A fetal brain age estimation method and device based on deep imbalance regression

Abstract

The present application relates to a kind of fetal brain brain age estimation method and device based on deep imbalance regression, the method includes: introducing label to fetal brain magnetic resonance image training set, using label distribution smoothing strategy to obtain effective label density distribution;Establish multi-scale hierarchical segmentation fetal brain feature extraction regression network;Regression focus mean square error loss function is constructed, and effective label density is used to reweighting regression focus mean square error loss function;Grade ordering similarity regularizer is constructed;With the reweighted focus mean square error regression loss function as main function, grade ordering similarity regularizer is used as balance regularization term, and the total loss function of fetal brain feature extraction regression network is constructed.The method can significantly improve the performance of regression model, can let model better learn the continuous information of regression age label, obtain lower mean absolute error, help to identify the abnormality of brain development and reduce the risk of adverse development.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of brain magnetic resonance image processing, and in particular to a fetal brain age estimation method and device based on deep imbalance regression. BACKGROUND

[0002] Fetal brain age estimation has been widely used to characterize normal brain development. The gap between the estimated age and the true age provides an index of deviation from the normal developmental trajectory, which can indicate neurodevelopmental disorders. Therefore, accurately estimating fetal age can provide important help for prenatal care throughout pregnancy and help identify complications. Fetal brain magnetic resonance imaging can provide detailed imaging of the developing fetus in utero, allowing detailed assessment of the fetal brain, including the shape, depth, and timing of the appearance of sulcal gyri between different gestational ages. Therefore, fetal brain magnetic resonance imaging can be used as a supplement to ultrasound examination, both to confirm the results of ultrasound examination and to obtain additional supplementary information.

[0003] In recent years, deep learning methods have achieved good results in predicting fetal brain age. Researchers have applied deep residual networks based on attention mechanisms to automatically locate fetal brain regions, and combined views from different sections to achieve high prediction accuracy and low mean absolute error. For example, the Chinese patent with the authorization publication number CN111415361B discloses a fetal brain age estimation and abnormality detection method and device based on deep learning. In the brain age estimation and abnormality detection method, first, a data set of normal fetal brain T2-weighted magnetic resonance images is established using T2-weighted images of pregnant women in utero collected by clinical routine. Second, a U-shaped network is used to segment the fetal brain from the uterus, and a deep residual network based on attention mechanisms is used to predict fetal brain age and generate uncertainty of brain age and confidence of fetal brain age estimation. Finally, a classifier is constructed according to the difference between actual gestational age and predicted brain age, uncertainty, confidence, and other indicators to determine whether the fetal brain development is abnormal.

[0004] However, the problem of deep imbalance regression in deep learning models has not been adequately addressed and effectively solved. In the task of fetal brain age estimation, the fetal brain magnetic resonance image data set often presents an imbalanced label distribution between different gestational ages, which leads to poor bias in the learned regression model for most labels. Therefore, the performance of the deep learning model will be significantly reduced on the imbalanced data set, resulting in inaccurate estimated fetal brain age. SUMMARY

[0005] The application aims to provide a fetal brain age estimation method and device based on deep imbalance regression, so as to solve the problem of inaccurate brain age estimation caused by the influence of deep imbalance regression on the performance of a deep learning model.

[0006] To solve the above technical problems, the technical scheme provided by the application is as follows:

[0007] The application relates to a fetal brain age estimation method based on deep imbalance regression.

[0008] S1. An original fetal brain magnetic resonance image dataset is acquired, and the original fetal brain magnetic resonance image is preprocessed through histogram equalization to obtain a fetal brain magnetic resonance image training set;

[0009] S2. Labels are introduced into the fetal brain magnetic resonance image training set, and an effective label density distribution is obtained by using a label distribution smoothing strategy;

[0010] S3. A multi-scale hierarchical segmentation fetal brain feature extraction regression network is established;

[0011] S4. A regression focus mean square error loss function is constructed, and the regression focus mean square error loss function is reweighted by using the effective label density;

[0012] S5. A rank order similarity regularizer is constructed;

[0013] S6. The reweighted focus mean square error regression loss function is used as a main function, and the rank order similarity regularizer is used as a balance regularization term, so as to construct a total loss function of the fetal brain feature extraction regression network;

[0014] S7. The fetal brain magnetic resonance image training set is substituted into the multi-scale hierarchical segmentation fetal brain feature extraction regression network for iterative training, the loss value of the total loss function is continuously updated, the best weight parameter is obtained and saved;

[0015] S8. The trained multi-scale hierarchical segmentation fetal brain feature extraction regression network is used to estimate the fetal brain age.

[0016] Preferably, the preprocessed fetal brain magnetic resonance image is further subjected to enhancement processing in step S1, and the enhancement processing includes rotation, translation, scaling, flipping and padding.

[0017] Preferably, the specific steps of step S1 include:

[0018] S1.1. The ROI region of a healthy fetal brain T2 weighted brain magnetic resonance image is cut and saved to obtain an original fetal brain magnetic resonance image;

[0019] S1.2. Preprocess the original fetal brain magnetic resonance images by histogram equalization;

[0020] S1.3. Rotate, scale, horizontally flip, and vertically flip the preprocessed fetal brain magnetic resonance images, and pad them with zeros according to the length of the longest side of each fetal brain magnetic resonance image to make all the fetal brain magnetic resonance images have the same height and width, and complete the enhancement processing of the fetal brain magnetic resonance images.

[0021] Preferably, the step S2 introduces labels into the fetal brain magnetic resonance image training set, and the specific steps of obtaining the effective label density distribution using the label distribution smoothing strategy include:

[0022] S2.1. Introduce labels into the fetal brain magnetic resonance image training set, and the expression is:

[0023]

[0024] wherein D represents the training set, i represents the i-th fetal brain magnetic resonance image, N represents the total number of fetal brain magnetic resonance images, x i represents the fetal brain magnetic resonance image input into the network, y i represents the continuous true fetal brain age label, R represents the set of true fetal brain age labels, R d represents the set of fetal brain magnetic resonance images input into the network;

[0025] S2.2. Discretize the continuous label space composed of fetal brain age labels into M boxes with equal interval Δm, form a box target space Ω, and obtain the original empirical label density:

[0026]

[0027] wherein w j represents the index of the j-th box, p(w j ) is the empirical label density of the box index w j , that is, the frequency of the label w j in the training set;

[0028] S2.3. Convolve the original empirical label density distribution using a Gaussian kernel function to obtain the effective label density distribution:

[0029]

[0030] In the formula, represents the effective label density of the smoothed label w j ′ j j ​​′ ) is a symmetric kernel function; the symmetric kernel function satisfies:

[0031] k(w j ,w j ′ )=k(w j ′ ,w j )

[0032]

[0033] k(w j ′ ,w j ) represents the similarity between two vectors w j ′ and w j , represents the gradient of the symmetric kernel function at w j , represents the gradient of the symmetric kernel function at w j ′ .

[0034] Preferably, the multi-scale hierarchical segmentation fetal brain feature extraction regression network established in S3 is: replacing the original 3x3 convolution in the ResNet50 network with a hierarchical segmentation fetal brain gyrus feature extraction module to form a feature extraction regression network.

[0035] Preferably, the specific way of reweighting the regression focus mean square error loss function with effective label density in step S4 is to multiply the inverse of the effective label density as a weight into the regression focus mean square error loss function, and the expression of the reweighted regression focus mean square error loss function is:

[0036]

[0037] wherein, represents the reweighted regression focus mean square error loss function, e i is the mean absolute error of the i-th fetal brain magnetic resonance image, β and γ are hyperparameters, and the scaling factor is a Sigmoid function that maps the absolute error to the range of [0, 1], represents the effective label density of the label w i ′ , and n represents the number of training samples loaded each time.

[0038] Preferably, the expression of the rank ordering similarity regularizer constructed in step S5 is:

[0039]

[0040] Rank() represents rank operation and forms rank Rank(t), the rank of the bth element is 1 plus the number of elements in this set of sequences that are greater than it, which can be expressed as:

[0041] Rank(t) b = 1 + |{k: t k > t b}|,

[0042] k is the number of elements greater than the element to be sorted t b , t k is the kth element, t b is the bth element;

[0043] For each subset of data, the expression is:

[0044]

[0045] x b represents the input fetal brain magnetic resonance image into the network, y b represents the label of the continuous real fetal brain age, b represents the bth fetal brain magnetic resonance image, T represents the number of subsets;

[0046] μ y represents the negative absolute distance, the similarity function is constructed to operate on the similarity matrix, the negative absolute distance μ y The pair-wise similarity matrix constructed by the similarity function is The expression is:

[0047]

[0048] μ y represents the negative absolute distance, y b represents the bth label, y k represents the kth label;

[0049] μ z represents the cosine similarity, the similarity function calculates the operation of the pair-wise similarity matrix, the cosine similarity μ z The pair-wise similarity matrix calculated by the similarity function is The expression is:

[0050]

[0051] μ z represents the cosine similarity, g() represents the feature extractor, z b represents the label xb corresponding feature, i.e. b ), z k denotes the label x k corresponding feature, i.e. k ), x b denotes the b-th label, x k denotes the k-th label.

[0052] denotes the cosine distance, [b,:] is the b-th row of all columns of the similarity matrix.

[0053] Preferably, the ranking operation of the ranking similarity regularizer is differentiable, and is a linear combination of the minimizer of the objective, so that the ranking similarity regularizer meets the requirement of updating the parameters in the backward propagation, i.e.

[0054] Rank(t)=argmint·π

[0055] t represents a continuous input vector, t∈R, R represents a set of continuous input vectors, π∈Π s , Π s is a set denoted as {1, 2,..., s}, and s represents the s-th discrete value.

[0056] Preferably, the total loss function of the fetal brain feature extraction regression network is The expression of the total loss function is as follows:

[0057]

[0058] wherein η is a weight to balance the total loss.

[0059] The application also relates to a device for implementing the fetal brain age estimation method based on deep imbalance regression.

[0060] The acquisition preprocessing module is configured to obtain an original fetal brain magnetic resonance image dataset, and preprocess the original fetal brain magnetic resonance image through histogram equalization to obtain a fetal brain magnetic resonance image training set.

[0061] The label distribution smoothing module is configured to introduce labels into the fetal brain magnetic resonance image training set, and obtain an effective label density distribution by using a label distribution smoothing strategy.

[0062] The regression network construction module is configured to establish a multi-scale hierarchical segmentation fetal brain feature extraction regression network.

[0063] The regression focus mean square error loss function construction module is configured to construct a regression focus mean square error loss function, and reweight the regression focus mean square error loss function by using the effective label density.

[0064] a rank-order similarity regularizer construction module configured to construct a rank-order similarity regularizer;

[0065] a total loss function calculation module configured to construct a total loss function of the fetal brain feature extraction regression network by taking the reweighted focal mean square error regression loss function as a main function and the rank-order similarity regularizer as a balance regularization term;

[0066] a training module configured to put a fetal brain magnetic resonance image training set into the multi-scale hierarchical segmentation fetal brain feature extraction regression network for iterative training, constantly update a loss value of the total loss function, obtain optimal weight parameters and save the same;

[0067] an estimation module configured to estimate a fetal brain brain age based on the trained multi-scale hierarchical segmentation fetal brain feature extraction regression network.

[0068] Compared with the prior art, the technical scheme provided by the present application has the following beneficial effects:

[0069] The fetal brain brain age estimation method based on deep imbalance regression provided by the present application uses a label distribution smoothing strategy to obtain an effective label density distribution, and then reweights a regression focal mean square error loss function, thereby improving the performance of the regression model. The rank-order similarity regularizer constructed can encode a global inductive bias, so that samples more similar in the label space are also similar in the corresponding feature space, and the model can better learn the continuous information of the regression age label. In summary, the algorithm can obtain a lower mean absolute error, which is helpful for identifying fetal brain development abnormalities and reducing the risk of adverse development in clinical practice. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 is a flowchart of the fetal brain brain age estimation method based on deep imbalance regression provided by the present application;

[0071] Figure 2 is a multi-scale fetal brain sulcus development feature extraction regression network of an embodiment of the present application;

[0072] Figure 3 is a mean absolute error (MAE) and regression fitting degree (R 2 ) result graph of a test data set of the present application. DETAILED DESCRIPTION

[0073] In order to further understand the content of the present application, the present application will be described in detail in conjunction with the embodiments, and the following embodiments are used to illustrate the present application but not to limit the scope of the present application.

[0074] Embodiment 1

[0075] Referring to the drawings Figure 1 The present application relates to a fetal brain age estimation method based on deep unbalanced regression, which comprises the following steps:

[0076] S1. Introduce healthy fetal brain T2-weighted brain magnetic resonance image data sets of different gestational ages (from 154 days to 244 days), and pre-process the original fetal brain magnetic resonance images by histogram equalization to make each fetal brain magnetic resonance image have a uniform gray probability density distribution, so that the image becomes clearer and has high contrast, to obtain a fetal brain magnetic resonance image training set. The pre-processed fetal brain magnetic resonance images are enhanced using data enhancement technology, and the enhancement processing includes rotation (0-360°), translation (0.2 pixels), scaling (96x96 pixels), flipping and padding processing, and the specific steps are:

[0077] S1.1. Cut and save the ROI region of the healthy fetal brain T2-weighted brain magnetic resonance image to obtain the original fetal brain magnetic resonance image;

[0078] S1.2. Pre-process the original fetal brain magnetic resonance image by histogram equalization;

[0079] S1.3. Rotate, translate, scale, horizontally flip and vertically flip the pre-processed fetal brain magnetic resonance image, and pad with zeros according to the length of the longest side of each fetal brain magnetic resonance image to make all fetal brain magnetic resonance images have equal height and width, and complete the enhancement processing of the fetal brain magnetic resonance image.

[0080] S2. Introduce labels into the fetal brain magnetic resonance image training set, and use a label distribution smoothing strategy to obtain an effective label density distribution, and the specific steps include:

[0081] S2.1. Introduce labels into the fetal brain magnetic resonance image training set, and the expression is:

[0082]

[0083] Where D represents the training set, i represents the i-th fetal brain magnetic resonance image, N represents the total number of fetal brain magnetic resonance images, x i represents the fetal brain magnetic resonance image input into the network, y i represents the continuous true fetal brain age label, R represents the set of true fetal brain age labels, R d represents the set of fetal brain magnetic resonance images input into the network;

[0084] S2.2. Discretize the continuous label space composed of fetal brain age labels into M boxes with equal interval Am, forming a box target space Ω, and obtain the original empirical label density:

[0085]

[0086] where w j represents the index of the jth box, p(w j ) is the empirical label density of the box index w j , that is, the frequency of the label w j appearing in the training set;

[0087] S2.3. Convolve the original empirical label density distribution with a Gaussian kernel function to obtain the effective label density distribution:

[0088]

[0089] In the formula, represents the effective label density of the smoothed label w j ′ , k(w j , w j ′ is a symmetric kernel function; the symmetric kernel function satisfies:

[0090] k(w j , w j ′ ) = k(w j ′ , w j )

[0091]

[0092] k(w j ′ , w j ) represents the similarity between two vectors w j ′ and w j , represents the gradient of the symmetric kernel function at w j , represents the gradient of the symmetric kernel function at w j ′ .

[0093] S3. Establish a multi-scale hierarchical segmentation fetal brain feature extraction regression network, that is, apply hierarchical segmentation and splicing operations to the standard residual structure 3×3 bottleneck layer: Figure 2(a) shows a bottleneck layer of Residual Network (ResNet), which is replaced by the hierarchical segmentation fetal brain sulci feature extraction module in the original 3x3 convolution in ResNet50 network, Figure 2 (b) shows the hierarchical segmentation fetal brain sulci feature extraction module to replace Figure 2 the original 3x3 convolution in (a); when the output feature map has a smaller receptive field, it can capture the development characteristics of fetal brain sulci; when the output feature map has a larger receptive field, it can obtain global information;

[0094] The hierarchical segmentation fetal brain sulci feature extraction module described above obtains multi-scale feature representation by using hierarchical segmentation and splicing operations: the feature map is divided into q groups and c channels, the first group is directly spliced to the next layer, the second group is convolved with a 3x3 filter, and then divided into two subgroups, one of which is directly spliced to the next layer, and the other is spliced with the incoming feature map; then, the spliced feature map is convolved with a 3x3 filter, and this process is repeated several times until all feature maps are operated as described above; finally, all incoming feature maps are spliced and convolved with a 1x1 filter, thereby reconstructing a new feature map.

[0095] S4. Construct a regression focal loss function, and reweight the regression focal loss function with effective label density, the expression of the reweighted regression focal loss function is:

[0096]

[0097] wherein, represents the reweighted regression focal loss function, e i is the mean absolute error of the i-th fetal brain magnetic resonance image, β and γ are hyperparameters, and the scaling factor is a Sigmoid function that maps the absolute error to the range of [0, 1], represents the effective label density of the label w i ′ n represents the number of training samples loaded each time. The reweighted focal loss function can better focus on difficult samples.

[0098] S5. Construct a rank order similarity regularizer, encode a global bias, and add it to the main function as a regularization term, so that samples with closer distances in the label space are also more similar in the feature space, and the relationship between close samples and distant samples can be captured at the same time, the rank order similarity regularizer is defined as:

[0099]

[0100] Rank() represents the rank operation and forms the ranks Rank(t), the rank of the b-th element is one plus the number of elements in this set of sequences that are larger than it, which can be expressed as:

[0101] Rank(t) b = 1 + |{k : t k > t b},

[0102] k is the number of elements larger than the element to be sorted t b , t k is the k-th element, t b is the b-th element;

[0103] For each subset in the batch, the expression is:

[0104]

[0105] x b represents the input fetal brain magnetic resonance images into the network, y b represents the label of the continuous true fetal brain age, b represents the b-th fetal brain magnetic resonance image, T represents the number of subsets;

[0106] represents the negative absolute distance μ y The similarity function is constructed to operate on the similarity matrix, the negative absolute distance μ y The pair-wise similarity matrix constructed by the similarity function is The expression is:

[0107]

[0108] μ y represents the negative absolute distance, y b represents the b-th label, y k represents the k-th label;

[0109] represents the cosine similarity μ z The similarity function is constructed to operate on the pair-wise similarity matrix, the cosine similarity μ z The pair-wise similarity matrix calculated by the similarity function is The expression is:

[0110]

[0111] μ z represents the cosine similarity, g() represents the feature extractor, zb denotes label x b corresponding feature, i.e. g(x b ), z k denotes label x k corresponding feature, i.e. g(x k ), x b denotes the b-th label, x k denotes the k-th label.

[0112] denotes cosine distance, [b, : ] is the b-th row of all columns of the similarity matrix.

[0113] The rank-wise similarity regularizer samples from each batch of data, ensuring that each sample appears at most once to reduce the connection between samples and improve the relative representation of infrequent labels.

[0114] In order to make the rank-wise similarity regularizer meet the requirements of updating parameters by back propagation, the rank operation of the rank-wise similarity regularizer needs to be differentiable, and it is a minimizer of a linear combination objective, i.e.

[0115] Rank(t) = arg min t·π

[0116] t represents a continuous input vector, t ∈ R, R represents a set of continuous input vectors, π ∈ Π s , Π s is a set denoted as {1, 2,..., s}, and s represents the s-th discrete value.

[0117] Then, a black box combination solver is used to perform back propagation and return a continuous interpolation gradient to update the parameters:

[0118]

[0119]

[0120] wherein, denotes the information of the input gradient, and λ represents the strength of interpolation.

[0121] In this way, the rank-wise similarity regularizer can be used, and can be added as a regularization term to the total loss function.

[0122] S6. The total loss function of the fetal brain feature extraction regression network is constructed by taking the reweighted focal mean square error regression loss function as the main function and the rank-wise similarity regularizer as the balance regularization term, and the expression of the total loss function is:

[0123]

[0124] wherein η is a weight to balance the total loss.

[0125] S7. The fetal brain magnetic resonance image training set is substituted into the multi-scale hierarchical segmentation fetal brain feature extraction regression network for iterative training, the loss value of the total loss function is constantly updated, the best weight parameter is obtained and saved, and the average absolute error and fitting degree of the final fetal brain magnetic resonance image regression prediction are calculated.

[0126] S8. The fetal brain brain age is estimated based on the trained multi-scale hierarchical segmentation fetal brain feature extraction regression network.

[0127] Experimental example

[0128] In this experimental example, 10 groups of random experiments are performed on the fetal brain magnetic resonance image as an example to test the effectiveness of the method, and the corresponding mean and standard deviation are calculated, as shown in Table 1.

[0129] Table 1: Statistical table of mean and standard deviation calculated by each method

[0130]

[0131] Secondly, this experimental example also specifically demonstrates the test results (mean absolute error (MAE) and regression fitting degree (R 2 )) of the present application by taking the fetal brain magnetic resonance image as an example, as shown in Figure 3 .

[0132] As shown in Table 1, compared with some popular regression models (ResNet, VGG, DenseNet, SE-Net and ConvNeXt), the hierarchical split residual regression network obtains the lowest mean absolute error of 0.925 weeks and the regression fitness of 0.871. By using the label distribution smoothing strategy and the rank similarity regularization technique, and using the hierarchical split residual regression network, our method obtains a very promising result: the mean absolute error of 0.760±0.066 weeks (5.323±0.464 days) and the regression fitness of 0.914±0.020, with a reduction of 0.165 weeks in mean absolute error, where the statistics are (mean ± standard deviation) of 10 groups of random experiments. These results prove that the label distribution smoothing strategy and the rank similarity regularization technique can effectively improve the performance of the hierarchical split residual regression network model on imbalanced data sets. In addition, compared with the method of fetal brain age estimation proposed by Liao et al., the mean absolute error of our deep imbalanced regression algorithm is reduced by 0.143 weeks. And our algorithm is also superior to the imbalanced regression algorithm proposed by Yang et al., with a reduction of 0.425 weeks in mean absolute error. Figure 3 The scatter plot shows the high correlation between the predicted brain age and the true brain age on the test set under different random experiments. The imbalanced regression algorithm we proposed effectively reduces the mean absolute error and achieves excellent performance in the fetal brain age estimation task.

[0133] Embodiment 2

[0134] The embodiment relates to an apparatus for implementing the fetal brain age estimation method based on deep imbalanced regression in embodiment 1, which comprises:

[0135] The acquisition preprocessing module: acquires an original fetal brain magnetic resonance image data set, pre-processes the original fetal brain magnetic resonance image through histogram equalization to obtain a fetal brain magnetic resonance image training set; the acquisition preprocessing module is used to realize the function of step S1 in embodiment 1.

[0136] The label distribution smoothing module is used to introduce labels into the fetal brain magnetic resonance image training set, and obtain an effective label density distribution by using the label distribution smoothing strategy; the label distribution smoothing module is used to realize the function of step S2 in embodiment 1.

[0137] The regression network construction module is used to establish a multi-scale hierarchical split fetal brain feature extraction regression network; the regression network construction module is used to realize the function of step S3 in embodiment 1.

[0138] A regression focus mean square error loss function construction module is configured to construct a regression focus mean square error loss function, and the effective label density is used to reweight the regression focus mean square error loss function.

[0139] A rank order similarity regularizer construction module is configured to construct a rank order similarity regularizer.

[0140] A total loss function calculation module is configured to construct a total loss function of the fetal brain feature extraction regression network by taking the reweighted focus mean square error regression loss function as a main function and taking the rank order similarity regularizer as a balance regularization term.

[0141] A training module is configured to input a fetal brain magnetic resonance image training set into the multi-scale hierarchical segmentation fetal brain feature extraction regression network for iterative training, constantly update a loss value of the total loss function, obtain optimal weight parameters, and save the optimal weight parameters.

[0142] An estimation module is configured to estimate a fetal brain age based on the trained multi-scale hierarchical segmentation fetal brain feature extraction regression network.

[0143] The above detailed description of the present application is based on the preferred embodiments of the present application, and the description should not be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still fall within the scope of the present application.

Claims

1. A deep imbalance regression based method for fetal brain age estimation, characterized in that: It comprises the following steps: S1. Obtain an original fetal brain magnetic resonance image dataset, preprocess the original fetal brain magnetic resonance image through histogram equalization to obtain a fetal brain magnetic resonance image training set; S2. Introduce labels into the fetal brain magnetic resonance image training set, and obtain an effective label density distribution using a label distribution smoothing strategy; S3. Establish a multi-scale hierarchical segmentation fetal brain feature extraction regression network; S4. Construct a regression focus mean square error loss function, and reweight the regression focus mean square error loss function using the effective label density, specifically, the reciprocal of the effective label density is multiplied as a weight into the regression focus mean square error loss function, and the expression of the reweighted regression focus mean square error loss function is: ， wherein, represents a re-weighted regression focus mean squared error loss function, is the mean absolute error for the i th fetal brain magnetic resonance image, and is a hyperparameter, the scaling factor is Sigmoid a function that maps the absolute error to the range [0, 1], represents the effective label density of the labels is the number of training samples loaded at each time, denotes the number of training samples loaded at each time; S5. Construct a rank similarity regularizer; S6. Take the reweighted focus mean square error regression loss function as the main function, and take the rank similarity regularizer as the balance regularization term to construct the total loss function of the fetal brain feature extraction regression network; S7. Substitute the fetal brain magnetic resonance image training set into the multi-scale hierarchical segmentation fetal brain feature extraction regression network for iterative training, constantly update the loss value of the total loss function, obtain the best weight parameter and save it; S8. Estimate the fetal brain age based on the trained multi-scale hierarchical segmentation fetal brain feature extraction regression network.

2. The deep imbalance regression based fetal brain age estimation method of claim 1, wherein: The preprocessed fetal brain magnetic resonance image in step S1 is also subjected to enhancement processing, and the enhancement processing includes rotation, translation, scaling, flipping and padding.

3. The deep imbalance regression based fetal brain age estimation method of claim 2, wherein: The specific steps of step S1 include: S1.

1. Cut and save the ROI region of the healthy fetal brain T2 weighted brain magnetic resonance image to obtain the original fetal brain magnetic resonance image; S1.

2. Preprocess the original fetal brain magnetic resonance image through histogram equalization; S1.

3. Rotate, translate, scale, horizontally flip and vertically flip the preprocessed fetal brain magnetic resonance image, and pad with zero according to the longest side length of each fetal brain magnetic resonance image to make all fetal brain magnetic resonance images equal in height and width, and complete the enhancement processing of the fetal brain magnetic resonance image.

4. The deep imbalance regression based fetal brain age estimation method of claim 1, wherein: The specific steps of step S2 include: S2.

1. Introduce labels into the fetal brain magnetic resonance image training set, and the expression is: ， wherein, D represent a training set, i represent a first i represent a fetal brain magnetic resonance image, N represent a total number of fetal brain magnetic resonance images, represent fetal brain magnetic resonance images input into the network, represent a label of a consecutive true fetal brain age, R represents a set of true fetal brain age labels, represent a set of fetal brain magnetic resonance images input into the network; S2.

2. Discretize the continuous label space consisting of fetal brain age labels into a box target space with equally spaced bins : ， wherein, represents the index of the j box. and obtaining raw experience label density i.e., label frequency of occurrence; S2.

3. Convolve the original empirical label density distribution using a Gaussian kernel function to obtain an effective label density distribution: ， In the formula, represents the smoothed label effective label density, is a symmetric kernel function; the symmetric kernel function satisfies: ， ， representing a similarity between two vectors and , representing a gradient of a symmetric kernel function at , representing a gradient of a symmetric kernel function at .

5. The deep imbalance regression based fetal brain age estimation method of claim 1, wherein: The multi-scale hierarchical segmentation fetal brain feature extraction regression network established in S3 is: replace the original 3x3 convolution in the ResNet50 network with a hierarchical segmentation fetal brain sulcus feature extraction module to form a feature extraction regression network.

6. The deep imbalance regression based fetal brain age estimation method of claim 1, wherein: The expression of the rank similarity regularizer constructed in step S5 is: ， Rank ( ) represents a rank ordering operation and forms a rank Rank(t), the b rank Rank(t) of the i-th element is one plus the number of elements in this set of sequences that are larger than it, and can be represented as: ， for the number of elements to be sorted, larger than the element to be sorted, for the first k element, for the first b element; For each subset in the batch of data, the expression is: ， representing fetal brain magnetic resonance images input into the network, representing labels of successive true fetal brain ages, b representing the first b fetal brain magnetic resonance image, T denotes the number of subsets; denotes negative absolute distance The similarity function is configured to operate on the similarity matrix, negative absolute distance The similarity function is configured to operate on the pair-wise similarity matrix as The expression is: ， Represents negative absolute distance. Indicates the first b A tag, Indicates the first k One tag; cosine similarity The similarity function computes the pair-wise similarity matrix, cosine similarity The pair-wise similarity matrix computed by the similarity function is denoted as The expression is: ， denotes a cosine similarity, denotes a feature extractor, denotes a label corresponding features, i.e. , denotes a label corresponding features, i.e. , denotes a first b label, denotes a first k label; denotes the cosine distance, is the i-th column of the similarity matrix S. b all columns of the i-th row of the similarity matrix S.

7. The deep imbalance regression based fetal brain age estimation method of claim 6, wherein: The ranking operation of the ranking similarity regularizer can be differentiated, and is a linear combination objective minimizer, so that the ranking similarity regularizer meets the requirement of updating parameters in reverse propagation, that is: ， t representing a vector of successive inputs, t ∈ R , R representing a set of vectors of successive inputs, , denoted by a set of s representing the s discrete value.

8. The deep imbalance regression based fetal brain age estimation method of claim 7, wherein: The total loss function of the fetal brain feature extraction regression network The expression is: ， wherein, is a weight to unbalance the total loss.

9. An apparatus for implementing the deep imbalance regression based fetal brain age estimation method of claim 1, characterized by: It comprises: The acquisition preprocessing module: acquires the original fetal brain magnetic resonance image dataset, pre-processes the original fetal brain magnetic resonance image through histogram equalization, and obtains the fetal brain magnetic resonance image training set; The label distribution smoothing module is used to introduce labels into the fetal brain magnetic resonance image training set, and uses a label distribution smoothing strategy to obtain an effective label density distribution; The regression network construction module is used to establish a multi-scale hierarchical segmentation fetal brain feature extraction regression network; The regression focal loss function construction module is used to construct a regression focal loss function, and the effective label density is used to reweight the regression focal loss function, and the specific method is to take the reciprocal of the effective label density as a weight and multiply it into the regression focal loss function, and the expression of the reweighted regression focal loss function is: ， wherein, represents a re-weighted regression focus mean squared error loss function, is the mean absolute error for the i th fetal brain magnetic resonance image, and is a hyperparameter, the scaling factor is Sigmoid a function that maps the absolute error to the range [0, 1], represents the effective label density of the labels , and denotes the number of training samples loaded at each time. The ranking similarity regularizer construction module is used to construct a ranking similarity regularizer; The total loss function calculation module takes the reweighted focal loss regression loss function as the main function, and takes the ranking similarity regularizer as the balance regularization term to construct the total loss function of the fetal brain feature extraction regression network; The training module is used to put the fetal brain magnetic resonance image training set into the multi-scale hierarchical segmentation fetal brain feature extraction regression network for iterative training, constantly update the loss value of the total loss function, obtain the best weight parameter and save it; The estimation module is used to estimate the fetal brain age based on the trained multi-scale hierarchical segmentation fetal brain feature extraction regression network.

Images