CN116740463B - SMRI image sequential multi-classification method - Google Patents

Abstract

The invention relates to a sMRI image sequential multi-classification method, which comprises the following steps: dividing a certain number sMRI of images into a training set, a verification set and a test set; constructing a network model, wherein the constructed network model comprises a feature extraction and classification model, a sorting model and a difficult sample identification correction model; inputting all sample images in the training set into the constructed network model in batches for training, and verifying the performance of the trained network model by using all images in the verification set; after multiple training and verification, screening out an optimal network model; and finally, inputting the to-be-tested image to be tested in the test set into an optimal network model to obtain a classification result of the to-be-tested image. The advantages are that: the distance between feature vectors obtained by the samples of different categories through the model and the distance in actual pathology are kept consistent through the sequencing model, so that the extracted features among sMRI images of different categories have sequential characteristics, and the network classification performance is improved.

Description

A sequential multi-classification method for sMRI images

Technical Field

The invention relates to the technical field of image processing, and in particular to a sequential multi-classification method for sMRI images.

Background Art

Structural Magnetic Resonance Imaging (sMRI) is an important tool for diagnosing brain diseases. It captures brain morphological changes caused by brain diseases in a non-invasive manner, such as shrinkage of the cerebral cortex and hippocampus, and enlargement of the ventricles.

With the continuous development of computer hardware and software, Convolutional Neural Network (CNN) has become a hot topic in the field of image processing and analysis with its excellent feature extraction and feature classification capabilities. In recent years, some CNN-based sMRI image classification methods have outperformed traditional machine learning methods in the diagnosis of brain diseases, such as CNN-based feature extraction and classification of regions of interest, and CNN-based and attention-based sMRI image classification methods. However, the above methods only consider the learning of a single sample, but ignore the relationship between samples. Pair-based methods consider the relationship between sample pairs consisting of positive and negative samples. Such methods can help the model better distinguish samples of different categories by encouraging the model to learn the true relationship between samples by reducing the distance between positive pairs and expanding the distance between negative pairs. Although most pair-based methods perform well in the binary classification task of sMRI images, they still face the problem of low accuracy for multi-classification of images, and are therefore rarely directly applied to the diagnosis of multi-stage brain diseases.

Related studies have shown that the multi-stage sequential development characteristics of the disease are conducive to the accurate extraction of image features at different stages and subsequent classification, and make the method output closer to the real form. The sequentiality of disease development is used to improve the performance of sMRI image multi-classification. However, there are currently few studies on the use of this sequential characteristic constraint in brain disease diagnosis, and the research is not in-depth. Therefore, further improvements to the existing technology are needed.

Summary of the invention

The technical problem to be solved by the present invention is to provide a sMRI image sequential multi-classification method with higher classification accuracy in view of the above-mentioned prior art.

The technical solution adopted by the present invention to solve the above technical problem is: a sMRI image sequential multi-classification method, characterized by comprising the following steps:

Step 1, obtaining a certain number of sMRI images and the labels corresponding to each sMRI image to form a sample set;

Step 2: Divide the sample set into training set, validation set and test set;

Step 3, constructing a network model; the constructed network model includes a feature extraction and classification model, a sorting model and a difficult sample recognition and correction model; wherein the feature extraction and classification model includes a feature extraction module and a classification module connected to the feature extraction module;

In each batch of training, at least one sample image with label c ₁ , at least one sample image with label c ₂ , ... at least one sample image with label c _n is selected, where n is the total number of labels; among them, c ₁ >c ₂ >... >c _n , and > is a partial order relationship;

The specific process of training the network model using any batch of sample images is as follows:

Step 4-1, preprocessing each sample image selected in the current batch to obtain a first image;

Step 4-2, inputting all sample images and all first images selected in the current batch into the feature extraction module at the same time to obtain a feature vector; wherein the feature vector includes the first feature vectors corresponding to all sample images and the second feature vectors corresponding to all first images;

Step 4-3, input the feature vector obtained in step 4-2 into the sorting model, the difficult sample identification and correction model and the classification module respectively, and calculate the first loss function corresponding to the sorting model, the second loss function corresponding to the difficult sample identification and correction model and the third loss function corresponding to the classification module respectively;

The specific calculation process of the first loss function L _rank is:

Step 4-31: Select a sample image with label _c1 , a sample image with label _c2 , ..., and a sample image with label _cn from all sample images selected in the current batch, and form an image set The initial value of k is 1. Corresponding to the sample image with label c ₁ selected for the kth time, Corresponding to the sample image with label c ₂ selected for the kth time, The corresponding sample image is the one with label c _n selected for the kth time;

and get The first image corresponding to The first image corresponding to The first image corresponding to

Calculate the distance relationship matrix S between sMRI images. The calculation formula of S is:

The values in the pth row of S represent and The corresponding label distance relationship metric value, p is an odd number, and takes values of 1, 3, 5, ... n respectively;

The values in the qth row of S represent and The corresponding label distance relationship metric value, q, is an even number and takes values of 2, 4, 6, ...n+1 respectively;

S ₁ , S ₂ , S ₃ …S _n are all the label distance relationship measurement values of any two images, and S ₁ >S ₂ >S ₃ >S _n ;

Step 4-32: Calculate the true distance matrix between sMRI images The calculation formula is:

in The values in row p represent and The true distance between them, p is an odd number, and takes values of 1, 3, 5, ... n respectively; The values in the qth row represent and The real distance between them, q is an even number, and takes the values of 2, 4, 6, ...n+1 respectively;

Right now: for The first feature vector obtained by inputting it into the feature extraction module F _θ ; e∈{1,…n}

Right now: for The second feature vector obtained by inputting into the feature extraction module F _θ ;

Step 4-33: Calculate the sorting loss corresponding to G _k in step 4-31

Among them, ∏(.) is the multiplication symbol, is the indicator function. If _Sa,b = _Sf , then II(Sa _,b = _Sf ) = 1; if Sa _,b ≠ _Sf , then II( _Sa,b = _Sf ) = 0; Sa _,b is the value of the ath row and bth column in S. for S _c,d is the value of the cth row and dth column in S;

Step 4-34, add 1 to the value of k, reconstruct the image set G _k+1 , and follow the same method as in steps 4-31 to 4-33 to obtain

Step 4-35, and so on to construct image sets G _k+2 , …G _g , wherein the image sets G ₁ , G ₂ , …G _g constructed each time are different, g=Nc ₁ ×Nc ₂ × …×Nc _n , Nc ₁ is the number of sample images with label c ₁ selected in the current batch training, Nc ₂ is the number of sample images with label c ₂ selected in the current batch training, …Nc _n is the number of sample images with label c _n selected in the current batch training;

And follow the same method as in step 4-31 to step 4-33 to obtain

Then the calculation formula of the first loss function L _rank is:

Step 4-4: Calculate the total loss function of the network model according to the first loss function, the second loss function and the third loss function, and update the parameters of the network model according to the total loss function, so as to obtain the network model after one training is completed;

Step 5: Input the image to be tested in the test set into the optimal network model to obtain the classification result of the image to be tested.

As an improvement, the specific calculation process of the second loss function L _hcr in step 4-3 is:

Step 4-a, all sample images selected in the current batch and all first images are combined into a sample set X″, X″=X∪X′, and the sample set X″ is input into the feature extraction module F _θ to obtain Z″=F _θ (X″), and the posterior probability Q(z _i ) of each sample z _i in Z″ belonging to its own category is calculated, z _i ∈ Z″, and the calculation formula of Q(z _i ) is as follows:

Where C is the label category set, C = {c ₁ , c ₂ , … c _n }, _Ni is the K nearest neighbor distance of z _i , whose value is calculated by Euclidean distance, K = |N _i |; z _j is the jth sample in Z″, y _j is the label of z _j , and N is the total number of sample images in the current batch;

Step 4-b: Use the cross entropy function to get the probability d _i that the label of z _i is indeed y _i . The calculation formula of d _i is as follows:

d _i = l(Q(z _i ),y _i );

l is the cross entropy function;

Step 4-c: Divide all d _i into n sets according to the sample labels, where the kth set is represented as

D _ck ＝{d _i |i＝1～2N&y _i ＝c _k ,c _k ∈C}

Arrange the elements in D _ck in descending order, and take the first α% of the elements in D _ck as difficult samples, denoted as The remaining 1-α% of the elements in D _ck are taken as simple samples and are recorded as T _ck ; α is an empirical value;

The difficult sample set Simple sample set

Step 4-d: randomly select a simple sample z _a from the simple sample set T as an anchor sample, and A difficult sample _zp with the same label as _za is selected as a positive sample from the set of simple samples T, and a simple sample _zn with a different label from _za is selected as a negative sample to construct a triplet _gr = { _za , _zp , _zn }. The calculation formula of the second loss function _Lhcr is:

in, is a set of triples composed of triples, N _G is the number of triples, is the L2 norm, and Φ is a marginal value.

Furthermore, the calculation process of the third loss function L _cls in step 4-3 is:

Wherein, _xe is any image among all sample images and all first images in the current batch, ye _is the label corresponding to x, and _Hθ ( _Fθ ( _xe )) is the classification result obtained by inputting _Fθ ( _xe ) into the classification module _Hθ .

Preferably, the calculation formula of the total loss function in step 4-4 is:

L _sum =L _cls +λ1*L _rank +λ2*L _hcr ;

Among them, λ1 is a preset first weighting coefficient, and λ2 is a preset second weighting coefficient.

Compared with the prior art, the advantages of the present invention are: through the sorting model, sMRI images with different labels are sequentially distributed in the feature embedding space, so that the distance between the feature vectors of these samples of different categories obtained through the model is consistent with the distance in the actual pathology, so that the features between the extracted sMRI images of different categories have sequential characteristics and improve the network classification performance; in addition, through the difficult sample recognition correction model, the feature representation of a small number of difficult samples that may exist in the training set is made to approach the feature representation of a large number of simple samples, thereby improving the network classification performance. Therefore, the classification method is simple and improves the classification accuracy.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a principle block diagram of a sequential multi-classification method for sMRI images according to an embodiment of the present invention;

FIG2 is a schematic diagram of a sorting model in an embodiment of the present invention;

FIG. 3 is a schematic diagram of a difficult sample recognition and correction model according to an embodiment of the present invention.

DETAILED DESCRIPTION

The present invention is further described in detail below with reference to the accompanying drawings.

As shown in FIGS. 1 to 3 , the sMRI image sequential multi-classification method in this embodiment includes the following steps:

Step 1, obtaining a certain number of sMRI images and the labels corresponding to each sMRI image to form a sample set;

In this embodiment, all sMRI images are preprocessed according to a unified operation. First, the skull stripping and affine registration operations of the original image are implemented using the FMRIB software tool 2. In the affine registration step, the sMRI image is linearly aligned with the MNI152 template to eliminate the global linear difference, and the sMRI image is resampled to a spatial resolution of 1×1×1mm; then the sMRI image is normalized to [0,1] and the background of each sMRI image is removed according to the minimum external stereo frame; finally, the size of all sMRI images is uniformly adjusted to 160×160×160 using the background zero filling method;

In addition, due to the relatively small amount of data, five-fold cross-validation was used in all experiments, and all sMRI images were randomly divided into five equal subsets, four of which were used for training and the remaining one for validation; data not in these five subsets were discarded;

Step 2: Divide the sample set into training set, validation set and test set;

Step 3, constructing a network model; the constructed network model includes a feature extraction and classification model, a sorting model and a difficult sample recognition and correction model; wherein the feature extraction and classification model includes a feature extraction module and a classification module connected to the feature extraction module;

The Plackett-Luce (PL) model is a classic ranking model, as shown in FIG2 . In this embodiment, a sequential distance relationship matrix is first designed to describe the sequential nature of sMRI images in different disease courses, and then a PL-constrained ranking loss is proposed to quantify the difference between the estimated sequential relationship between sMRI images with different labels and the actual sequential relationship;

As shown in FIG3 , in this embodiment, the posterior probability estimation based on KNN and cross entropy loss is used to identify difficult and simple samples, and the method of simple samples is used to induce difficult samples with the same labels to approximate simple samples;

Step 4: Input all sample images in the training set into the network model constructed in step 3 in batches for training, and use all images in the validation set to verify the performance of the trained network model; after multiple training and validation, select the optimal network model;

In each batch of training, at least one sample image with label c ₁ , at least one sample image with label c ₂ , ... at least one sample image with label c _n is selected, where n is the total number of labels; among them, c ₁ >c ₂ >... >c _n , and > is a partial order relationship;

The specific process of training the network model using any batch of sample images is as follows:

Step 4-1, preprocessing each sample image selected in the current batch to obtain a first image;

Step 4-2, inputting all sample images and all first images selected in the current batch into the feature extraction module at the same time to obtain a feature vector; wherein the feature vector includes the first feature vectors corresponding to all sample images and the second feature vectors corresponding to all first images;

Step 4-3, input the feature vector obtained in step 4-2 into the sorting model, the difficult sample identification and correction model and the classification module respectively, and calculate the first loss function corresponding to the sorting model, the second loss function corresponding to the difficult sample identification and correction model and the third loss function corresponding to the classification module respectively;

The specific calculation process of the first loss function L _rand is:

The specific calculation process of the first loss function L _rank is:

Step 4-31: Select a sample image with label _c1 , a sample image with label _c2 , ..., and a sample image with label _cn from all sample images selected in the current batch, and form an image set The initial value of k is 1. Corresponding to the sample image with label c ₁ selected for the kth time, Corresponding to the sample image with label c ₂ selected for the kth time, The corresponding sample image is the one with label c _n selected for the kth time;

and get The first image corresponding to The corresponding first image The corresponding first image

The distance relationship matrix S between the sMRI images (the sMRI images are the sample image and the first image) is calculated. The calculation formula of S is:

The values in the pth row of S represent and The corresponding label distance relationship metric value, p is an odd number, and takes values of 1, 3, 5, ... n respectively;

The values in the qth row of S represent and The corresponding label distance relationship metric value, q, is an even number, and takes values of 2, 4, 6, ...n+1 respectively;

S ₁ , S ₂ , S ₃ …S _n are all the label distance relationship measurement values of any two images, and S ₁ >S ₂ >S ₃ >S _n ;

Step 4-32: Calculate the true distance matrix between sMRI images The calculation formula is:

in The values in row p represent and The true distance between them, p is an odd number, and takes values of 1, 3, 5, ... n respectively; The values in the qth row represent and The real distance between them, q is an even number, and takes the values of 2, 4, 6, ...n+1 respectively;

Right now: for The first feature vector obtained by inputting it into the feature extraction module F _θ ; e∈{1,…n}

Right now: for The second feature vector obtained by inputting into the feature extraction module F _θ ;

Step 4-33: Calculate the sorting loss corresponding to G _k in step 4-31

Among them, ∏(.) is the multiplication symbol, is the indicator function. If _Sa,b = _Sf , then II(Sa _,b = _Sf ) = 1; if Sa _,b ≠ _Sf , then II( _Sa,b = _Sf ) = 0; Sa _,b is the value of the ath row and bth column in S. for S _c,d is the value of the cth row and dth column in S;

Step 4-34, add 1 to the value of k, reconstruct the image set G _k+1 , and follow the same method as in steps 4-31 to 4-33 to obtain

Step 4-35, and so on to construct image sets G _k+2 , …G _g , wherein the image sets G ₁ , G ₂ , …G _g constructed each time are different, g=Nc ₁ ×Nc ₂ × …×Nc _n , Nc ₁ is the number of sample images with label c ₁ selected in the current batch training, Nc ₂ is the number of sample images with label c ₂ selected in the current batch training, …Nc _n is the number of sample images with label c _n selected in the current batch training;

And follow the same method as in step 4-31 to step 4-33 to obtain

Then the calculation formula of the first loss function L _rank is:

The specific calculation process of the second loss function L _hcr is as follows:

Step 4-a, all sample images selected in the current batch and all first images are combined into a sample set X″, X″=X∪X′, and the sample set X″ is input into the feature extraction module F _θ to obtain Z″=F _θ (X″), and the posterior probability Q(z _i ) of each sample z _i in Z″ belonging to its own category is calculated, z _i ∈ Z″, and the calculation formula of Q(z _i ) is as follows:

Wherein, C is a set of label categories, C = {c ₁ , c ₂ , ... c _n }, _Ni is the K nearest neighbor distance of z _i , and its value is calculated by Euclidean distance, K = |N _i |; z _j is the jth sample in Z″, y _j is the label of z _j , and N is the total number of sample images in the current batch; in this embodiment, K = 8;

Step 4-b: Use the cross entropy function to get the probability d _i that the label of z _i is indeed y _i . The calculation formula of d _i is as follows:

d _i = l(Q(z _i ),y _i );

l is the cross entropy function;

Step 4-c: Divide all d _i into n sets according to the sample labels, where the kth set is represented as

D _ck ＝{d _i |i＝1～2N&y _i ＝c _k ,c _k ∈C}

Arrange the elements in D _ck in descending order, and take the first α% of the elements in D _ck as difficult samples, denoted as The remaining 1-α% of the elements in D _ck are taken as simple samples and are recorded as T _ck ; α is an empirical value;

The difficult sample set Simple sample set

Step 4-d: randomly select a simple sample z _a from the simple sample set T as an anchor sample, and A difficult sample _zp with the same label as _za is selected as a positive sample from the set of simple samples T, and a simple sample _zn with a different label from _za is selected as a negative sample to construct a triplet _gr = { _za , _zp , _zn }. The calculation formula of the second loss function _Lhcr is:

in, is a set of triples composed of triples, N _G is the number of triples, is the L2 norm, Φ is a marginal value; in this embodiment, Φ = 0.2;

In this embodiment, an online data enhancement method is used at the image input, including a simple random rotation and random translation method. Specifically, in any small batch, for the input image dataset X, an enhanced dataset X′ with the same data volume and size can be generated after using the data enhancement method. The original dataset X and the enhanced dataset X′ are simultaneously input into the network, and the corresponding prediction results can be obtained after model inference, and finally the model is optimized according to the classification loss calculated by the cross entropy loss;

The calculation process of the third loss function L _cls is:

Wherein, x _e is any image among all sample images and all first images in the current batch, ye _is the label corresponding to x _e , and H _θ (F _θ (x _e )) is the classification result obtained by inputting F _θ (x _e ) into the classification module H _θ ;

In this embodiment, the network structure of the feature extraction and classification model is shown in Table 1 below, where the feature extraction module F _θ is a feature extractor based on 3D-ResNet18, and the classification module H _θ is a linear classifier that maps high-dimensional feature vectors to the number of categories. It should be noted that a feature mapping layer is inserted between the global average pooling layer and the classifier of 3D-ResNet18, which is essentially a 512×128 fully connected layer. Through this feature mapping layer, all images can be mapped into feature representations in a unified feature space. The proposed sorting model and difficult sample recognition and correction model take these feature representations as input and calculate the corresponding losses;

Table 1 Network structure of feature extraction and classification model

Step 4-4: Calculate the total loss function of the network model according to the first loss function, the second loss function and the third loss function, and update the parameters of the network model according to the total loss function, so as to obtain the network model after one training is completed;

The calculation formula of the total loss function is:

L _sum =L _cls +λ1*L _rank +λ2*L _hcr ;

Wherein, λ1 is a preset first weighting coefficient, and λ2 is a preset second weighting coefficient; in this embodiment, λ1=0.01 and λ2=0.001;

Step 5: Input the image to be tested in the test set into the optimal network model to obtain the classification result of the image to be tested.

Considering that brain disease is a progressive process, it is assumed that c ₁ , c ₂ and c ₃ represent the labels of the three progressive processes of Alzheimer's disease, that is, C = {c ₁ , c ₂ , c ₃ } and c ₁ >c ₂ >c ₃ (this partial order relationship represents the three specific severity levels of Alzheimer's disease, such as c ₁ represents normal, c ₂ represents mild cognitive impairment, and c ₃ represents Alzheimer's disease. The higher the label level, the more serious the current disease). The label extraction method corresponding to each SMRI image can refer to the content disclosed in the invention patent "Alzheimer's disease classification method based on improved 3D CNN network" with patent number ZL202010772776.0 (authorization announcement number CN111738363B), then G ₁ ′ is the first image set corresponding to the image set G ₁ , where and The labels are the same, that is, _cr , r∈{1,2,3}.

In the progression of brain diseases, it can be found that the distance between the sMRI image at a certain stage and the enhanced image at the same stage is the shortest, the distance between the sMRI image with adjacent labels is the second shortest, and the distance between the sMRI image with non-adjacent labels is the farthest; the calculation formula of S in this embodiment is:

Among them, S ₁ , S ₂ and S ₃ respectively represent the closest, second closest and third closest distance relationships between labels of different sMRI images in pathological stages, and S ₁ >S ₂ >S ₃ ; it is worth noting that S _i is not a distance value, but a measure of the distance relationship, and what is important is not the value of S _i , but the relationship between S ₁ , S ₂ and S _3. It can be seen that the designed sequential distance relationship matrix S describes the sequential nature of the distance between sMRI images.

exist In , the meanings of the rows from top to bottom and the columns from left to right are the same as those in S. The goal of this model is to drive the distance relationship between image features extracted by the feature extraction module F _θ using the PL model Satisfy the distance relation in S.

Among them, P represents the probability that the predicted true distance matrix V ^Ω satisfies the order distance relationship in S, Quantified The score of the element in S is S _f . Specifically, By averaging The values of all elements with distance relationship value S _f in are obtained.

The present embodiment further relates to a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the sMRI image sequential multi-classification method as described above is executed.

Claims (4)

1. A sequential multi-classification method for sMRI images, comprising the following steps: Step 1, obtaining a certain number of sMRI images and the labels corresponding to each sMRI image to form a sample set; Step 2: Divide the sample set into training set, validation set and test set; Step 3, constructing a network model; the constructed network model includes a feature extraction and classification model, a sorting model and a difficult sample recognition and correction model; wherein the feature extraction and classification model includes a feature extraction module and a classification module connected to the feature extraction module; Step 4: Input all sample images in the training set into the network model constructed in step 3 in batches for training, and use all images in the validation set to verify the performance of the trained network model; after multiple training and validation, select the optimal network model; In each batch of training, at least one sample image with label c ₁ , at least one sample image with label c ₂ , ... at least one sample image with label c _n is selected, where n is the total number of labels; among them, c ₁ >c ₂ >…>c _n , and > is a partial order relationship; The specific process of training the network model using any batch of sample images is as follows: Step 4-1, preprocessing each sample image selected in the current batch to obtain a first image; Step 4-2, inputting all sample images and all first images selected in the current batch into the feature extraction module at the same time to obtain a feature vector; wherein the feature vector includes the first feature vectors corresponding to all sample images and the second feature vectors corresponding to all first images; Step 4-3, input the feature vector obtained in step 4-2 into the sorting model, the difficult sample identification and correction model and the classification module respectively, and calculate the first loss function corresponding to the sorting model, the second loss function corresponding to the difficult sample identification and correction model and the third loss function corresponding to the classification module respectively; The specific calculation process of the first loss function L _rank is: Step 4-31: Select a sample image with label _c1 , a sample image with label _c2 , ..., and a sample image with label _cn from all sample images selected in the current batch, and form them into an image set The initial value of k is 1. Corresponding to the sample image with label c ₁ selected for the kth time, Corresponding to the sample image with label c ₂ selected for the kth time, The corresponding sample image is the one with label c _n selected for the kth time; and get The first image corresponding to The corresponding first image The corresponding first image Calculate the distance relationship matrix S between sMRI images. The calculation formula of S is: The values in the pth row of S represent and The corresponding label distance relationship metric value, p is an odd number, and takes values of 1, 3, 5, ... n respectively; The values in the qth row of S represent and The corresponding label distance relationship metric value, q, is an even number and takes values of 2, 4, 6, ...n+1 respectively; S ₁ , S ₂ , S ₃ …S _n are all the label distance relationship measurement values of any two images, and S ₁ >S ₂ >S ₃ >S _n ; Step 4-32: Calculate the true distance matrix between sMRI images The calculation formula is: in The values in row p represent and The true distance between them, p is an odd number, and takes values of 1, 3, 5, ... n respectively; The values in the qth row represent and The real distance between them, q is an even number, and takes the values of 2, 4, 6, ...n+1 respectively; Right now: for The first feature vector obtained by inputting it into the feature extraction module F _θ ; e∈{1,…n} Right now: for The second feature vector obtained by inputting into the feature extraction module F _θ ; Step 4-33: Calculate the sorting loss corresponding to G _k in step 4-31 Among them, ∏(.) is the multiplication symbol, II(.) is the indicator function. If Sa _,b = _Sf , then II(Sa _,b = _Sf ) = 1; if Sa _,b ≠ _Sf , then II(Sa _,b = _Sf ) = 0; Sa _,b is the value of the ath row and bth column in S. for S _c,d is the value of the cth row and dth column in S; Step 4-34: Add 1 to the value of k, reconstruct the image set G _k+1 , and follow the same method as in steps 4-31 to 4-33 to obtain Step 4-35, and so on to construct image sets G _k+2 , …G _g , wherein the image sets G ₁ , G ₂ , …G _g constructed each time are different, g=Nc ₁ ×Nc ₂ × …×Nc _n , Nc ₁ is the number of sample images with label c ₁ selected in the current batch training, Nc ₂ is the number of sample images with label c ₂ selected in the current batch training, …Nc _n is the number of sample images with label c _n selected in the current batch training; And follow the same method as in step 4-31 to step 4-33 to obtain Then the calculation formula of the first loss function L _rank is: Step 4-4: Calculate the total loss function of the network model according to the first loss function, the second loss function and the third loss function, and update the parameters of the network model according to the total loss function, so as to obtain the network model after one training is completed; Step 5: Input the image to be tested in the test set into the optimal network model to obtain the classification result of the image to be tested. 2. The sMRI image sequential multi-classification method according to claim 1, characterized in that: the specific calculation process of the second loss function L _hcr in step 4-3 is: Step 4-a, all sample images selected in the current batch and all first images are combined into a sample set X″, X″=X∪X′, and the sample set X″ is input into the feature extraction module F _θ to obtain Z″=F _θ (X″), and the posterior probability Q(z _i ) of each sample z _i in Z″ belonging to its own category is calculated, z _i ∈ Z″, and the calculation formula of Q(z _i ) is as follows: Where C is the label category set, C = {c ₁ , c ₂ , … c _n }, _Ni is the K nearest neighbor distance of z _i , whose value is calculated by Euclidean distance, K = |N _i |; z _j is the jth sample in Z″, y _j is the label of z _j , and N is the total number of sample images in the current batch; Step 4-b: Use the cross entropy function to get the probability d _i that the label of z _i is indeed y _i . The calculation formula of d _i is as follows: d _i = l(Q(z _i ),y _i ); l is the cross entropy function; Step 4-c: Divide all d _i into n sets according to the sample labels, where the kth set is represented as D _ck ＝{d _i |i＝1～2N&y _i ＝c _k ,c _k ∈C} Arrange the elements in D _ck in descending order, and take the first α% of the elements in D _ck as difficult samples, denoted as The remaining 1-α% of the elements in D _ck are taken as simple samples and recorded as T _ck ; α is an empirical value; The difficult sample set Simple sample set Step 4-d: randomly select a simple sample z _a from the simple sample set T as an anchor sample, and A difficult sample _zp with the same label as _za is selected as a positive sample from the set of simple samples T, and a simple sample _zn with a different label from _za is selected as a negative sample to construct a triplet _gr = { _za , _zp , _zn }. The calculation formula of the second loss function _Lhcr is: in, is a set of triples composed of triples, N _G is the number of triples, is the L2 norm, and Φ is a marginal value. 3. The sMRI image sequential multi-classification method according to claim 2, characterized in that: the calculation process of the third loss function L _cls in step 4-3 is: Wherein, _xe is any image among all sample images and all first images in the current batch, ye _is the label corresponding to _xe , and _Hθ ( _Fθ ( _xe )) is the classification result obtained by inputting _Fθ ( _xe ) into the classification module _Hθ . 4. The sMRI image sequential multi-classification method according to claim 3, characterized in that: the calculation formula of the total loss function in step 4-4 is: L _sum =L _cls +λ1*L _rank +λ2*L _hcr ; Among them, λ1 is a preset first weighting coefficient, and λ2 is a preset second weighting coefficient.