetus to be diagnosed through the improved U_Net neural network, and obtains stable segmentation accuracy through supervised training. The diagnostic accuracy of doctors without experience in fetal brain development diagnosis is improved. For experienced doctors, it also has reference significance and can improve the clinical diagnosis efficiency.

[0128] Those skilled in the art will appreciate that embodiments of the present application can be readily used as software, hardware, or a combination of software and hardware. In a software embodiment, the methods can be tangibly embodied in a machine-readable storage medium having stored thereon instructions that can be used to program a computer to perform any of the methods. The software implementation can be initialized by loading and executing a set of instructions arranged to perform one of the methods into the computer's memory. Alternatively, hard-wired circuitry can be used in place of, or in combination with, software instructions. Thus, the

[0129] The present application is described in reference to the 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 block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can 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 illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for performing each of the functions specified in one or more of the flowchart illustrations and / or block diagrams.

[0130] These computer program instructions can 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 instructions which implement the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for performing each of the functions specified in one or more of the flowchart illustrations and / or block diagrams.

[0131] These computer program instructions can 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 flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for performing each of the functions specified in one or more of the flowchart illustrations and / or block diagrams.

[0132] While the present application has been described with reference to specific implementations thereof, it should be understood by those skilled in the art that a variety of equivalents for the specific embodiments described and illustrated herein can be made without departing from the spirit of the present application, which is limited only by the scope of the appended claims.

Claims

1. A method for intracranial cavity segmentation based on a U_Net model, characterized in that, include: Step 10: Obtain MRI images of the fetus in the mother's body and create labeled images based on medical diagnostic statements, then perform preprocessing to obtain a training dataset; Step 20: Construct a brain MRI intracavitary segmentation model. The model includes a downsampling module, an improved ECA attention module, an upsampling module, and a spatial pyramid fusion module connected in sequence. The downsampling module first uses depthwise separable convolutions as convolutional and pooling layers to reduce image size and increase image dimensionality, obtaining feature images of a specified dimension through multiple convolutions. The improved ECA attention module adds a log operation before the dot product operation. This improved ECA attention module facilitates layered attention by adding a log operation, which has high sensitivity at low values. The upsampling module restores the image size layer by layer through deconvolution, reducing the feature dimension and adding it to the feature images of the same dimension and size from the downsampling process. The spatial pyramid fusion module is used to perform feature fusion on the image obtained after the upsampling part has been processed. Step 30: Input the training dataset into the model for training to obtain a trained intracranial magnetic resonance imaging cavity segmentation model. Step 40: Obtain the magnetic resonance image of the fetus to be diagnosed and input it into the trained intracranial magnetic resonance image cavity segmentation model to obtain the intracranial cavity region segmentation map.

2. The method of claim 1, wherein: In step 10, the magnetic resonance images of the fetus in the mother's body are acquired by magnetic resonance imaging, and a set of transverse, sagittal and coronal images are obtained simultaneously for the same fetus.

3. The method of claim 1, wherein: In step 20, the up-sampling module first restores the picture size and reduces the feature dimension layer by layer through deconvolution, and element-wise add is used to fuse the features of the top three layers. The outputs of the two features of the same layer of the down-sampling and up-sampling are combined to generate a fused feature set: the information of the down-sampling and the information of the same layer of the up-sampling For each layer of the down-sampling information, different weights are given The add fusion formula is as follows: 。 4. The method of claim 1, wherein: In step 20, the feature fusion of the spatial pyramid fusion module uses the Adam optimizer to accelerate the model convergence speed, and the dice loss is used as a loss function to evaluate the difference between the segmented samples and the labeled samples. 5.A brain intracranial cavity segmentation device based on a U_Net model, characterized in that, include: The system includes a dataset processing module, a model building module, a model training module, and a segmentation module. The dataset processing module is used to acquire magnetic resonance images of the fetus in the mother's body and create labeled images based on medical diagnostic statements, and then perform preprocessing to obtain a training data set; The model construction module is used to construct a segmentation model for intracavitary cavities in cranial magnetic resonance imaging (MRI). The model includes a downsampling module, an improved ECA attention module, an upsampling module, and a spatial pyramid fusion module connected in sequence. The downsampling module first uses depthwise separable convolutions as convolutional and pooling layers to reduce image size and increase image dimensionality, obtaining feature images of a specified dimension through multiple convolutions. The improved ECA attention module adds a log operation before the dot product operation. This log operation, which is highly sensitive at low values, facilitates layered attention. The upsampling module restores the image size layer by layer through deconvolution to reduce feature dimensions and adds these features to the same dimension and size feature images obtained during downsampling. The spatial pyramid fusion module performs feature fusion on the image obtained after the upsampling process. The model training module is used to input the training data set into the model for training, and obtain a trained intracranial magnetic resonance imaging cavity segmentation model. The segmentation module is used to acquire the magnetic resonance imaging of the fetus to be diagnosed and input it into a trained intracranial magnetic resonance imaging cavity segmentation model to obtain a segmentation map of the intracranial cavity region.

6. The apparatus of claim 5, wherein: In the dataset processing module, the magnetic resonance images of the fetus in the mother's body are acquired by magnetic resonance imaging, and a set of transverse, sagittal and coronal images are obtained simultaneously for the same fetus.

7. The apparatus according to claim 5, characterized in that: In the model construction module, the up-sampling module first restores the picture size and reduces the feature dimension layer by layer through deconvolution, and the top three layers use element-wise add to fuse the features; the outputs of the two features of the same layer of the down-sampling and the up-sampling are combined to generate a fused feature set: the information of the down-sampling and the information of the same layer of the up-sampling For the information of each layer of the down-sampling, different weights are given The add fusion formula is as follows: 。 8. The apparatus of claim 5, wherein: In the model building module, the feature fusion of the spatial pyramid fusion module uses the Adam optimizer to accelerate the model convergence speed, and the dice loss is used as a loss function to evaluate the difference between segmented samples and labeled samples.

Images