A method, system, device and medium for fetal MRI brain tissue segmentation

Abstract

The present invention discloses a method, system, device and medium for fetal MRI brain tissue segmentation, belonging to the segmentation of brain tissue images in the field of artificial intelligence technology. The purpose is to solve the problems in the prior art that the segmentation method based on deep learning cannot be applied to multi-resolution data, and the manual measurement of the width of the lateral ventricle and single-shot measurement result in low measurement efficiency. It includes obtaining sample data and labels, constructing and training a lateral ventricle segmentation and classification model, and real-time segmentation and classification. The lateral ventricle segmentation and classification model includes a U-Net segmentation module and a Resnet classification module. The U-Net segmentation module includes a DWMF block and a decoder block, and the DWMF block and the decoder block are connected through a skip connection block; the DWMF block includes two depthwise separable convolution blocks and four depthwise separable channel dilated convolution blocks, and the output of one depthwise separable convolution block is used as the input of the last depthwise separable channel dilated convolution block after passing through three depthwise separable channel dilated convolution blocks.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, relates to the segmentation of brain tissue images, and particularly relates to a method, a system, a device and a medium for segmenting fetal MRI brain tissue. Background Art

[0002] Prenatal examination is crucial for early detection and treatment of fetal congenital diseases, and it is particularly important to measure and analyze key biological indicators of the developing fetal brain. Ultrasonography is a commonly used method for prenatal diagnosis, but due to factors such as cranial sound attenuation, ultrasonography has limitations in judging and analyzing the development of the fetal brain. In recent years, magnetic resonance imaging (MRI) has become increasingly important in evaluating fetal brain development and growth. It has the characteristics of superior image resolution and tissue contrast and plays a crucial role in solving difficult diagnoses. Analyzing important biological indicators based on MRI images and determining the degree to which a case deviates from the normal brain development trajectory can be an important means of measuring fetal brain abnormalities. Among them, the lateral ventricles in fetal brain tissue are parts that need to be focused on. Many congenital diseases can cause changes in the morphology of the lateral ventricles compared to the normal situation, and lateral ventricle dilation is typical among them. Therefore, examining fetal brain tissue based on MRI images and analyzing whether there is lateral ventricle dilation is an important method for prenatal examination of pregnant women.

[0003] In the prior art, when performing an MRI examination on a fetus, a pregnant woman will, under the guidance of a professional physician in the radiology department, undergo a full - scale three - dimensional scan of the fetus in the uterus by a magnetic resonance machine in the coronal plane, sagittal plane, and axial plane directions according to a certain scanning sequence. After the scanning is completed, the doctor reads and analyzes each piece of MRI image data obtained under this sequence, manually measures the width of the lateral ventricles on the imaging display device, analyzes its morphological characteristics graph by graph, and determines whether there is an abnormal situation of lateral ventricle dilation. Due to the huge workload, long time consumption, and low efficiency of manual image reading and diagnosis, in recent years, people have tried many automated and semi - automated methods to detect and segment brain tissue. Deep learning methods, especially convolutional neural networks (CNNs), have been popular in some object recognition and biomedical image segmentation challenges. They directly learn complex and representative features automatically from the data itself and achieve better performance than traditional machine learning methods. Segmenting the lateral ventricle organs of fetal brain tissue based on a convolutional neural network and measuring the width of the lateral ventricles is a new method developed in the context of deep learning in recent years.

[0004] The invention patent with the application number 202010337402.6 discloses a deep learning-based MRI whole brain tissue segmentation system, including: an image data preprocessing module for preprocessing MR brain images; an image block division module for dividing the preprocessed images into blocks; a multi-prior feature network model construction module for constructing a multi-prior feature network model based on symmetric prior, edge prior, and neighborhood prior feature information; a training module for training the multi-prior feature network model using a training set; a label fusion module for determining the final segmentation result of each pixel in the MR brain image; and an inverse affine transformation module for performing a transformation operation on the MR brain image from the MNI space to the original image space after determining the final segmentation result of each pixel in the MR brain image. By introducing multiple prior features, the present invention makes the segmentation result more accurate, the edge between brain tissues clearer, and the details more abundant.

[0005] The invention patent with the application number 202110902377.6 discloses a method for segmenting infant brain tissue images based on U-net and attention mechanism. The method includes the following steps: randomly selecting a certain number of samples from the training samples according to the pixel size of 32×32×32 to form a training set; building a Dense-Unet network model; using the Dense-Unet network model to train and adjust the parameters of the training set to obtain the best segmentation model and save it; comparing the model prediction result with the result of the manual label to obtain a distribution map of the areas that are easily misclassified; using a Gaussian function to blur the distribution map of the areas that are easily misclassified for each sample and calculate the average value; forming a training set from the training samples according to a certain ratio; designing a self-attention mechanism including space and areas that are easily misclassified and adding it to the network model; using the new network model to train and adjust the parameters of the training set to obtain the best segmentation model and save it.

[0006] Similar to the above-mentioned invention patents, most of the existing deep learning-based fetal brain tissue and organ segmentation methods are aimed at datasets with unified resolution under regular imaging quality. However, in practice, the MRI devices in different medical centers are different, the habits of device operators are different, the conditions and situations for examining pregnant women are different, and the imaging accuracy is different. Therefore, there are differences in spatial resolution between the imaging sequences of actual fetal brain MRI images. An MRI dataset containing multiple spatial resolutions is called a multi-resolution dataset, and the existing segmentation techniques are difficult to automatically adapt to multi-resolution data. Secondly, in the existing methods, doctors need to manually measure based on the width of the lateral ventricle to determine whether the fetus has lateral ventricle dilation, and the measurement method is limited to a single MRI image and does not effectively utilize voxel information, resulting in low measurement efficiency. Summary of the Invention

[0007] The object of the present invention is to provide a fetal MRI brain tissue segmentation method, system, device and medium to solve the problems in the prior art that the segmentation method based on deep learning cannot be applied to multi-resolution data, and the manual measurement and single-image measurement of the width of the lateral ventricle result in low measurement efficiency.

[0008] The present invention specifically adopts the following technical solutions to achieve the above object:

[0009] A fetal MRI brain tissue segmentation method includes the following steps:

[0010] Step S1, obtaining sample data and labels;

[0011] Obtain fetal brain MRI image data and label data, where the label data includes the lateral ventricle segmentation label and the classification label of whether there is lateral ventricle dilation for each fetal brain MRI image;

[0012] Step S2, constructing a lateral ventricle segmentation and classification model;

[0013] Construct a lateral ventricle segmentation and classification model, which includes a U-Net segmentation module and a Resnet classification module, and the output of the U-Net segmentation module is used as the input of the Resnet classification module;

[0014] The U-Net segmentation module uses a U-Net network, which includes multiple DWMF blocks for downsampling and multiple decoder blocks for upsampling. The DWMF blocks and the decoder blocks are connected through skip connection blocks; each DWMF block includes two depthwise separable convolution blocks and four depthwise separable channel dilated convolution blocks. The image data is used as the input of the two depthwise separable convolution blocks. The output of one depthwise separable convolution block is respectively input into three depthwise separable channel dilated convolution blocks, and the outputs of these three depthwise separable channel dilated convolution blocks are jointly used as the input of the last depthwise separable channel dilated convolution block. The output of the last depthwise separable channel dilated convolution block is fused with the output of the other depthwise separable convolution block and used as the output of the DWMF block;

[0015] Step S3, training the lateral ventricle segmentation and classification model;

[0016] Use the sample data and labels obtained in step S1 to train the lateral ventricle segmentation and classification model constructed in step S2 to obtain a mature lateral ventricle segmentation and classification model;

[0017] Step S4, real-time segmentation and classification;

[0018] Real-time obtain the fetal brain MRI image to be measured and input it into the lateral ventricle segmentation and classification model, and the lateral ventricle segmentation and classification model outputs the segmentation result and the classification result.

[0019] Further, in step S1, after obtaining the fetal brain MRI image data, the voxel information of each fetal brain MRI image is statistically analyzed, and the statistically analyzed voxel information is preprocessed; according to the results of the data preprocessing, the parameters of the lateral ventricle segmentation and classification model are configured.

[0020] Furthermore, when preprocessing the statistically analyzed voxel information, the specific preprocessing method is as follows:

[0021] Step S1-1-1, determining the target voxels for data preprocessing based on the statistically analyzed voxel information;

[0022] Step S1-1-2, performing preprocessing based on the target voxels;

[0023] Based on the difference between the original voxel information and the target voxels, calculate the new shape of the fetal brain MRI image after scaling the original voxels to the target voxels, and scale the fetal brain MRI image to the size of the new shape by trilinear interpolation to obtain the target image shape size of the trilinear interpolation;

[0024] The calculation formula for the target image shape size of the trilinear interpolation is:

[0025]

[0026] where H, W, and D respectively represent the width, height, and slice spacing size of the slices in the MRI sequence, represents the height of each voxel unit in the original voxels, reflecting the physical distance in the real world corresponding to each pixel point in the slice height dimension (H), represents the height of each voxel unit in the target voxels, represents the width of each voxel unit in the original voxels, reflecting the physical distance in the real world corresponding to each pixel point in the slice width dimension (W) the width of each voxel unit in the target voxels, represents the depth of each voxel unit in the original voxels, reflecting the physical distance in the real world corresponding to each pixel point in the slice spacing dimension (D), the depth of each voxel unit in the target voxels.

[0027] Furthermore, in step S1-1-1, the method for determining the target voxels is as follows:

[0028] Step S1-1-1-1, determining the median of the voxel information of all samples according to the statistically analyzed voxel information;

[0029] Step S1-1-1-2: Based on the median of the voxel information statistics, determine whether the voxel depth representing the slice spacing (D) of each slice in the MRI sequence exceeds twice the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W) of a single slice, and judge whether there is anisotropy in this data set;

[0030] If there is anisotropy, go to step S1-1-1-3; if there is no anisotropy, go to step S1-1-1-4;

[0031] Step S1-1-1-3: Statistically list the voxel depths representing the slice spacing (D) of each slice in the MRI sequence of all samples, sort the list from small to large, and take the mode of the top 10% values as the target depth value corresponding to the target voxel to reduce distortion during the interpolation process; for the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W) of a single slice, take the median of all samples statistically, and recombine it with the target depth value into the target voxel. Among them, the height and width of the target voxel are the medians of the voxel widths and heights of all sample original images, and the depth of the target voxel is the mode of the top 10% values after statistically all samples;

[0032] Step S1-1-1-4: Statistically list the voxel height corresponding to the height (H) of a single slice in the MRI sequence, the voxel width corresponding to the width (W) of a single slice, and the voxel depth representing the slice spacing (D) of each slice in the original images of all samples, take the median of each statistical list, and recombine it into the target voxel. Further, in step S1, when configuring the parameters of the lateral ventricle segmentation and classification model, the specific method is as follows:

[0033] Step S1-2-1: Set the minimum feature map size for the U-Net segmentation module of the lateral ventricle segmentation and classification model;

[0034] Step S1-2-2: Calculate the mode of the sizes of all images of the fetal brain MRI image data for each axis, and record it as the mode size after recombination; then calculate the required downsampling times for each axis according to the mode size and the minimum feature map size;

[0035] Step S1-2-3: Determine the input size of the lateral ventricle segmentation and classification model according to the downsampling times;

[0036] Step S1-2-4: Calculate the configuration of the convolutional layer and the max pooling layer during the downsampling process according to the required downsampling times for each axis, the target voxel, the input size, and the minimum feature map size.

[0037] Furthermore, in step S1-2-4, when calculating the configuration of the convolutional layer and the max pooling layer, the specific steps are as follows:

[0038] Step S1-2-4-1: Establish a loop with the maximum downsampling times. The starting value of the loop is 0, and the ending value is the downsampling times. The loop index is , and the loop index is ;

[0039] Step S1-2-4-2: Calculate the voxel differences between axes in the voxel, and check whether there is an axis whose voxel value is twice that of other axes;

[0040] Step S1-2-4-3: Check the difference between the current feature map size and the minimum feature map size, and check whether there is an axis whose feature map size is less than twice the minimum feature map size;

[0041] Step S1-2-4-4: If there is an axis that meets the voxel requirements, adjust the kernel size of this axis to 1 in the convolutional layer and the max pooling layer; if it does not meet the voxel requirements but meets the feature map requirements, adjust the kernel size of this axis to 1 in the max pooling layer, and the convolutional layer uses the standard convolutional layer; if neither the voxel requirements nor the feature map requirements are met, each axis is processed according to the standard convolutional layer and pooling layer, that is, the convolutional kernel size of the convolutional layer is (3,3,3), and the convolutional kernel size of the pooling layer is (2,2,2);

[0042] Step S1-2-4-5: Multiply the elements in the max pooling layer convolutional kernel by the elements of the voxels in the i-th layer according to the index to obtain the voxels in the (i + 1)-th layer, and divide the feature map size of the i-th layer by the elements of the max pooling layer convolutional kernel according to the index to obtain the feature map size of the (i + 1)-th layer; repeat the calculation of the configurations of the convolutional layer and the max pooling layer in the (i + 1)-th layer according to steps S1-2-4-2, S1-2-4-3, and S1-2-4-4.

[0043] Furthermore, in step S2, the Resnet classification module adopts the Resnet network, including a convolutional layer, a max pooling layer, four Resnet residual blocks, a global average pooling layer, and a fully connected layer set in sequence.

[0044] A fetal MRI brain tissue segmentation system includes:

[0045] A sample data and label acquisition module, which is used to acquire fetal brain MRI image data and label data. The label data includes the lateral ventricle segmentation label of each fetal brain MRI image and the classification label of whether there is lateral ventricle dilation;

[0046] A lateral ventricle segmentation and classification model construction module, which is used to construct a lateral ventricle segmentation and classification model. The lateral ventricle segmentation and classification model includes a U-Net segmentation module and a Resnet classification module. The output of the U-Net segmentation module serves as the input of the Resnet classification module;

[0047] The U-Net segmentation module adopts the U-Net network, including multiple DWMF blocks for downsampling and multiple decoder blocks for upsampling. The DWMF blocks and the decoder blocks are connected through skip connection blocks; each DWMF block includes two depthwise separable convolution blocks and four depthwise separable channel dilated convolution blocks. The image data is used as the input of the two depthwise separable convolution blocks. The output of one of the depthwise separable convolution blocks is respectively input into three depthwise separable channel dilated convolution blocks. The outputs of these three depthwise separable channel dilated convolution blocks are jointly used as the input of the last depthwise separable channel dilated convolution block. The output of the last depthwise separable channel dilated convolution block is fused with the output of the other depthwise separable convolution block and used as the output of this DWMF block;

[0048] The lateral ventricle segmentation and classification model training module is used to train the lateral ventricle segmentation and classification model constructed by the lateral ventricle segmentation and classification model construction module with the sample data and labels obtained by the sample data and label acquisition module to obtain a mature lateral ventricle segmentation and classification model;

[0049] The real-time segmentation and classification module is used to obtain the fetal brain MRI image to be measured in real time and input it into the lateral ventricle segmentation and classification model. The lateral ventricle segmentation and classification model outputs the segmentation result and the classification result.

[0050] A computer device includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the above method.

[0051] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the steps of the above method.

[0052] The beneficial effects of the present invention are as follows:

[0053] 1. The self-configuring fetal MRI brain tissue segmentation and classification method based on deep learning proposed by the present invention can quickly segment the brain tissue in the fetal MRI image in a short time, judge whether there is a problem of lateral ventricle dilation, and at the same time has good robustness for multi-resolution data, reducing a large amount of repetitive work of doctors, and effectively solving the problems that the segmentation method based on deep learning cannot be applied to multi-resolution data and the manual measurement of the lateral ventricle width and single-shot measurement result in low measurement efficiency.

[0054] 2. After specifying a small number of hyperparameters in the preprocessing stage, the present invention can automatically complete the network configuration, greatly shortening the training and learning cycle of the model, and improving the analysis and judgment efficiency of the brain tissue and whether there is a problem of lateral ventricle dilation; after the model is trained, only the brain MRI image of this patient needs to be input, and the segmentation of the fetal brain tissue and the diagnosis of whether there is a lateral ventricle dilation can be automatically performed. Brief Description of the Drawings

[0055] Figure 1 is a schematic flow chart of the present invention;

[0056] Figure 2 is a schematic structural diagram of the U-Net network in the present invention;

[0057] Figure 3 is a schematic structural diagram of the DWMF block in the present invention;

[0058] Figure 4 is a schematic structural diagram of the Resnet network in the present invention;

[0059] Figure 5 is a voxel distribution diagram of the sample data in the present invention;

[0060] Figure 6 is a schematic flow chart of obtaining the target voxel in the present invention;

[0061] Figure 7 is a schematic flow chart of preprocessing the target voxel in the present invention;

[0062] Figure 8 is a schematic flow chart of calculating the input size in the present invention;

[0063] Figure 9 is a schematic flow chart of calculating the configuration of each layer of the network in the present invention. Detailed Embodiments

[0064] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention.

[0065] Therefore, based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0066] Embodiment 1

[0067] Due to problems such as multi - resolution (i.e., inconsistent voxel data sizes) and anisotropy (i.e., large differences in voxel axes) in fetal MRI brain tissue image data, in this embodiment, taking the multi - resolution coronal plane data as an example, a method for segmenting fetal MRI brain tissue is provided. First, the voxel data is statistically analyzed to obtain the distribution information of voxel sizes; then, based on the distribution information of voxel sizes, the sample data is pre - processed to unify it into a single voxel distribution and perform pixel normalization; then, an automatic configuration network is set up for the pre - processed data and pre - set feature icon annotations, and the network is trained; finally, the trained network is used to process fetal MRI brain tissue images to obtain the segmentation result and classification result of the fetal brain tissue. As Figure 1 shown, the specific steps are as follows:

[0068] Step S1, obtain sample data and labels;

[0069] Obtain fetal brain MRI image data and label data. The label data includes the lateral ventricle segmentation label and the classification label indicating whether there is lateral ventricle dilation for each fetal brain MRI image.

[0070] The fetal brain MRI image data comes from 186 cases of fetal brain coronal plane MRI images provided by West China Second University Hospital / West China Women's and Children's Hospital of Sichuan University, which includes samples with lateral ventricle dilation and samples with normal lateral ventricles. In addition, for each fetal brain MRI image, there is a lateral ventricle segmentation label and a classification label indicating whether there is lateral ventricle dilation, and these labels are also annotated by physicians in the Department of Radiology of West China Second University Hospital / West China Women's and Children's Hospital of Sichuan University.

[0071] After obtaining the fetal brain MRI image data, the voxel information of each fetal brain MRI image is statistically analyzed, and the statistically analyzed voxel information is pre - processed; according to the results of data pre - processing, the parameters of the lateral ventricle segmentation and classification model are configured. Specifically:

[0072] Traverse all sample data, read the voxel information and three - dimensional shape information of the MRI images, and perform statistics. The voxel size distribution of the obtained dataset is as Figure 5 shown.

[0073] After obtaining the voxel information of the sample data, pre - processing is performed to unify it into a single voxel distribution and perform pixel normalization. The specific pre - processing method is:

[0074] Step S1 - 1 - 1, determine the target voxel for data pre - processing based on the statistically analyzed voxel information. When determining the target voxel, as Figure 6 shown, the specific steps are:

[0075] Step S1 - 1 - 1 - 1, according to the statistically analyzed voxel information, determine the median of the voxel information of all samples.

[0076] Step S1-1-1-2: Based on the median of the statistically obtained voxel information, determine whether the voxel depth representing the slice spacing (D) of each slice in the MRI sequence exceeds twice the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W) of a single slice in the MRI sequence, and judge whether there is anisotropy in this data set.

[0077] If there is anisotropy, go to step S1-1-1-3; if there is no anisotropy, go to step S1-1-1-4.

[0078] Step S1-1-1-3: Statistically obtain the voxel depths representing the slice spacing (D) of each slice in the MRI sequence of all samples as a list, sort the list from small to large, and take the mode of the top 10% values as the target depth value corresponding to the target voxel to reduce distortion during the interpolation process; for the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W) of a single slice in the MRI sequence, take the median of all statistically obtained samples, and recombine it with the target depth value into the target voxel. Among them, the height and width of the target voxel are the medians of the voxel widths and heights of the original images of all samples, and the depth of the target voxel is the mode of the top 10% values after statistically obtaining all samples.

[0079] Step S1-1-1-4: Statistically obtain the voxel height corresponding to the height (H) of a single slice in the MRI sequence, the voxel width corresponding to the width (W) of a single slice in the MRI sequence, and the voxel depth representing the slice spacing (D) of each slice in the original images of all samples as a list, take the median of each statistical list, and recombine it into the target voxel.

[0080] According to Figure 5 the voxel distribution shown, the median of the voxels in this embodiment is: (0.9375, 0.9375, 4.19998932), indicating that the axis with the index D representing the slice spacing of each slice in the MRI sequence has a large difference from the other two axes. After statistically obtaining the values of the voxel depths representing the slice spacing (D) of each slice in all samples, select 3.9999969 as the target value, and recombine it with the median of the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W) of a single slice in the MRI sequence into (0.9375, 0.9375, 3.9999969) as the target voxel.

[0081] Step S1-1-2: Perform preprocessing based on the target voxel.

[0082] As Figure 7 shown, based on the difference between the original voxel information and the target voxel, calculate the new shape of the shape of the fetal brain MRI image after scaling the original voxel to the target voxel, and scale the fetal brain MRI image to the size of the new shape by trilinear interpolation to obtain the target image shape size of the trilinear interpolation.

[0083] The calculation formula for the shape and size of the target image of trilinear interpolation is as follows:

[0084]

[0085] Among them, H, W, and D respectively represent the width, height, and slice spacing size of the slices in the MRI sequence. represents the height of each voxel unit in the original voxel, reflecting the physical distance in the real world corresponding to each pixel point in the slice height dimension (H). represents the height of each voxel unit in the target voxel. represents the width of each voxel unit in the original voxel, reflecting the physical distance in the real world corresponding to each pixel point in the slice width dimension (W). The width of each voxel unit in the target voxel. represents the depth of each voxel unit in the original voxel, reflecting the physical distance in the real world corresponding to each pixel point in the slice spacing dimension (D). The depth of each voxel unit in the target voxel.

[0086] In this embodiment, taking the original voxel as (1., 1., 4.2000076) and the original image with the (H, W, D) dimensions of the original shape being (300, 320, 20) as a demonstration, the new shape and size obtained after interpolating it to the target voxel using the above calculation method is (320, 341, 21).

[0087] The purpose of the above preprocessing of the voxel is to establish an automatic configuration scheme for the model, that is, the automatic configuration of some parameters of the model, so as to shorten the subsequent training time of the model.

[0088] Downsampling is an important step in convolutional neural networks. While maintaining important information, it gradually reduces the size of the feature map, thereby capturing more abstract features at a higher level and improving the receptive field of the network. For multi-resolution anisotropic MRI datasets, a fixed downsampling process may lose too much information in a certain axis. For the fetal MRI brain tissue segmentation and classification task, the model uses an improved Unet segmentation network and a Resnet network as the backbone respectively, and automatically configures the downsampling parts of the two networks based on the data preprocessing results. When configuring the parameters of the lateral ventricle segmentation and classification model, the specific method is as follows:

[0089] Step S1-2-1, set the minimum feature map size for the U-Net segmentation module of the lateral ventricle segmentation and classification model.

[0090] Set the minimum feature map size for the (improved) U-Net network, corresponding to the bottom layer of the U-Net network, which will affect the network depth. This feature map size is manually set as a hyperparameter, mainly depending on the video memory of the GPU device used for training, the desired training time, and the desired training effect. Generally, the smaller the minimum feature map size, the larger the video memory required for the GPU device, the longer the training time, and the better the training effect may be.

[0091] Step S1-2-2: Calculate the mode of the sizes of each axis of all images in the fetal brain MRI image data, and record it as the mode size after recombination; then calculate the required downsampling times for each axis according to the mode size and the minimum feature map size.

[0092] For an input image with a size of (H, W, D), the calculation process of the downsampling times for each axis is as follows, where represents the minimum feature map size.

[0093]

[0094] In this embodiment, the mode size of the preprocessed dataset is (320, 320, 25). For the Unet network, FeatureMapMinSize is set to (4, 4, 4), and the calculated downsampling times for each axis are (6, 6, 2); for the Resnet network, since the minimum feature map size in the standard structure of Resnet is 7, and because the dataset in this example has anisotropy, the voxel depth corresponding to the D axis in the image is smaller than the voxel height corresponding to H and the voxel width corresponding to W, FeatureMapMinSize is set to (7, 7, 3), and the calculated downsampling times for each axis of the Resnet network are (5, 5, 3).

[0095] Step S1-2-3: Determine the input size of the lateral ventricle segmentation classification model according to the downsampling times.

[0096] As Figure 8 shown, according to the characteristics of downsampling, the input size of the model should satisfy that the input size of each axis can be divisible by the power of 2 of the downsampling times of each axis, which is expressed as:

[0097]

[0098] The method for calculating and regularizing the input size to meet the above conditions is:

[0099]

[0100] Among them, represents the downsampling times of each axis, represents the corresponding value of each axis in the image size.

[0101] In this embodiment, for the adversarial Unet network, the number of downsampling times for each axis is (6, 6, 2), the corresponding divisibility requirements are (64, 64, 4), the mode size used to calculate the number of downsampling times is (320, 320, 25), and the size is regularized to (320, 320, 28) to meet the conditions. After calculating the regularization conditions, the estimated video memory capacity consumed is calculated. In this method, the video memory occupied by a single MRI data during training should not exceed a certain value, so as to be used on more devices. The calculation method of the video memory is completed with the help of a common code library. This method adjusts the input size according to the video memory. When the video memory at the current input size needs to be greater than the set requirement, the ratio of the current size to the mode size is compared, and a value of the divisibility requirement is subtracted from the axis with the largest ratio to form a new input size, and the above work is repeated with the new input size until the requirement is met. In this method, for the improved Unet, the maximum video memory estimate for a single data is 4.5 GB, and for Resnet, the maximum video memory estimate for a single data is 1.5 GB. The input size of the improved Unet is (256, 256, 24), and the input size of the Resnet network is (224, 224, 24).

[0102] Step S1-2-4, according to the number of downsampling times required for each axis and the target voxel and the input size and the minimum feature map size calculate the configuration of the convolutional layer and the max pooling layer during the downsampling process.

[0103] When calculating the configuration of the convolutional layer and the max pooling layer, as Figure 9 shown, the specific steps are as follows:

[0104] Step S1-2-4-1, establish a loop with the maximum number of downsampling times as the starting value of the loop is 0, the ending value is , and the loop index is ;

[0105] Step S1-2-4-2, calculate the voxel difference between the axes in the voxel and check whether there is an axis whose voxel value is twice that of other axes;

[0106] Step S1-2-4-3, check the difference between the current feature map size and the minimum feature map size, and check whether there are some axes whose feature map size is less than twice the minimum feature map size;

[0107] Step S1-2-4-4, if there is an axis that meets the voxel requirements, adjust the kernel size of this axis to 1 in the convolutional layer and the max pooling layer; if it does not meet the voxel requirements but meets the feature map requirements, adjust the kernel size of this axis to 1 in the max pooling layer, and the convolutional layer uses a standard convolutional layer; if neither the voxel requirements nor the feature map requirements are met, each axis is processed according to the standard convolutional layer and pooling layer, that is, the convolutional kernel size of the convolutional layer is (3, 3, 3), and the convolutional kernel size of the pooling layer is (2, 2, 2);

[0108] Step S1-2-4-5, multiply the elements in the convolutional kernel of the max pooling layer by the elements of the voxels of the i-th layer according to the index to obtain the voxels of the (i + 1)-th layer, and divide the size of the feature map of the i-th layer by the elements of the convolutional kernel of the max pooling layer according to the index to obtain the size of the feature map of the (i + 1)-th layer; repeat to calculate the configurations of the convolutional layer and the max pooling layer of the (i + 1)-th layer according to steps S1-2-4-2, S1-2-4-3, and S1-2-4-4.

[0109] In this embodiment, the voxels of the first cycle are equivalent to the target voxels (0.9375, 0.9375, 3.9999969). The D axis has a two-fold relationship compared to the H axis and the W axis, and the size of this axis in the current feature map, which is 24, is greater than the minimum feature map size of 4. Therefore, in the first cycle, the kernel size will be set to (3, 3, 1), and the pooling layer configuration kernel size is (2, 2, 1) with a stride of (2, 2, 1). Finally, multiply the elements in the downsampling kernel size by the elements of the voxels of the layer according to the index to obtain the voxels of the layer, and divide the size of the feature map of the layer by the downsampling kernel size according to the index to obtain the size of the feature map of the layer, which is used for the Calculation of the convolutional layer and pooling layer of the layer network. In this example, the voxel of the 0th layer is (0.9375, 0.9375, 3.9999969), and the downsampling size is (2, 2, 1). Then the voxel of the 1st layer is (1.8750, 1.8750, 3.9999969), and the feature map size is (128, 128, 24). Finally, the convolutional layer configuration of the improved Unet network in this example is (3, 3, 1), (3, 3, 1), (3, 3, 3), (3, 3, 3), (3, 3, 3), (3, 3, 3), and the downsampling configuration is (2, 2, 1), (2, 2, 1), (2, 2, 2), (2, 2, 2), (2, 2, 1), (2, 2, 1). Since the convolutional kernel downsampling combination of the Resnet standard configuration is different from that of Unet, and the size of the first convolutional kernel is (7, 7, 7), the convolutional layer configuration of the Resnet network in this example is (7, 7, 1), (3, 3, 1), (3, 3, 3), (3, 3, 3), (3, 3, 3), and the downsampling configuration is (2, 2, 1), (2, 2, 1), (2, 2, 2), (2, 2, 2), (2, 2, 2).

[0110] Step S2, construct a lateral ventricle segmentation and classification model;

[0111] Construct a lateral ventricle segmentation and classification model. The lateral ventricle segmentation and classification model includes a U-Net segmentation module and a Resnet classification module. The output of the U-Net segmentation module is used as the input of the Resnet classification module;

[0112] As Figure 2 shown, the U-Net segmentation module uses a U-Net network, which includes multiple DWMF blocks for downsampling and multiple decoder blocks for upsampling. The DWMF blocks and the decoder blocks are connected by skip link blocks. Through the skip link blocks, feature maps of the same size are spliced, supplementing the spatial detail information required for upsampling to restore the image, and making full use of the rich spatial detail information in the encoder and the semantic information in the decoder. As Figure 3 shown, each DWMF block includes two depthwise separable convolution blocks and four depthwise separable channel dilated convolution blocks. The image data is used as the input of the two depthwise separable convolution blocks. The output of one of the depthwise separable convolution blocks is respectively input into three depthwise separable channel dilated convolution blocks. The outputs of these three depthwise separable channel dilated convolution blocks are jointly used as the input of the last depthwise separable channel dilated convolution block. The output of the last depthwise separable channel dilated convolution block is fused with the output of the other depthwise separable convolution block and used as the output of this DWMF block.

[0113] As Figure 4As shown in the figure, the Resnet classification module adopts the Resnet network, which includes a convolutional layer, a max pooling layer, four Resnet residual blocks, a global average pooling layer, and a fully connected layer arranged in sequence. The input layer consists of a convolutional layer with a size of 7×7×1. To ensure the correct Resnet network structure, its stride is set to (2, 2, 1) according to downsampling, followed by a max pooling layer with a size of 2×2×1 and a stride of 2, which is used for preliminary feature extraction and downsampling of the input image. Next are 4 stages, each stage is composed of multiple Resnet Block residual learning modules. The first layer of each stage is a max pooling layer. The convolutional kernel sizes of each stage are (3, 3, 3), (3, 3, 3), (3, 3, 3), (3, 3, 3). The first, second, third, and fourth stages contain 3, 4, 6, and 3 residual learning modules respectively. In each residual learning module, batch normalization (BatchNormalization) is used for standardization processing and the ReLU activation function is used to enhance the stability and non-linear expression ability of the network. Finally, the feature map of the last stage is converted into a vector of a fixed size through global average pooling, and it is mapped to the final number of output categories through a fully connected layer. Usually, a Softmax function is added before the fully connected layer to convert the output of the network into a probability distribution for image classification. In this example, there are two categories to be classified, namely "the lateral ventricle is dilated" and "the lateral ventricle is not dilated". Therefore, the regression module maps the finally extracted features to between 0 and 1 to reflect whether there is dilation of the lateral ventricle in the MRI image of the input case.

[0114] Step S3: Train the lateral ventricle segmentation and classification model;

[0115] Use the sample data and labels obtained in Step S1 to train the lateral ventricle segmentation and classification model constructed in Step S2 to obtain a mature lateral ventricle segmentation and classification model.

[0116] When training the lateral ventricle segmentation and classification model, the training method is not the innovation point of this application. Those skilled in the art can adopt existing training methods, loss functions, etc. according to the actual situation for training without creative labor.

[0117] Step S4: Real-time segmentation and classification;

[0118] Real-time obtain the fetal brain MRI image to be measured and input it into the lateral ventricle segmentation and classification model, and the lateral ventricle segmentation and classification model outputs the segmentation result and classification result.

[0119] Embodiment 2

[0120] This embodiment provides a fetal MRI brain tissue segmentation system, including:

[0121] A sample data and label acquisition module, which is used to acquire fetal brain MRI image data and label data. The label data includes the lateral ventricle segmentation label of each fetal brain MRI image and the classification label indicating whether there is lateral ventricle dilation.

[0122] The fetal brain MRI image data comes from 186 cases of fetal brain coronal plane MRI images provided by West China Second University Hospital / West China Women's and Children's Hospital, which includes samples with lateral ventricle dilation and samples with normal lateral ventricles. In addition, for each fetal brain MRI image, there is a lateral ventricle segmentation label and a classification label indicating whether there is lateral ventricle dilation, and the above labels are also marked by physicians in the imaging department of West China Second University Hospital / West China Women's and Children's Hospital.

[0123] After acquiring the fetal brain MRI image data, the voxel information of each fetal brain MRI image is counted, and the counted voxel information is preprocessed; according to the results of the data preprocessing, the parameters of the lateral ventricle segmentation classification model are configured. Specifically:

[0124] Traverse all sample data, read the voxel information and three-dimensional shape information of the MRI images, and perform statistics. The voxel size distribution of the obtained dataset is as Figure 5 shown.

[0125] After obtaining the voxel information of the sample data, preprocessing is performed to unify it into a voxel distribution and perform pixel normalization. The specific preprocessing method is:

[0126] Step S1-1-1, determine the target voxel for data preprocessing based on the counted voxel information. When determining the target voxel, as Figure 6 shown, the specific steps are:

[0127] Step S1-1-1-1, according to the counted voxel information, determine the median of the voxel information of all samples.

[0128] Step S1-1-1-2, according to the median of the counted voxel information, determine whether the voxel depth representing the slice spacing (D) of each slice in the MRI sequence exceeds twice the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W) of a single slice, and judge whether there is anisotropy in this dataset;

[0129] If there is anisotropy, go to step S1-1-1-3; if there is no anisotropy, go to step S1-1-1-4;

[0130] Step S1-1-1-3: Statistically analyze the voxel depths representing the slice spacing (D) of each slice in the MRI sequence for all samples as a list, sort the list from smallest to largest, and take the mode of the top 10% of the values to obtain the target depth value corresponding to the target voxel, so as to reduce distortion during the interpolation process; for the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W), take the median of all the samples statistically analyzed, and recombine it with the target depth value to form the target voxel. Among them, the height and width of the target voxel are the medians of the voxel widths and heights of the original images of all samples, and the depth of the target voxel is the mode of the top 10% of the values after statistically analyzing all samples;

[0131] Step S1-1-1-4: Statistically analyze the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W) of all the original images of the samples, and the voxel depths representing the slice spacing (D) of each slice in the MRI sequence as a list, take the median of each statistical list, and recombine it to form the target voxel.

[0132] According to Figure 5 the voxel distribution shown, in this embodiment, the median of the voxels is: (0.9375, 0.9375, 4.19998932), indicating that the axis with the index D representing the slice spacing of each slice in the MRI sequence has a relatively large difference from the other two axes. After statistically analyzing the values of the voxel depths representing the slice spacing (D) of each slice in all samples, select 3.9999969 as the target value, and recombine it with the median of the voxel height corresponding to the height (H) of a single slice in the MRI sequence and the voxel width corresponding to the width (W) to form (0.9375, 0.9375, 3.9999969) as the target voxel.

[0133] Step S1-1-2: Perform preprocessing based on the target voxel.

[0134] As Figure 7 shown, based on the difference between the original voxel information and the target voxel, calculate the new shape of the shape of the fetal brain MRI image after scaling the original voxel to the target voxel, and scale the fetal brain MRI image to the size of the new shape through trilinear interpolation to obtain the target image shape size of trilinear interpolation;

[0135] The calculation formula for the target image shape size of trilinear interpolation is:

[0136]

[0137] Among them, H, W, and D respectively represent the width, height, and slice spacing size of the slices in the MRI sequence, represents the height of each voxel unit in the original voxel, reflecting the physical distance in the real world corresponding to each pixel point in the slice height size (H), Represents the height of each voxel unit in the target voxel, Represents the width of each voxel unit in the original voxel, reflecting the physical distance in the real world corresponding to each pixel in the slice width dimension (W). The width of each voxel unit in the target voxel, Represents the depth of each voxel unit in the original voxel, reflecting the physical distance in the real world corresponding to each pixel in the slice spacing dimension (D), The depth of each voxel unit in the target voxel.

[0138] In this embodiment, taking the original voxel as (1., 1., 4.2000076) and the original image with the (H, W, D) dimensions of the original shape being (300, 320, 20) as a demonstration, after interpolating it to the target voxel using the above calculation method, the new shape dimensions obtained are (320, 341, 21).

[0139] The purpose of the above preprocessing of the voxel is to establish an automatic configuration scheme for the model, that is, the automatic configuration of some parameters of the model, so as to shorten the subsequent training time of the model.

[0140] Downsampling is an important step in convolutional neural networks. While maintaining important information, it gradually reduces the size of the feature map, thereby capturing more abstract features at a higher level and improving the receptive field of the network. For multi-resolution anisotropic MRI datasets, a fixed downsampling process may lose too much information in a certain axis. For the fetal MRI brain tissue segmentation and classification task, the model uses an improved Unet segmentation network and a Resnet network as the backbone respectively, and automatically configures the downsampling part of the two networks based on the data preprocessing results. When configuring the parameters of the lateral ventricle segmentation and classification model, the specific method is as follows:

[0141] Step S1-2-1, set the minimum feature map size for the U-Net segmentation module of the lateral ventricle segmentation and classification model.

[0142] Setting the minimum feature map size for the (improved) U-Net network, corresponding to the bottom layer of the U-Net network, will affect the network depth. This feature map size is set manually as a hyperparameter, mainly depending on the video memory of the GPU device used for training, the desired training time, and the desired training effect. Generally, the smaller the minimum feature map size, the larger the video memory of the GPU device required, the longer the training time, and the better the training effect may be.

[0143] Step S1-2-2, calculate the mode of the size of each axis of all images in the fetal brain MRI image data, and record it as the mode size after recombination; then calculate the required downsampling times for each axis according to the mode size and the minimum feature map size.

[0144] For an input image with dimensions (H, W, D), the downsampling times for each axis are calculated as follows, where represents the minimum feature map size.

[0145]

[0146] In this embodiment, the mode size of the preprocessed dataset is (320, 320, 25). For the Unet network, FeatureMapMinSize is set to (4, 4, 4), and the calculated downsampling times for each axis are (6, 6, 2). For the Resnet network, since the minimum feature map size in the standard structure of Resnet is 7, and because the dataset in this example has anisotropy, the voxel depth corresponding to the D axis in the image is smaller than the voxel height corresponding to H and the voxel width corresponding to W, FeatureMapMinSize is set to (7, 7, 3), and the calculated downsampling times for each axis of the Resnet network are (5, 5, 3).

[0147] Step S1-2-3, determine the input size of the lateral ventricle segmentation classification model according to the downsampling times.

[0148] As Figure 8 shown, according to the characteristics of downsampling, the input size of the model should satisfy that the input size of each axis can be divided by the power of 2 of the downsampling times of each axis, which is expressed as:

[0149]

[0150] The method for calculating and regularizing the input size to meet the above conditions is:

[0151]

[0152] where are the downsampling times for each axis, represents the corresponding value of each axis in the image size.

[0153] In this embodiment, for the Unet network, the number of downsampling times for each axis is (6, 6, 2), the corresponding divisibility requirements are (64, 64, 4), the mode size used to calculate the number of downsampling times is (320, 320, 25), and the size regularized to meet the conditions is (320, 320, 28). After calculating the regularization conditions, the estimated video memory capacity consumed is calculated. In this method, the video memory occupied by a single MRI data during training should not exceed a certain value, so as to be used on more devices. The calculation method of the video memory is completed with the help of a common code library. This method adjusts the input size according to the video memory. When the video memory under the current input size needs to be greater than the set requirement, compare the ratio of the current size to the mode size, subtract a value of the divisibility requirement from the axis with the largest ratio to form a new input size, and repeat the above work with the new input size until the requirement is met. In this method, for the improved Unet, the maximum video memory of a single data is estimated to be 4.5GB, and for Resnet, the maximum video memory of a single data is estimated to be 1.5GB. The input size of the improved Unet is (256, 256, 24), and the input size of the Resnet network is (224, 224, 24).

[0154] Step S1-2-4, according to the number of downsampling times required for each axis and the target voxel and the input size and the minimum feature map size calculate the configurations of the convolutional layer and the max pooling layer during the downsampling process.

[0155] When calculating the configurations of the convolutional layer and the max pooling layer, as Figure 9 shown, the specific steps are as follows:

[0156] Step S1-2-4-1, establish a loop with the maximum number of downsampling times as the starting value of the loop is 0, the ending value is , and the loop index is ;

[0157] Step S1-2-4-2, calculate the voxel differences between the axes in the voxel and check whether there is an axis whose voxel value is twice that of other axes;

[0158] Step S1-2-4-3, check the difference between the current feature map size and the minimum feature map size and check whether there is a feature map size of some axes that is less than twice the minimum feature map size;

[0159] Step S1-2-4-4: If there is an axis that meets the voxel requirements, adjust the kernel size of this axis to 1 in the convolutional layer and the max pooling layer; if it does not meet the voxel requirements but meets the feature map requirements, adjust the kernel size of this axis to 1 in the max pooling layer, and use the standard convolutional layer in the convolutional layer; if neither the voxel requirements nor the feature map requirements are met, each axis is processed according to the standard convolutional layer and pooling layer, that is, the convolutional kernel size of the convolutional layer is (3, 3, 3), and the convolutional kernel size of the pooling layer is (2, 2, 2).

[0160] Step S1-2-4-5: Multiply the elements in the convolutional kernel of the max pooling layer by the elements of the voxels in the i-th layer according to the index to obtain the voxels of the (i + 1)-th layer, and divide the size of the feature map of the i-th layer by the elements of the convolutional kernel of the max pooling layer according to the index to obtain the size of the feature map of the (i + 1)-th layer; repeat to calculate the configurations of the convolutional layer and the max pooling layer of the (i + 1)-th layer according to Steps S1-2-4-2, S1-2-4-3, and S1-2-4-4.

[0161] In this embodiment, the voxels of the first cycle are equivalent to the target voxels (0.9375, 0.9375, 3.9999969). There is a two-fold relationship between the D axis and the H axis and the W axis, and the size of this axis in the current feature map is 24, which is greater than the minimum feature map size of 4. Therefore, in the first cycle, the convolutional kernel size will be set to (3, 3, 1), and the pooling layer configuration kernel size is (2, 2, 1), and the stride is (2, 2, 1). Finally, multiply the elements in the downsampling kernel size by the elements of the voxels in the layer according to the index to obtain the voxels of the layer, and divide the size of the feature map of the layer by the downsampling kernel size according to the index to obtain the size of the feature map of the layer, which is used for the Calculation of the convolutional layer and pooling layer of the layer network. In this example, the voxel of the 0th layer is (0.9375, 0.9375, 3.9999969), and the downsampling size is (2, 2, 1). Then the voxel of the 1st layer is (1.8750, 1.8750, 3.9999969), and the feature map size is (128, 128, 24). Finally, the convolutional layer configuration of the improved Unet network in this example is (3, 3, 1), (3, 3, 1), (3, 3, 3), (3, 3, 3), (3, 3, 3), (3, 3, 3), and the downsampling configuration is (2, 2, 1), (2, 2, 1), (2, 2, 2), (2, 2, 2), (2, 2, 1), (2, 2, 1). Since the convolutional kernel downsampling combination of the Resnet standard configuration is different from that of Unet, and the size of the first convolutional kernel is (7, 7, 7), the convolutional layer configuration of the Resnet network in this example is (7, 7, 1), (3, 3, 1), (3, 3, 3), (3, 3, 3), (3, 3, 3), and the downsampling configuration is (2, 2, 1), (2, 2, 1), (2, 2, 2), (2, 2, 2), (2, 2, 2).

[0162] Lateral ventricle segmentation and classification model construction module, used to construct a lateral ventricle segmentation and classification model. The lateral ventricle segmentation and classification model includes a U-Net segmentation module and a Resnet classification module. The output of the U-Net segmentation module is used as the input of the Resnet classification module.

[0163] As Figure 2 shown, the U-Net segmentation module uses a U-Net network, including multiple DWMF blocks for downsampling and multiple decoder blocks for upsampling. The DWMF blocks and the decoder blocks are connected by skip link blocks. Through the skip link blocks, feature maps of the same size are concatenated, supplementing the spatial detail information required for upsampling to restore the image, and making full use of the rich spatial detail information in the encoder and the semantic information in the decoder. As Figure 3 shown, each DWMF block includes two depthwise separable convolution blocks and four depthwise separable channel dilation convolution blocks. The image data is used as the input of the two depthwise separable convolution blocks. The output of one of the depthwise separable convolution blocks is respectively input into three depthwise separable channel dilation convolution blocks. The outputs of these three depthwise separable channel dilation convolution blocks are jointly used as the input of the last depthwise separable channel dilation convolution block. The output of the last depthwise separable channel dilation convolution block is fused with the output of the other depthwise separable convolution block and used as the output of the DWMF block.

[0164] As Figure 4As shown in the figure, the Resnet classification module uses the Resnet network, including a convolutional layer, a max pooling layer, four Resnet residual blocks, a global average pooling layer, and a fully connected layer arranged in sequence. The input layer consists of a convolutional layer with a size of 7×7×1. To ensure the correct Resnet network structure, its stride is configured as (2,2,1) according to downsampling, followed by a max pooling layer with a size of 2×2×1 and a stride of 2, which is used for preliminary feature extraction and downsampling of the input image. Next are 4 stages, each stage is composed of multiple Resnet Block residual learning modules. The first layer of each stage is a max pooling layer. The convolutional kernel sizes of each stage are (3,3,3), (3,3,3), (3,3,3), (3,3,3). The first, second, third, and fourth stages contain 3, 4, 6, and 3 residual learning modules respectively. Batch Normalization is used for standardization processing and the ReLU activation function is used in each residual learning module to enhance the stability and non-linear expression ability of the network. Finally, the feature map of the last stage is converted into a vector of a fixed size through global average pooling, and it is mapped to the final number of output categories through a fully connected layer. A Softmax function is usually added before the fully connected layer to convert the output of the network into a probability distribution for image classification. In this example, there are two categories to be classified, namely "the lateral ventricle is dilated" and "the lateral ventricle is not dilated". Therefore, the regression module maps the finally extracted features to between 0 and 1 to reflect whether there is dilation of the lateral ventricle in the MRI image of the input case.

[0165] The lateral ventricle segmentation and classification model training module is used to train the lateral ventricle segmentation and classification model constructed by the lateral ventricle segmentation and classification model construction module using the sample data and labels obtained by the sample data and label acquisition module, and obtain a mature lateral ventricle segmentation and classification model.

[0166] When training the lateral ventricle segmentation and classification model, the training method is not the innovation point of this application. Those skilled in the art can adopt existing training methods, loss functions, etc. according to the actual situation for training without creative labor.

[0167] The real-time segmentation and classification module is used to obtain the fetal brain MRI image to be measured in real time and input it into the lateral ventricle segmentation and classification model, and the lateral ventricle segmentation and classification model outputs the segmentation result and the classification result.

[0168] Embodiment 3

[0169] A computer device includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the fetal MRI brain tissue segmentation method.

[0170] Among them, the computer device may be a computing device such as a desktop computer, a notebook, a palm computer, or a cloud server. The computer device can perform human-computer interaction with the user through a keyboard, a mouse, a remote control, a touchpad, or a voice control device, etc.

[0171] The memory at least includes one type of readable storage medium, and the readable storage medium includes flash memory, a hard disk, a multimedia card, a card-type memory (such as an SD or D interface display memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk equipped on the computer device, a SmartMedia Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Of course, the memory may also include both the internal storage unit and the external storage device of the computer device. In this embodiment, the memory is commonly used to store the operating system and various application software installed on the computer device, such as the program code of the fetal MRI brain tissue segmentation method. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0172] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor is generally used to control the overall operation of the computer device. In this embodiment, the processor is used to run the program code stored in the memory or process data, such as running the program code of the fetal MRI brain tissue segmentation method.

[0173] Embodiment 4

[0174] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor is caused to execute the steps of the fetal MRI brain tissue segmentation method.

[0175] Among them, the computer-readable storage medium stores an interface display program, and the interface display program can be executed by at least one processor so that the at least one processor executes the steps of the fetal MRI brain tissue segmentation method as described above.

[0176] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation method. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the fetal MRI brain tissue segmentation method described in the embodiments of the present application.

Claims

1. A fetal MRI brain tissue segmentation method, characterized in that: The following steps are involved: Step S1, obtaining sample data and labels; Obtaining fetal brain MRI image data and label data, the label data including a lateral ventricle segmentation label of each fetal brain MRI image and a classification label of whether lateral ventricle dilatation exists; Step S2, constructing a lateral ventricle segmentation classification model; Constructing a lateral ventricle segmentation and classification model, which includes a U-Net segmentation module and a Resnet classification module. The output of the U-Net segmentation module is used as the input of the Resnet classification module. The U-Net segmentation module adopts a U-Net network, including multiple DWMF blocks for downsampling and multiple decoder blocks for upsampling. The DWMF blocks and the decoder blocks are connected by jump link blocks; each DWMF block includes two depth separation convolution blocks and four depth separation channel dilated convolution blocks. The image data is used as the input of the two depth separation convolution blocks, and the output of one depth separation convolution block is respectively input into three depth separation channel dilated convolution blocks. The outputs of the three depth separation channel dilated convolution blocks are used as the input of the last depth separation channel dilated convolution block. The output of the last depth separation channel dilated convolution block is fused with the output of another depth separation convolution block as the output of the DWMF block; Step S3, training a lateral ventricle segmentation classification model; The sample data and labels obtained in step S1 are used to train the lateral ventricle segmentation classification model constructed in step S2 to obtain a mature lateral ventricle segmentation classification model; Step S4, real-time segmentation and classification; The MRI image of the fetal brain to be tested is acquired in real time and input into the lateral ventricle segmentation classification model, and the lateral ventricle segmentation classification model outputs the segmentation result and the classification result.

2. The fetal MRI brain tissue segmentation method according to claim 1, characterized in that: In step S1, after obtaining the fetal brain MRI image data, the voxel information of each fetal brain MRI image is counted, and the counted voxel information is preprocessed; according to the result of the data preprocessing, the parameters of the lateral ventricle segmentation classification model are configured.

3. The fetal MRI brain tissue segmentation method according to claim 2, characterized in that: When preprocessing the statistical voxel information, the specific preprocessing method is: Step S1-1-1, determining the target voxel for data preprocessing based on the statistical voxel information; Step S1-1-2, preprocessing based on the target voxel; Based on the difference between the original voxel information and the target voxel, the new shape of the fetal brain MRI image is calculated after the original voxel is scaled to the target voxel, and the fetal brain MRI image is scaled to the size of the new shape by trilinear interpolation to obtain the target image shape size of trilinear interpolation; The calculation formula for the shape and size of the target image of trilinear interpolation is: Where H, W, and D represent the width, height, and slice spacing of the slice in the MRI sequence, respectively. Indicates the height of each voxel unit in the original voxel, reflecting the physical distance in the real world corresponding to each pixel point of the slice height dimension H. represents the height of each voxel unit in the target voxel, Represents the width of each voxel unit in the original voxel, reflecting the physical distance of each pixel in the slice width size W corresponding to the real world The width per voxel unit in the target voxel, Indicates the depth of each voxel unit in the original voxel, reflecting the physical distance of each pixel point in the slice spacing size D corresponding to the real world. The depth per voxel unit in the target voxel.

4. The fetal MRI brain tissue segmentation method according to claim 3, characterized in that: In step S1-1-1, the method for determining the target voxel is: Step S1-1-1-1, determining the median of the voxel information of all samples according to the statistical voxel information; Step S1-1-1-2, judging whether the sample is anisotropic according to the median of the statistical voxel information, whether the voxel depth representing the interval dimension D of each slice of the MRI sequence exceeds twice the voxel height corresponding to the height dimension H and the voxel width corresponding to the width dimension W of a single slice of the MRI sequence; If anisotropy exists, proceed to step S1-1-1-3; if anisotropy does not exist, proceed to step S1-1-1-4; Step S1-1-1-3, count the voxel depths of all samples representing the spacing size D of each slice of the MRI sequence as a list, and sort the list from small to large and take the mode of the top 10% values ​​as the target depth value corresponding to the target voxel to reduce distortion in the interpolation process; for the voxel height corresponding to the height size H and the voxel width corresponding to the width size W representing a single slice of the MRI sequence, take the median of all the samples counted, and reorganize them with the target depth value into the target voxel; wherein, the height and width of the target voxel are the medians of the voxel width and height of the original image of all samples, and the depth of the target voxel is the mode of the top 10% values ​​after counting all the samples; Step S1-1-1-4, count all sample original images to represent the voxel height corresponding to the height dimension H of a single slice of the MRI sequence, the voxel width corresponding to the width dimension W, and the voxel depth of the spacing dimension D of each slice of the MRI sequence as a list, take the median of each statistical list, and reorganize it into the target voxel.

5. The fetal MRI brain tissue segmentation method according to claim 2, characterized in that: In step S1, when configuring the parameters of the lateral ventricle segmentation classification model, the specific method is as follows: Step S1-2-1, setting a minimum feature map size for a U-Net segmentation module of a lateral ventricle segmentation classification model; Step S1-2-2, calculating the mode of the size of each axis of all images in the fetal brain MRI image data, and recording it as the mode size after reorganization; and then calculating the required number of downsampling times of each axis according to the mode size and the minimum feature map size; Step S1-2-3, determining the input size of the lateral ventricle segmentation classification model according to the number of downsampling times; Step S1-2-4, calculates the configuration of the convolution layer and the maximum pooling layer during the downsampling process according to the number of downsampling times required for each axis, the target voxel, the input size, and the minimum feature map size.

6. The fetal MRI brain tissue segmentation method according to claim 5, characterized in that: In step S1-2-4, when calculating the configuration of the convolution layer and the maximum pooling layer, the specific steps are: Step S1-2-4-1, establish a loop with the maximum number of downsampling times, the loop start value is 0, and the end value is the number of downsampling times , the loop index is ; Step S1-2-4-2, calculating the voxel differences between the axes in the voxel, and checking whether there is an axis whose voxel value is twice that of other axes; Step S1-2-4-3, check the difference between the current feature map size and the minimum feature map size, and check whether the feature map size of some axes is less than twice the minimum feature map; Step S1-2-4-4, if there is an axis that meets the voxel requirement, the kernel size of the axis is adjusted to 1 in the convolution layer and the maximum pooling layer; if it does not meet the voxel requirement but meets the feature map requirement, the kernel size of the axis is adjusted to 1 in the maximum pooling layer, and the convolution layer uses the standard convolution layer; if neither the voxel requirement nor the feature map requirement is met, each axis is processed according to the standard convolution layer and pooling layer, that is, the convolution kernel size of the convolution layer is (3,3,3), and the convolution kernel size of the pooling layer is (2,2,2); Step S1-2-4-5, multiply the elements in the convolution kernel of the maximum pooling layer by the elements of the voxels of the i-th layer by index to obtain the voxels of the i+1-th layer, divide the feature map size of the i-th layer by the elements of the convolution kernel of the maximum pooling layer by index to obtain the feature map size of the i+1-th layer; repeat steps S1-2-4-2, S1-2-4-3, and S1-2-4-4 to calculate the configuration of the convolution layer and the maximum pooling layer of the i+1-th layer.

7. The fetal MRI brain tissue segmentation method according to claim 1, characterized in that: In step S2, the Resnet classification module adopts a Resnet network, including a convolutional layer, a maximum pooling layer, four Resnet residual blocks, a global average pooling layer and a fully connected layer arranged in sequence.

8. A fetal MRI brain tissue segmentation system, characterized in that: include: A sample data and label acquisition module is used to acquire fetal brain MRI image data and label data, wherein the label data includes a lateral ventricle segmentation label of each fetal brain MRI image and a classification label of whether there is lateral ventricle dilatation; A lateral ventricle segmentation classification model construction module is used to construct a lateral ventricle segmentation classification model. The lateral ventricle segmentation classification model includes a U-Net segmentation module and a Resnet classification module. The output of the U-Net segmentation module serves as the input of the Resnet classification module. The U-Net segmentation module adopts a U-Net network, including multiple DWMF blocks for downsampling and multiple decoder blocks for upsampling. The DWMF blocks and the decoder blocks are connected by jump link blocks; each DWMF block includes two depth separation convolution blocks and four depth separation channel dilated convolution blocks. The image data is used as the input of the two depth separation convolution blocks, and the output of one depth separation convolution block is respectively input into three depth separation channel dilated convolution blocks. The outputs of the three depth separation channel dilated convolution blocks are used as the input of the last depth separation channel dilated convolution block. The output of the last depth separation channel dilated convolution block is fused with the output of another depth separation convolution block as the output of the DWMF block; The lateral ventricle segmentation classification model training module is used to train the lateral ventricle segmentation classification model constructed by the lateral ventricle segmentation classification model construction module using the sample data and labels obtained by the sample data and label acquisition module to obtain a mature lateral ventricle segmentation classification model; The real-time segmentation and classification module is used to obtain the fetal brain MRI image to be tested in real time and input the lateral ventricle segmentation and classification model, and the lateral ventricle segmentation and classification model outputs the segmentation result and the classification result.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images