Method, device, medium and equipment for predicting fetal age

Abstract

The present invention discloses a method, apparatus, storage medium, and computer device for predicting gestational age based on fetal brain magnetic resonance images and prior information. A three-dimensional convolutional network is proposed to extract brain structural information from fetal MRI images to obtain fetal brain structural features. The network, composed of asymmetric convolutional layers and attention enhancement layers, uses the entire 3D image as input and selectively emphasizes key information features, adapting to random variations in the position and orientation of the fetal brain. A discrete area distribution vector of the fetal brain is proposed to represent the distribution of brain region areas in MRI images, improving the robustness and effectiveness of the introduced prior knowledge. The method also proposes a fusion regression module to integrate the fetal brain structural features and fetal brain area distribution features for gestational age estimation based on fetal brain MRI images. The proposed gestational age prediction method achieves optimal accuracy and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear magnetic resonance image processing, and in particular to a method, device, storage medium and computer equipment for predicting fetal age based on nuclear magnetic resonance images and prior information. Background Art

[0002] Gestational age prediction is widely used in downstream tasks to evaluate the quality of fetal brain MRI reconstructions, as reconstructions lack real-world results to assess them. However, unpredictable fetal movements and abdominal movements caused by maternal breathing can cause artifacts and blurring in fetal brain MRI images, posing a significant challenge to MRI-based gestational age prediction.

[0003] Among the existing methods, gestational age prediction based on MRI images directly uses classic natural image classification networks to extract features from fetal MRI images to predict gestational age. This does not take into account the small sample size problem and the inherent characteristics of fetal brain MRI images, resulting in inaccurate gestational age prediction and insufficient interpretability. Summary of the Invention

[0004] The present invention aims to address the aforementioned deficiencies in the prior art by providing a method, apparatus, storage medium, and computer device for predicting gestational age based on fetal brain magnetic resonance images and prior information. The method first uses localization and segmentation to extract brain regions from magnetic resonance images. The area of the brain regions at each MRI image slice is calculated to form a discrete area distribution vector of the fetal brain. A multilayer perceptron is then used to extract features from these discrete area distribution vectors to obtain brain area distribution features of the magnetic resonance images. A three-dimensional fetal convolutional network is then used to extract features from magnetic resonance images containing only the brain regions to obtain brain structural features of the magnetic resonance images. Finally, a fusion regression module is used to fuse the brain area distribution features and brain structural features of the magnetic resonance images and regress them to output a predicted value for the fetal gestational age.

[0005] A first object of the present invention is to provide a method for predicting gestational age based on nuclear magnetic resonance images and prior information, the method comprising the following steps:

[0006] S1. Obtain a fetal MRI image dataset with known gestational age results, and preprocess the dataset, wherein the preprocessing is to extract brain regions from the MRI images by localization and segmentation;

[0007] S2. Based on the brain regions extracted in step S1, the discrete area distribution vector of the fetal brain is calculated based on the brain area of each MRI image layer. The discrete area distribution vector is extracted with the help of a multi-layer perceptron to obtain the brain area distribution feature of the MRI image. The brain area distribution feature is expressed as

[0008] S3, based on the brain region extracted in step S1, a nuclear magnetic resonance image containing only the brain region is generated, and the nuclear magnetic resonance image containing only the brain region is subjected to feature extraction by means of a three-dimensional fetal convolutional network to obtain brain structural features of the nuclear magnetic resonance image. The brain structural features are expressed as

[0009] The three-dimensional fetal convolutional network includes a first, second, and third fetal feature extraction modules, a global pooling layer, and a fully connected layer connected in sequence. The first, second, and third fetal feature extraction modules have exactly the same structure. Each fetal feature extraction module includes two asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer connected in sequence. The input of the first fetal feature extraction module is a magnetic resonance image containing only the brain region. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and serves as the input of the second fetal feature extraction module. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and the output of the second fetal feature extraction module and serves as the input of the third fetal feature extraction module.

[0010] S4. Fusion of the discrete area distribution vector features and the nuclear magnetic resonance image features by a fusion regression module, and regression output of a predicted value of gestational age.

[0011] Furthermore, in step S1, positioning refers to using a box to locate the brain position at each level, and segmentation refers to extracting a set of pixel points in the box belonging to the brain area. Positioning and segmentation can eliminate background noise and allow the model to focus more on the brain area. Positioning and segmentation can be completed by manual annotation by a doctor or by positioning and segmentation software.

[0012] Furthermore, the process of step S2 is as follows:

[0013] S21. Assume that the MRI image contains n slices. The brain area calculated from the n slices is combined into an n×1 discrete area distribution vector S=[s1,...,s i ,...,s n ], adding prior knowledge in the form of a vector can improve the robustness and effectiveness of the added prior knowledge. The calculation formula of the brain area of the i-th slice is expressed as where N i is the number of pixels occupied by the brain area, L×W×H mm 3 is the voxel size of the MRI image, L, W, and H are the length, width, and height of the voxel, respectively. Introducing L and W in the formula can enhance the generalization of the model across different datasets.

[0014] S22. After calculating the discrete area distribution vector of the fetus, a multilayer perceptron is used to extract features of the discrete area distribution vector. The multilayer perceptron has a two-layer structure, the first layer is a hidden layer, and the input of the hidden layer is the discrete area distribution vector S = [s1, ..., s i ,...,s n ], the output is C=[c1,...,c j ,...,c 12 ],in in is the connection parameter matrix of the hidden layer, p ij represents the element in row i and column j in P, g() is the activation function, and the activation function is selected from one of the ReLU function, sigmoid function and tanh function; the second layer is the output layer, the input of the output layer is C, and the output is the brain area distribution feature of the magnetic resonance imaging. in is the connection parameter matrix of the output layer, q jk Represents the j-th row and k-th column in Q.

[0015] Furthermore, the input of the first fetal feature extraction module in the three-dimensional fetal convolutional network is a magnetic resonance image containing only the brain area. The input image is output after passing through two sequentially connected asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer in the first fetal feature extraction module. The asymmetric convolutional layer decomposes the standard 3×3 convolution into 1×3 convolution and 3×1 convolution and uses three-dimensional convolution to ensure the global information of the entire brain. The attention enhancement layer uses Squeeze-and Excitation Block, which is a public technology from "Squeeze-and-Excitation Networks" published by authors Jie Hu and Li Shen and Samuel Albanie and Gang Sun and Enhua Wu in the journal "IEEE Transactions on Pattern Analysis and Machine Intelligence, year: 2017, volume: 42, pages: 2011-2023". The Squeeze-and Excitation Block filters the input global information and emphasizes important information output. The attention enhancement layer selectively emphasizes key information features and suppresses useless features, thereby adapting to the random changes in the position and direction of the fetal brain. The step size of the maximum pooling layer is set to 2. After the image passes through the maximum pooling layer, the length and width will be reduced by half. The length, width, and height of the MRI image containing only the brain area are L 1 、W 1 、H 1 The output of the first fetal feature extraction module is the length, width, and height corresponding to the size L 2 、W 2 、H 2 images;

[0016] The MRI image containing only the brain region and the output of the first fetal feature extraction module are jump-joined as the input of the second fetal feature extraction module. The structure of the second fetal feature extraction module is the same as that of the first fetal feature extraction module. The MRI image containing only the brain region, the output of the first fetal feature extraction module, and the output of the second fetal feature extraction module are jump-joined as the input of the third fetal feature extraction module. The structure of the third fetal feature extraction module is the same as that of the first fetal feature extraction module.

[0017] Among them, jump stitching is to stitch the high dimension of the image. During stitching, since the length and width of the image are different, the length and width need to be transformed before stitching. The size conversion operations are pooling, upsampling and transposed convolution.

[0018] The global pooling layer and the fully connected layer connected in sequence at the end of the three-dimensional fetal convolutional network can ensure that the output format of step S3 is consistent with the output format of step S2, so as to ensure that the subsequent steps can better integrate the brain area distribution characteristics and brain structure characteristics of the magnetic resonance imaging.

[0019] Furthermore, the processing process of the fusion regression module is as follows:

[0020] The fusion regression module includes a feature splicing layer and a first and a second fully connected layer connected in sequence. The feature splicing layer combines the brain area distribution features of the magnetic resonance image and brain structural characteristics Spliced into a multimodal feature vector The first and second fully connected layers extract features from the multimodal feature vector Y and regress and output the predicted value of gestational age. The use of two layers of full connection can better integrate and regress these two features. The expression formula is: Prediction value = in is the connection parameter matrix of the first fully connected layer, r uv represents the element in row u and column v in R, W = [w1,w2,w3,w4] is the connection parameter matrix of the second fully connected layer, w v Represents the vth vector in W, g() is the activation function, and the activation function is selected from one of the ReLU function, sigmoid function and tanh function.

[0021] A second object of the present invention is to provide a gestational age prediction device based on nuclear magnetic resonance images and prior information, wherein:

[0022] a data set preprocessing unit, which obtains a fetal magnetic resonance image data set with known gestational age results and preprocesses the data set, wherein the preprocessing extracts brain regions from the magnetic resonance image by positioning and segmenting;

[0023] The brain area distribution feature extraction unit calculates the discrete area distribution vector of the fetal brain composed of the brain area of each MRI image layer based on the brain area extracted by the data set preprocessing unit. The multi-layer perceptron is used to extract the features of the discrete area distribution vector to obtain the brain area distribution feature of the MRI image. The brain area distribution feature is expressed as

[0024] The brain structure feature extraction unit generates a MRI image containing only the brain area based on the brain area extracted by the data set preprocessing unit. The MRI image containing only the brain area is extracted with the help of a three-dimensional fetal convolutional network to obtain the brain structure features of the MRI image. The brain structure features are expressed as

[0025] The three-dimensional fetal convolutional network includes a first, second, and third fetal feature extraction modules, a global pooling layer, and a fully connected layer connected in sequence. The first, second, and third fetal feature extraction modules have exactly the same structure. Each fetal feature extraction module includes two asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer connected in sequence. The input of the first fetal feature extraction module is a magnetic resonance image containing only the brain region. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and serves as the input of the second fetal feature extraction module. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and the output of the second fetal feature extraction module and serves as the input of the third fetal feature extraction module.

[0026] The prediction unit fuses the discrete area distribution vector features and the nuclear magnetic resonance image features through a fusion regression module and regresses to output a predicted value of the gestational age.

[0027] The third object of the present invention is to provide a computer device comprising a processor and a memory for storing a program executable by the processor, wherein when the processor executes the program stored in the memory, the method for predicting fetal age based on magnetic resonance imaging and prior information is implemented.

[0028] A fourth object of the present invention is a storage medium storing a program, which, when executed by a processor, implements the method for predicting fetal age based on magnetic resonance imaging and prior information.

[0029] The present invention has the following advantages and effects compared to the prior art:

[0030] 1. The present invention proposes a new three-dimensional fetal convolutional network for extracting structural and spatial information from magnetic resonance images. The three-dimensional fetal convolutional network consists of asymmetric convolutional layers and attention enhancement layers, which can utilize the global information of the entire fetal brain and adapt to random changes in the position and direction of the fetal brain.

[0031] 2. This invention also proposes a discrete area distribution vector for the fetal brain to represent the distribution of brain regions in MRI images. This improves the robustness and effectiveness of the introduced prior knowledge and enhances the interpretability of the entire gestational age prediction method.

[0032] 3. This paper also proposes a fusion regression module for fusing discrete area distribution vector features with brain structural features from MRI images to achieve gestational age prediction based on MRI images. The dual streams consist of a volume stream that uses prior knowledge to learn prior features and an image stream that learns fetal brain structural features. This gestational age prediction method was tested on two major benchmarks: Dataset-GD and Dataset-SH. The method achieved state-of-the-art results on both datasets, with mean gestational age errors of 0.624 and 0.812 weeks, respectively, and was more robust in handling noise and small sample size learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0034] Figure 1 This is a flowchart of a method for predicting fetal age based on magnetic resonance imaging and prior information disclosed in the present invention;

[0035] Figure 2 This is a principle block diagram of the three-dimensional fetal convolutional network module disclosed in the present invention;

[0036] Figure 3 It is a principle block diagram of the multilayer perceptron disclosed in the present invention;

[0037] Figure 4 This is a data comparison chart of the predicted gestational age and the actual gestational age in Example 1 of the present invention;

[0038] Figure 5 This is a data comparison chart of the predicted gestational age and the actual gestational age in Example 2 of the present invention;

[0039] Figure 6 is a structural block diagram of a gestational age prediction device in Example 3 of the present invention;

[0040] Figure 7 This is a structural block diagram of the computer device in Example 4 of the present invention. DETAILED DESCRIPTION

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0042] Example 1

[0043] This embodiment further discloses a method for predicting fetal age based on nuclear magnetic resonance images and prior information. The specific process is as follows:

[0044] S1. Data acquisition: This example selects the Dataset-GD dataset, which contains 40 samples in the training set and 10 samples in the test set. The gestational age range is 24 to 36 weeks of gestation, and the image size is 560×560×30 pixels. 3 , the voxel size is 0.71×0.71×3.00mm 3 ;

[0045] S2, data preprocessing: extracting brain regions from MRI images through localization and segmentation, where localization and segmentation are performed by localization and segmentation software;

[0046] S3, based on the brain regions extracted in step S1, calculate the brain region area of each layer in the MRI image to form the discrete area distribution vector S = [s1, ..., s i ,...,s 30 ], where the calculation formula for the brain area at the i-th level is expressed as

[0047] S4, generating a magnetic resonance imaging image containing only the brain region based on the brain region extracted in step S1;

[0048] S5, based on the discrete area distribution vectors generated in steps S3 and S4 and the MRI images containing only the brain area, the multilayer perceptron, the three-dimensional fetal convolutional network and the fusion regression module are trained. The training process is to use the gradient descent method to train the hidden layer connection parameter matrix of the multilayer perceptron. The output layer connection parameter matrix of the multilayer perceptron The connection parameter matrix of each convolution kernel and the first fully connected layer of the fusion regression module in the 3D fetal convolutional network And the connection parameter matrix W = [w1,w2,w3,w4] of the second fully connected layer, thus obtaining a method that can predict fetal brain age; the prediction method uses As the loss function, where y i ′ is the predicted age, y i is the real age, and the learning rate is set to 1e -3 to 2e -3 The experimental method used five-fold three-fold cross validation, and the experimental evaluation method used three evaluation indicators, namely, the mean gestational age error, R 2 Correlation coefficient, accuracy, among which R 2The correlation coefficient is the Pearson correlation coefficient between the predicted gestational age and the actual gestational age. N e<1.5 The number of gestational age errors in the data set is less than 1.5, and N is the total number of data in the data set.

[0049] Final mean gestational age prediction error, R 2 The correlation coefficient and accuracy can reach 0.624, 0.934, and 90%, respectively. Table 1 is the comparison results of the classification method provided in this embodiment on Dataset-GD with other methods based on convolutional neural networks and some of their derivatives. It can be seen that the method proposed in the present invention achieves the best results. Figure 4 This is a comparison chart of the predicted gestational age and the actual gestational age of 50 data on Dataset-GD using the classification method provided in this embodiment. The horizontal axis is the actual gestational age and the vertical axis is the predicted gestational age. It can be seen that the gap between the gestational age predicted by the method proposed in the present invention and the actual gestational age is very small.

[0050] Table 1. Comparison of experimental results of the gestational age prediction method in Example 1 and other methods

[0051] method Mean gestational age error <![CDATA[R 2 Correlation coefficient]]> Accuracy TSAN 1.343 0.709 0.68 VGG-reg 1.074 0.773 0.70 ResNet-reg 0.971 0.768 0.78 SENet-reg 0.905 0.847 0.76 DLVGG-reg 0.880 0.865 0.82 The method proposed by the present invention 0.624 0.934 0.90

[0052] Example 2

[0053] This embodiment further discloses a method for predicting fetal age based on nuclear magnetic resonance images and prior information. The specific process is as follows:

[0054] S1. Data acquisition: This example uses the Dataset-SH dataset, which contains 200 training and 50 test samples, with a gestational age range of 21 to 36 weeks of gestation and an image size of 256×256×24 pixels. 3 , the voxel size is 0.55×0.55×4.4mm 3 ;

[0055] S2, data preprocessing: extracting brain regions from MRI images through localization and segmentation. The localization and segmentation here are manually annotated by doctors.

[0056] S3, based on the brain regions extracted in step S1, calculate the brain region area of each MRI image layer to form the discrete area distribution vector S of the fetal brain = [s1, ..., s i ,...,s 24 ], where the calculation formula for the brain area at the i-th level is expressed as

[0057] S4, generating a magnetic resonance imaging image containing only the brain region based on the brain region extracted in step S1;

[0058] S5, based on the discrete area distribution vectors generated in steps S3 and S4 and the MRI images containing only the brain area, the multilayer perceptron, the three-dimensional fetal convolutional network and the fusion regression module are trained. The training process is to use the gradient descent method to train the hidden layer connection parameter matrix of the multilayer perceptron. The output layer connection parameter matrix of the multilayer perceptron The connection parameter matrix of each convolution kernel and the first fully connected layer of the fusion regression module in the 3D fetal convolutional network And the connection parameter matrix W = [w1,w2,w3,w4] of the second fully connected layer, thus obtaining a method that can predict fetal brain age; the prediction method uses As the loss function, where y i ′ is the predicted age, y i is the real age, and the learning rate is set at 5e -4 to 2e -3 The experimental method used five-fold three-fold cross validation, and the experimental evaluation method used three evaluation indicators, namely, the mean gestational age error, R 2 Correlation coefficient, accuracy, among which R 2 The correlation coefficient is the Pearson correlation coefficient between the predicted age and the actual age, N e<1.5 The number of gestational age errors in the data set is less than 1.5, and N is the total number of data in the data set.

[0059] Final mean gestational age prediction error, R 2 The correlation coefficient and accuracy can reach 0.812, 0.933, and 85%, respectively. Table 2 is the comparison results of the classification method provided in this embodiment on Dataset-SH with other methods based on convolutional neural networks and some of their derivatives. It can be seen that the method proposed in the present invention achieves the best results. Figure 5 This is a comparison chart of the predicted gestational age and the actual gestational age of 250 data on Dataset-SH using the classification method provided in this embodiment. The horizontal axis is the actual gestational age and the vertical axis is the predicted gestational age. It can be seen that the gap between the gestational age predicted by the method proposed in the present invention and the actual gestational age is very small.

[0060] Table 2. Comparison of experimental results of the gestational age prediction method in Example 2 and other methods

[0061] method Mean gestational age error R2 correlation coefficient Accuracy TSAN 1.126 0.7147 0.62 VGG-reg 1.022 0.7158 0.67 ResNet-reg 1.005 0.7822 0.66 SENet-reg 0.993 0.7553 0.70 DLVGG-reg 0.831 0.8608 0.79 The method proposed by the present invention 0.812 0.933 0.85

[0062] Example 3

[0063] like Figure 6As shown, this embodiment provides a fetal age prediction device based on magnetic resonance imaging and prior information. The device includes a data set preprocessing unit 601, a brain area distribution feature extraction unit 602, a brain structure feature extraction unit 603, and a prediction unit 604. The specific functions of each unit are as follows:

[0064] The data set preprocessing unit 601 obtains a fetal MRI image data set with known gestational age results and preprocesses the data set, wherein the preprocessing extracts brain regions from the MRI images by positioning and segmenting;

[0065] The brain area distribution feature extraction unit 602 calculates the discrete area distribution vector of the fetal brain composed of the brain area of each MRI image layer based on the brain area extracted by the data set preprocessing unit, and extracts the features of the discrete area distribution vector using a multi-layer perceptron to obtain the brain area distribution feature of the MRI image. The brain area distribution feature is expressed as

[0066] The brain structure feature extraction unit 603 generates a MRI image containing only the brain region based on the brain region extracted by the data set preprocessing unit. The MRI image containing only the brain region is subjected to feature extraction by a three-dimensional fetal convolutional network to obtain the brain structure feature of the MRI image. The brain structure feature is expressed as

[0067] The three-dimensional fetal convolutional network includes a first, second, and third fetal feature extraction modules, a global pooling layer, and a fully connected layer connected in sequence. The first, second, and third fetal feature extraction modules have exactly the same structure. Each fetal feature extraction module includes two asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer connected in sequence. The input of the first fetal feature extraction module is a magnetic resonance image containing only the brain region. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and serves as the input of the second fetal feature extraction module. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and the output of the second fetal feature extraction module and serves as the input of the third fetal feature extraction module.

[0068] The prediction unit 604 fuses the discrete area distribution vector features and the nuclear magnetic resonance image features through a fusion regression module and regresses to output a predicted value of the gestational age.

[0069] The specific implementation of each unit in this embodiment can be found in the above-mentioned embodiment 1, and will not be described one by one here; it should be noted that the device provided in this embodiment is only illustrated by the division of the above-mentioned functional units. In actual applications, the above-mentioned functions can be assigned to different functional units as needed, that is, the internal structure can be divided into different functional units to complete all or part of the functions described above.

[0070] Example 4

[0071] This embodiment provides a computer device, which can be a computer, such as Figure 7 As shown, a processor 702, a memory, an input device 703, a display 704, and a network interface 705 are connected via a system bus 701. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium 706 and an internal memory 707. The non-volatile storage medium 706 stores an operating system, a computer program, and a database. The internal memory 707 provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. When the processor 702 executes the computer program stored in the memory, a method for predicting fetal age based on nuclear magnetic resonance images and prior information proposed in the above-mentioned embodiment 1 is implemented as follows:

[0072] S1. Obtain a fetal MRI image dataset with known gestational age results, and preprocess the dataset, wherein the preprocessing is to extract brain regions from the MRI images by localization and segmentation;

[0073] S2. Based on the brain regions extracted in step S1, the discrete area distribution vector of the fetal brain is calculated based on the brain area of each MRI image layer. The discrete area distribution vector is extracted with the help of a multi-layer perceptron to obtain the brain area distribution feature of the MRI image. The brain area distribution feature is expressed as

[0074] S3, based on the brain region extracted in step S1, a nuclear magnetic resonance image containing only the brain region is generated, and the nuclear magnetic resonance image containing only the brain region is subjected to feature extraction by means of a three-dimensional fetal convolutional network to obtain brain structural features of the nuclear magnetic resonance image. The brain structural features are expressed as

[0075] The three-dimensional fetal convolutional network includes a first, second, and third fetal feature extraction modules, a global pooling layer, and a fully connected layer connected in sequence. The first, second, and third fetal feature extraction modules have exactly the same structure. Each fetal feature extraction module includes two asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer connected in sequence. The input of the first fetal feature extraction module is a magnetic resonance image containing only the brain region. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and serves as the input of the second fetal feature extraction module. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and the output of the second fetal feature extraction module and serves as the input of the third fetal feature extraction module.

[0076] S4. Fusion of the discrete area distribution vector features and the nuclear magnetic resonance image features by a fusion regression module, and regression output of a predicted value of gestational age.

[0077] Example 5

[0078] This embodiment provides a storage medium, which is a computer-readable storage medium and stores a computer program. When the computer program is executed by a processor, the method for predicting fetal age based on magnetic resonance imaging and prior information implemented in the above embodiment 1 is implemented as follows:

[0079] S1. Obtain a fetal MRI image dataset with known gestational age results, and preprocess the dataset, wherein the preprocessing is to extract brain regions from the MRI images by localization and segmentation;

[0080] S2. Based on the brain regions extracted in step S1, the discrete area distribution vector of the fetal brain is calculated based on the brain area of each MRI image layer. The discrete area distribution vector is extracted with the help of a multi-layer perceptron to obtain the brain area distribution feature of the MRI image. The brain area distribution feature is expressed as

[0081] S3, based on the brain region extracted in step S1, a nuclear magnetic resonance image containing only the brain region is generated, and the nuclear magnetic resonance image containing only the brain region is subjected to feature extraction by means of a three-dimensional fetal convolutional network to obtain brain structural features of the nuclear magnetic resonance image. The brain structural features are expressed as

[0082] The three-dimensional fetal convolutional network includes a first, second, and third fetal feature extraction modules, a global pooling layer, and a fully connected layer connected in sequence. The first, second, and third fetal feature extraction modules have exactly the same structure. Each fetal feature extraction module includes two asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer connected in sequence. The input of the first fetal feature extraction module is a magnetic resonance image containing only the brain region. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and serves as the input of the second fetal feature extraction module. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and the output of the second fetal feature extraction module and serves as the input of the third fetal feature extraction module.

[0083] S4. Fusion of the discrete area distribution vector features and the nuclear magnetic resonance image features by a fusion regression module, and regression output of a predicted value of gestational age.

[0084] The storage medium described in this embodiment can be a magnetic disk, an optical disk, a computer memory, a random access memory (RAM), a USB flash drive, a mobile hard disk, or other media.

[0085] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for predicting gestational age based on fetal magnetic resonance images and prior information, characterized in that: The method for predicting fetal gestational age comprises the following steps: S1. Obtain a fetal MRI image dataset with known gestational age results, and preprocess the dataset, wherein the preprocessing is to extract brain regions from the MRI images by localization and segmentation; S2. Based on the brain regions extracted in step S1, the discrete area distribution vector of the fetal brain is calculated based on the brain area of each MRI image layer. The discrete area distribution vector is extracted with the help of a multi-layer perceptron to obtain the brain area distribution feature of the MRI image. The brain area distribution feature is expressed as S3, based on the brain region extracted in step S1, a nuclear magnetic resonance image containing only the brain region is generated, and the nuclear magnetic resonance image containing only the brain region is subjected to feature extraction by means of a three-dimensional fetal convolutional network to obtain brain structural features of the nuclear magnetic resonance image. The brain structural features are expressed as The three-dimensional fetal convolutional network includes a first, second, and third fetal feature extraction modules, a global pooling layer, and a fully connected layer connected in sequence. The first, second, and third fetal feature extraction modules have exactly the same structure. Each fetal feature extraction module includes two asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer connected in sequence. The input of the first fetal feature extraction module is a magnetic resonance image containing only the brain region. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and serves as the input of the second fetal feature extraction module. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and the output of the second fetal feature extraction module and serves as the input of the third fetal feature extraction module. S4. A fusion regression module is used to fuse the discrete area distribution vector features and the nuclear magnetic resonance image features and regress and output a predicted value of gestational age. The processing process of the fusion regression module is as follows: The fusion regression module includes a feature splicing layer and a first and a second fully connected layer connected in sequence. The feature splicing layer combines the brain area distribution features of the magnetic resonance image and brain structural characteristics Spliced into a multimodal feature vector The first and second fully connected layers extract features from the multimodal feature vector Y and regress and output the predicted value of fetal brain age, which is expressed as: in is the connection parameter matrix of the first fully connected layer, r uv represents the element in row u and column v in R, W = [w1,w2,w3,w4] is the connection parameter matrix of the second fully connected layer, w v Represents the vth vector in W, g() is the activation function, and the activation function is selected from one of the ReLU function, sigmoid function and tanh function.

2. The method for predicting gestational age based on fetal magnetic resonance images and prior information according to claim 1, characterized in that: In step S1, positioning refers to locating the brain position of each layer using a box, and segmentation refers to extracting the set of pixel points belonging to the brain area in the box. Positioning and segmentation are completed by manual marking by a doctor or by positioning and segmentation software.

3. The method for predicting gestational age based on fetal magnetic resonance images and prior information according to claim 1, wherein: The process of step S2 is as follows: S21. Assume that the MRI image contains n slices. The brain area calculated from the n slices is combined into an n×1 discrete area distribution vector of the fetal brain S = [s 1, ...,s i, ...,s n ], the calculation formula of the brain area of the i-th slice is expressed as where N i is the number of pixels occupied by the brain area, L×W×H mm 3 is the voxel size of the MRI image, where L, W, and H are the length, width, and height of the voxel, respectively; S22. After calculating the discrete area distribution vector of the fetus, a multilayer perceptron is used to extract features of the discrete area distribution vector. The multilayer perceptron has a two-layer structure, the first layer is a hidden layer, and the input of the hidden layer is the discrete area distribution vector S = [s1, ..., s i ,...,s n ], the output is C=[c1,...,c j ,...,c 12 ], is the connection parameter matrix of the hidden layer, p ij represents the element in row i and column j in P, g() is the activation function, and the activation function is selected from one of the ReLU function, sigmoid function and tanh function; the second layer is the output layer, the input of the output layer is C, and the output is the brain area distribution feature of the magnetic resonance imaging. in is the connection parameter matrix of the output layer, q jk Represents the j-th row and k-th column in Q.

4. The method for predicting gestational age based on fetal magnetic resonance images and prior information according to claim 1, wherein: The input of the first fetal feature extraction module in the three-dimensional fetal convolutional network is a nuclear magnetic resonance image containing only the brain area. The input image is output after passing through two sequentially connected asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer in the first fetal feature extraction module. The asymmetric convolutional layer decomposes the standard 3×3 convolution into 1×3 convolution and 3×1 convolution. The attention enhancement layer uses Squeeze-and Excitation Block. Squeeze-and Excitation Block filters and emphasizes important information from the input global information and outputs it. The step size of the maximum pooling layer is set to 2. After passing through the maximum pooling layer, the length and width of the image will be reduced by half. The length, width, and height of the nuclear magnetic resonance image containing only the brain area are L 1 、W 1 、H 1 The output of the first fetal feature extraction module is the length, width, and height corresponding to the size L 2 、W 2 、H 2 images; The MRI image containing only the brain region and the output of the first fetal feature extraction module are jump-joined as the input of the second fetal feature extraction module. The structure of the second fetal feature extraction module is the same as that of the first fetal feature extraction module. The MRI image containing only the brain region, the output of the first fetal feature extraction module, and the output of the second fetal feature extraction module are jump-joined as the input of the third fetal feature extraction module. The structure of the third fetal feature extraction module is the same as that of the first fetal feature extraction module. Among them, jump stitching is to stitch the high dimension of the image. During stitching, since the length and width of the image are different, the length and width need to be transformed before stitching. The size conversion operations are pooling, upsampling and transposed convolution.

5. A gestational age prediction device based on the gestational age prediction method based on fetal magnetic resonance images and prior information according to any one of claims 1 to 4, characterized in that: The gestational age prediction device comprises: a data set preprocessing unit, which obtains a fetal magnetic resonance image data set with known gestational age results and preprocesses the data set, wherein the preprocessing extracts brain regions from the magnetic resonance image by positioning and segmenting; The brain area distribution feature extraction unit calculates the discrete area distribution vector of the fetal brain composed of the brain area of each MRI image layer based on the brain area extracted by the data set preprocessing unit. The multi-layer perceptron is used to extract the features of the discrete area distribution vector to obtain the brain area distribution feature of the MRI image. The brain area distribution feature is expressed as The brain structure feature extraction unit generates a MRI image containing only the brain area based on the brain area extracted by the data set preprocessing unit. The MRI image containing only the brain area is extracted with the help of a three-dimensional fetal convolutional network to obtain the brain structure features of the MRI image. The brain structure features are expressed as The three-dimensional fetal convolutional network includes a first, second, and third fetal feature extraction modules, a global pooling layer, and a fully connected layer connected in sequence. The first, second, and third fetal feature extraction modules have exactly the same structure. Each fetal feature extraction module includes two asymmetric convolutional layers, an attention enhancement layer, and a maximum pooling layer connected in sequence. The input of the first fetal feature extraction module is a magnetic resonance image containing only the brain region. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and serves as the input of the second fetal feature extraction module. The magnetic resonance image containing only the brain region is jump-concatenated with the output of the first fetal feature extraction module and the output of the second fetal feature extraction module and serves as the input of the third fetal feature extraction module. The prediction unit fuses the discrete area distribution vector features and the nuclear magnetic resonance image features through a fusion regression module and regresses to output a predicted value of the gestational age.

6. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, the gestational age prediction method based on fetal magnetic resonance images and prior information as described in any one of claims 1 to 4 is implemented.

7. A storage medium storing a program, characterized in that: When the program is executed by a processor, the method for predicting gestational age based on fetal magnetic resonance images and prior information according to any one of claims 1 to 4 is implemented.

Images