MRI Hippocampus Segmentation Method and System Based on Hypergraph Numerical Neuron Membrane System

Abstract

The present invention provides an MRI hippocampus segmentation method and system based on a hypergraph numerical neural membrane system, including: acquiring an MR image to be segmented; inputting the MR image to be segmented into the hypergraph numerical neural membrane system to obtain a hippocampus segmentation result; wherein, the hypergraph numerical neural membrane system includes hyper neurons, as well as an encoder and a decoder composed of multiple basic neurons, and the encoder and the decoder are used to learn the upsampling and downsampling features of the hippocampus in the MR image, and the hyper neurons are used to perform skip and concatenation fusion between the encoder and the decoder. By combining the high accuracy of the deep learning model in semantic segmentation with the good parallelism and robustness of the membrane system, the performance of hippocampus segmentation is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation, and in particular relates to an MRI hippocampus segmentation method and system based on a hypergraph numerical neural membrane system. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] The hippocampus is an important brain structure that affects human cognition, memory, and emotion. Accurate segmentation of the hippocampus can help identify early atrophy of the hippocampus, which helps doctors predict and diagnose various neurological diseases such as Alzheimer's disease (AD), epilepsy, and schizophrenia, and formulate treatment plans. However, manual delineation by radiologists takes a lot of time and is often not repeatable. Therefore, an accurate and effective automatic hippocampus segmentation method is needed. However, the following problems make automatic segmentation difficult: First, the hippocampus is located between the thalamus and the medial temporal lobe, and its intensity is similar to that of other brain structures; second, the amygdala and the white matter located below the hippocampus make the boundary of the hippocampus blurred; in addition, the shape and volume of the hippocampus vary from patient to patient.

[0004] Currently, convolutional neural networks (CNNs) have been widely used in medical image segmentation. When performing image segmentation operations, CNNs have excellent feature extraction capabilities and good feature expression capabilities, and do not require manual extraction of image features or excessive preprocessing of images. In recent years, CNNs have achieved great success in the field of medical image segmentation and auxiliary diagnosis. Among them, the neural membrane system is a membrane computing model that adopts the characteristics of spiking neural networks encoding information in time and membrane computing processing information in parallel. The neural membrane system is usually constructed as a two-dimensional graph structure, where vertices represent neurons and edges represent corresponding synapses between neurons. In recent years, the neural membrane system and its extensions have shown excellent convergence, robustness, and parallelism in different fields. In addition, the membrane system is simulated using a graphics processing unit (GPU), making full use of the parallelism in the model.

[0005] The numerical neural membrane system further expands the learning ability of the neural membrane system. However, the current neural membrane system is a two-dimensional graph structure, where neurons can only exchange information with connected neurons, cannot use a hierarchical structure for computing and temporarily store intermediate calculation results, which also limits the application of the neural membrane system in hippocampus segmentation. Summary of the Invention

[0006] To solve the technical problems existing in the above-mentioned background art, the present invention provides an MRI hippocampus segmentation method and system based on a hypergraph numerical neural membrane system, which combines the high precision of deep learning models in semantic segmentation with the good parallelism and robustness of membrane systems, improving the performance of hippocampus segmentation.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] The first aspect of the present invention provides an MRI hippocampus segmentation method based on a hypergraph numerical neural membrane system, which includes:

[0009] Obtain the MR image to be segmented;

[0010] Input the MR image to be segmented into the hypergraph numerical neural membrane system to obtain the hippocampus segmentation result;

[0011] Among them, the hypergraph numerical neural membrane system includes hyperneurons, and an encoder and a decoder composed of multiple basic neurons. The encoder and decoder are used to learn the upsampling and downsampling features of the hippocampus in the MR image, and the hyperneurons are used to perform jump and serial fusion between the encoder and decoder.

[0012] Further, the encoder includes a first basic neuron, a second basic neuron, a third basic neuron, a fourth basic neuron, and a fifth basic neuron connected in sequence;

[0013] The decoder includes a sixth basic neuron, a seventh basic neuron, an eighth basic neuron, and a ninth basic neuron connected in sequence;

[0014] The first basic neuron and the ninth basic neuron, the second basic neuron and the eighth basic neuron, the third basic neuron and the eighth basic neuron, and the fourth basic neuron and the ninth basic neuron are respectively included in the first hyperneuron, the second hyperneuron, the third hyperneuron, and the fourth hyperneuron;

[0015] The learning information of all basic neurons is transmitted to the fifth hyperneuron; the fifth hyperneuron is also the input neuron and the output neuron.

[0016] Further, after the MR image to be segmented is converted into an initial variable, the fifth hyperneuron uses the H rule to send the initial variable to the first basic neuron;

[0017] The first basic neuron performs two convolution operations on the initial variable through the V rule to obtain two post-convolution features, and connects the two post-convolution features through a residual connection to obtain the output feature of the first basic neuron.

[0018] Furthermore, the second basic neuron, the third basic neuron, the fourth basic neuron and the fifth basic neuron all perform maximum pooling on the input through the V rule, and after obtaining the pooled features, execute two or three convolution operation functions to obtain the convolution features, and connect the pooled features and the convolution features through the residual to obtain the output features.

[0019] Furthermore, the sixth basic neuron, the seventh basic neuron, the eighth basic neuron and the ninth basic neuron respectively perform a deconvolution function on the output features of their adjacent basic neurons and the multi-level fusion attention features output by their super neurons through multiple residual connections, and then use the V rule to obtain deconvolution features, and perform two or three convolution functions on the deconvolution features to obtain post-convolution features, and connect the deconvolution features and the post-convolution features through residual connections to obtain output features.

[0020] Furthermore, the first super neuron uses the E rule and the H rule to obtain the multi-level fusion attention features of the output of the first basic neuron.

[0021] Furthermore, the second super neuron, the third super neuron and the fourth super neuron respectively use E rules and H rules for the output features of the basic neurons of the encoders they contain and the multi-level fused attention features output by their adjacent super neurons to obtain the multi-level fused attention features of their outputs.

[0022] A second aspect of the present invention provides an MRI hippocampus segmentation system based on a hypergraph numerical neural membrane system, comprising:

[0023] An image acquisition module is configured to: acquire an MR image to be segmented;

[0024] The hippocampus segmentation module is configured to: input the MR image to be segmented into the hypergraph-based numerical neural membrane system to obtain a hippocampus segmentation result;

[0025] Among them, the hypergraph-based numerical neural membrane system includes a super neuron, and an encoder and a decoder composed of multiple basic neurons. The encoder and the decoder are used to learn the upsampling and downsampling features of the hippocampus in MR images, and the super neuron is used to perform jump and series fusion between the encoder and the decoder.

[0026] The third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system as described above.

[0027] The fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps in the MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system as described above are implemented.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] The present invention provides an MRI hippocampus segmentation method based on a hypergraph numerical neural membrane system, which combines the high precision of deep learning models in semantic segmentation with the good parallelism and robustness of membrane systems, improving the performance of hippocampus segmentation.

[0030] The present invention provides an MRI hippocampus segmentation method based on a hypergraph numerical neural membrane system, which designs three new types of rules, namely the V rule, H rule, and E rule between vertices and hyperedges, to realize the learning ability based on the hypergraph numerical neural membrane system and improve the potential of the hypergraph numerical neural membrane system in hippocampus segmentation.

[0031] The present invention provides an MRI hippocampus segmentation method based on a hypergraph numerical neural membrane system, which makes full use of the high-order correlation of hypergraphs, enabling the membrane system to perform calculations and communications at the hierarchical and planar levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0033] Figure 1 is the flowchart of the method in the first embodiment of the present invention;

[0034] Figure 2 is a schematic diagram of the hypergraph and the corresponding membrane structure in the first embodiment of the present invention;

[0035] Figure 3 is the hippocampus segmentation result diagram before radiotherapy in the first embodiment of the present invention;

[0036] Figure 4 is the hippocampus segmentation result diagram after radiotherapy in the first embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0038] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0039] It should be noted that the terms used herein are for the purpose of describing specific embodiments only and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0040] Term Explanation:

[0041] Numeric neural membrane system: A numeric neural membrane system composed of N > 1 neurons is constructed as follows:

[0042] ∏=(σ 1 ,σ 2 ,…,σ N ,syn,in,out) (1)

[0043] The following cases apply:

[0044] (1) The neuron σ i (1 ≤ i ≤ N) is in the form of σ i =(δ i (0),δ i ,P i ). Among them, is a set of variables existing in the neuron σ i , that is, the vector δ i (0) is a set of initial variables in the neuron σ i ; P i is a set of production functions associated with σ i , and there are two types: non-threshold form and threshold form T q , where T q is the threshold; in the stage f q,i calculate the value P(t)=f q,i (t), and immediately transmit P(t) to all postsynaptic neurons σ j , where (i,j) ∈ syn.

[0045] (2) represents the set of synapses, which is a two-dimensional graphical structure. For each connection (i,j) ∈ syn, 1 ≤ i,j ≤ N and i = j.

[0046] (3) are the input neuron and the output neuron respectively.

[0047] Hypergraph: In hypergraph theory, objects with common attributes belong to the same set, and objects at different levels belong to hyper-sets. These hyper-sets contain special logical structures that can be used to organize complex relationships between objects. Therefore, different from traditional graphs or tree-based structures, hypergraphs can provide more flexible hierarchical relationships for neurons.

[0048] Given a hypergraph G = (V, E, W), H is a matrix representing the hypergraph structure:

[0049]

[0050] where V represents the vertex set of the graph, E represents the edge set of the graph, W is a diagonal matrix of hyper-edge weights, diag(w) = [w(e 1 ), w(e 2 ), …, w(e |E| )].

[0051] Example 1

[0052] As Figure 1 shown, this example provides an MRI hippocampus segmentation method based on a hypergraph numerical neural membrane system, which specifically includes the following steps:

[0053] Step 1: Obtain the MR image to be segmented.

[0054] Step 2: Input the MR image to be segmented into a hypergraph numerical neural (HNN) membrane system to obtain the hippocampus segmentation result. Among them, the hypergraph numerical neural membrane system includes hyper-neurons, as well as an encoder and a decoder composed of multiple basic neurons. The encoder and decoder are used to learn the upsampling and downsampling features of the hippocampus in the MR image, and the hyper-neurons are used to perform skip and concatenation fusion between the encoder and the decoder, or input the image and output the result (only e5 is used to input the MR image and output the segmentation result).

[0055] As Figure 2 shown, Figure 2 the left figure in 1 is an example of a hypergraph. The hypergraph G contains V = {v 2 , v 3} and E = {e 1 , e 2 , e 3}, where e 1 = {v 1 , v 2}, e 2 = {v 2 , v 3}, e 3 = {v 1 , v 2, v 3}. The corresponding neuron structure is as shown in the right figure of Figure 2 . Among them, the neuron without external neurons is the output neuron; the neuron v without internal neurons 1 , v 2 , v 3 is the basic neuron; the neuron e 1 , e 2 , e 3 is the hyperneuron, which respectively includes its subneurons {v 1 , v 2}, {v 2 , v 3} and {v 1 , v 2 , v 3}. Neurons like v 1 can have several different parent neurons, such as e 1 , e 2 , which is defined as the hyper - sub relationship; neurons like v 1 and v 2 have the same hyperneuron e 1 , and there are no external / internal neurons between them, which are called adjacent neurons; neurons like e 1 and e 2 that have external neurons but no internal hyperneurons are also called adjacent neurons. In the HNN membrane system Π with N > 1 neurons, the neuron structure is represented by Hsyn:

[0056] ∏=(σ 1 , σ 2 , …, σ N , Hsyn, in, out) (3)

[0057] Among them, are the input neuron and the output neuron respectively. Each neuron σ i (1 ≤ i ≤ N) has the form σ i =(δ i (0), δ i , R i , P i ), where δ i (0), δ i , R i , P i respectively represent the initial variable, the variable set, the rule type, and the production function of the neuron σ i .

[0058] According to the three relationships in the HNN membrane system, that is, the hyper - sub relationship (H) between v and e and v i , vj With the adjacent relationships (V) and (E) between e i , e j , three new rule types are defined:

[0059] (1) H rule: The basic neuron v generates and sends P v (δ v ) to its super neuron e. At the same time, the super neuron e generates and sends P e (δ e ) to its sub neuron v:

[0060] [v, e]: [(δ v , P v , up); (P e (δ e ), in)] → [((δ e , P e , down); (P v (δ v ), in)] (4) where δ v and δ v are variables in neurons v and e respectively, and P e and P v are related functions.

[0061] (2) V rule: When the basic neurons v i and v j contain the variables δ i and δ j , the V rule is used; the functions P i and P j convert δ i and δ j into P i (δ i ) and P j (δ j ); v i , v j exchange their outputs simultaneously; if i = j, the variable δ i will be changed by P i in neuron v i .

[0062] [v i , v j : [(δ i , P i , forward); (P j (δ j ), in)] → [((δ j , P j , backward); (Pi (δ i ),in)] (5) Among them, δ i and δ j are variables in neurons v i and v j respectively, and P i and P j are related functions.

[0063] (3) E rule: When the hyper neuron e i contains the variable δ i , the E rule is activated. The function P i converts δ i into P i (δ i ), and transmits P i (δ i ) to the connected neuron e j . If i = j, the variable δ i will be changed by P i in the neuron e i .

[0064] [e i ,e j :[δ i ,P i →[{P i (δ i ),δ j}] (6)

[0065] Among them, δ i and δ j are variables in neurons e i and e j respectively, and P i is a related function.

[0066] In the initial state, the HNN membrane system Π includes the structure Hsyn and the neuron σ i =(δ i (0),δ i ,R i ,P i ). In the maximum parallel mode, Π transfers its state from one configuration to another by using the H, V, and E rules. When no rule can be applied or the calculated variables cannot meet these values, Π will stop. When Π terminates, the variable contained in the output neuron out is regarded as the final result of the HNN membrane system Π.

[0067] The present invention proposes an HNN membrane system for hippocampal segmentation. Inspired by the semantic segmentation network (U-Net), the present invention designs the membrane structure (neuron) of the HNN membrane system, such asFigure 1 As shown, an encoder (the first basic neuron 1, the second basic neuron 2, the third basic neuron 3, the fourth basic neuron 4, and the fifth basic neuron 5) and a decoder (the sixth basic neuron 6, the seventh basic neuron 7, the eighth basic neuron 8, and the ninth basic neuron 9) are composed of 9 basic neurons 1 - 9, implementing a U-Net with attention and multi-residual learning (UAM), aiming to accurately and automatically learn the upsampling and downsampling features of the hippocampus. The learning information of neurons 1 - 9 is transmitted to their parent neuron, i.e., the fifth super neuron e 5 . In addition, each pair of basic neurons (i.e., 1, 9; 2, 8; 3, 7; 4, 6) is included in super neurons (i.e., the first super neuron e 1 , the second super neuron e 2 , the third super neuron e 3 , and the fourth super neuron e 4 ) to perform skip and concatenation fusion between the encoder and the decoder. Neuron e 5 is also the input and output neuron.

[0068] The initialization of the HNN membrane system is specifically as follows:

[0069] To accurately segment the hippocampus, the HNN membrane system will learn features and determine the classification (i.e., slices) of a given patient.

[0070] The initial variable δ in the system will be represented as representing a slice (i.e., a pixel matrix) or a related label. Each variable δ in the neuron is regarded as a two-dimensional vector, representing the pixels of a slice or label of size g×h:

[0071] After the MR image to be segmented is converted into the initial variable, that is, when all initial variables δ are generated, the input neuron e 5 uses the H rule to send all its initial variables to the first basic neuron 1 to start the HNN membrane system. After neuron 1 is initialized, the HNN membrane system starts to perform calculations in the maximum parallel mode.

[0072] The calculation process of the HNN membrane system is specifically as follows:

[0073] The encoding path is composed of neurons 1 - 5, and each neuron performs two or three convolution operation functions P through the V rule c . The obtained features are connected by multi-residual within neurons 1 - 5.

[0074] The first basic neuron performs two convolution operations on the initial variable through the V rule to obtain two features, and connects the two features through a residual connection to obtain the output of the first basic neuron. Specifically, as Figure 1As shown, in neuron 1, δ 11 and δ 12 are features generated by P c . The output δ(1) of neuron 1 is the result of δ 11 and δ 12 through a residual connection. Expressed by the formula:

[0075] δ 11 = P c (δ), δ 12 = P c (δ 11 )

[0076]

[0077] After that, the outputs of neurons 1 - 4, namely δ(1) - δ(4), are obtained through max - pooling (kernel size 2×2), group normalization, rectified linear unit, and dropout calculations performed by Pm. The second basic neuron, the third basic neuron, the fourth basic neuron, and the fifth basic neuron all perform max - pooling on the input through the V rule. After obtaining the pooled features, they execute two or three convolution operation functions to get the convolved features, and then connect the pooled features and the convolved features through a residual connection to obtain the output features. Specifically, expressed by the formula:

[0078] Second basic neuron

[0079] δ 21 = P m (δ(1)), δ 22 = P c (δ 21 )

[0080] δ 23 = P c (δ 22 )

[0081]

[0082] Third basic neuron

[0083] δ 31 = P m (δ(2)), δ 32 = P c (δ 31 )

[0084]

[0085]

[0086] Fourth basic neuron

[0087] δ41 = P m (δ(3)), δ 42 = P c (δ 41 )

[0088]

[0089]

[0090] The fifth basic neuron

[0091] δ 51 = P m (δ(4)), δ 52 = P c (δ 51 )

[0092]

[0093]

[0094] The decoding path includes neurons 6 - 9, and each neuron executes a deconvolution function P using the V rule d and multiple convolution functions P with multi - residual connections as shown in c , such as Figure 1 shown. Specifically, the sixth basic neuron, the seventh basic neuron, the eighth basic neuron, and the ninth basic neuron respectively perform multi - residual connection on the output features of their adjacent basic neurons and the multi - level fusion attention features output by their super - neurons, and then use the V rule to execute a deconvolution function to obtain deconvolution features, and perform two or three convolution functions on the deconvolution features to obtain post - convolution features. The deconvolution features and the post - convolution features are connected through a residual connection to obtain output features. Specifically, it is expressed by the formula as:

[0095] The sixth basic neuron

[0096]

[0097] δ 62 = P c (δ 61 ), δ 63 = P c (δ 62 ),

[0098]

[0099]

[0100] The seventh basic neuron

[0101]

[0102] δ 72 = P c (δ 71 ), δ 73 = P c (δ 72 ),

[0103]

[0104]

[0105] The eighth basic neuron

[0106]

[0107] δ 82 = P c (δ 81 ), δ 83 = P c (δ 82 ),

[0108]

[0109] The ninth basic neuron

[0110]

[0111] δ 92 = P c (δ 91 ), δ 93 = P c (δ 92 ),

[0112]

[0113] Finally, the feature δ e51 is calculated by the sigmoid activation function P 5 in the out neuron e s and supervised by the cross-entropy loss P l to obtain the final output δ f . Specifically, it is expressed by the formula:

[0114] δ e51 = P s (δ(9)),

[0115] δ f = P l (δ e51 )

[0116] To enhance the boundary and content features of the hippocampus, the hyperneuron e 1 -e 4 implements a multi-level fusion attention mechanism for the features of adjacent neurons in the encoder and corresponding neurons in the decoder ( Figure 1 neuron e 1 -e 4 ). Specifically:

[0117] For the output features of the first basic neuron by the first hyperneuron, using the E rule and the H rule, the multi-level fusion attention features output by the first hyperneuron are obtained. Specifically, it is expressed by the formula:

[0118]

[0119] For the output features of the basic neurons of the encoder included in the second hyperneuron, the third hyperneuron, and the fourth hyperneuron respectively, and the multi-level fusion attention features output by their adjacent hyperneurons, using the E rule and the H rule, the multi-level fusion attention features output by them are obtained. For example, for neurons 1 and 2, for adjacent neurons in the encoder, neuron e 2 uses the E rule and the H rule to generate the feature δ(e2), and is connected by δ e21 and δ e22 . Specifically, δ e21 and δ e22 are the features obtained through the operations in neuron 2 and the operation of the convolution function P c on δ(e1) respectively. Is implemented by the operation in neuron 1. Therefore, δ(e2) is the feature fusion of adjacent neurons 1 and 2. For the neurons in the encoder and the corresponding neurons in the decoder, such as neurons 2 and 8, the feature δ(e2) is sent to neuron 8 to be fused with other features using the H rule. Specifically, it is expressed by the formula:

[0120] Second hyperneuron

[0121]

[0122] δ e22 =P c (δ(e1))

[0123]

[0124] Third hyperneuron

[0125]

[0126] δ e32 =P c (δ(e2))

[0127]

[0128] Fourth super neuron

[0129]

[0130] δ e42 = P c (δ(e3))

[0131]

[0132] The stop condition and output of the HNN membrane system are specifically as follows:

[0133] The above calculation process is supervised by the cross-entropy loss P l and optimized by the Adam algorithm. The HNN membrane system will be iteratively updated until a predetermined number of iterations is reached. When the HNN membrane system stops, the variable in the out neuron e 5 is regarded as the final result of the HNN membrane system.

[0134] The HNN membrane system proposed by the present invention is implemented on an NVIDIA Tesla V100 GPU with 32GB RAM using the TensorFlow framework. The Adam algorithm iteratively updates the weights of the convolutional kernels to optimize the model. The initial learning rate is 0.003, and exponential decay is performed every 500 iterations. To avoid overfitting, the dropout rate is set to 0.7. The Batch size is 8. During the training process, the network is trained for 40,000 iterations. Each patient in the test dataset is evaluated within only 13 seconds. The dataset (SCH dataset) used is collected from Shandong Tumor Hospital. The dataset consists of 184 patients, including 93 cases before radiotherapy and 91 cases after radiotherapy. The experiment of the present invention adopts 4-fold cross-validation (CV-4).

[0135] For comparison, the method of the present invention is evaluated on the SCH dataset collected from Shandong Tumor Hospital. The evaluation metrics in the experiment are respectively: Dice similarity coefficient (DSC), average surface distance (ASD), positive predictive value (PPV), and sensitivity (SEN). After 4-fold cross-validation, the mean values obtained on the patient data before and after radiotherapy are 0.917±0.009, 0.208±0.028, 0.921±0.010, 0.919±0.012 and 0.916±0.018, 0.174±0.015, 0.889±0.030, 0.945±0.007, which are better than the previous methods.

[0136] Figure 3 and Figure 4Pre - radiotherapy ([ Figure 3 ) and post - radiotherapy ([ Figure 4 ) hippocampal segmentation results for three cases (Case 1, Case 2, and Case 3) on the SCH dataset. Figure 3 and Figure 4 In, the left side is the original MR image, the middle column is the MR image with labels, and the right - hand column is the display diagram of the HNN membrane system segmentation result, where the index values are marked on the corresponding figure.

[0137] Convolutional neural networks have superior capabilities in automatically learning hierarchical information and preserving spatial information compared to traditional multi - layer perceptrons, achieving advanced results in medical image segmentation tasks. In addition, to understand more detailed information about the hippocampus, residual learning is incorporated into the U - net, and residual learning solves the problem of performance degradation of deep convolutional neural networks to a certain extent. To highlight useful features in the hippocampus, attention mechanisms are widely adopted in CNNs, which focus on regions of interest while suppressing irrelevant background regions. Therefore, the present invention uses a U - Net with attention and residual learning as the core computational mechanism for hippocampal segmentation.

[0138] Different from graph - based structures, hypergraphs can provide more complex data relationships, and hyper - edges connect any number of vertices. Therefore, developing a hypergraph - structured neural membrane system can extend the membrane structure to a high - dimensional and more complete non - linear space, endowing the membrane system with more powerful computational capabilities. The present invention proposes a hypergraph numerical neural (HNN) membrane system for semantic segmentation of the hippocampus from magnetic resonance imaging (MRI), which combines the advantages of neural membrane systems and deep learning models.

[0139] Example Two

[0140] This example provides an MRI hippocampal segmentation system based on a hypergraph numerical neural membrane system, which specifically includes the following modules:

[0141] An image acquisition module, which is configured to: acquire the MR image to be segmented;

[0142] A hippocampal segmentation module, which is configured to: input the MR image to be segmented into the hypergraph numerical neural membrane system to obtain the hippocampal segmentation result;

[0143] Among them, the hypergraph numerical neural membrane system includes hyper - neurons, and an encoder and a decoder composed of multiple basic neurons. The encoder and decoder are used to learn the up - sampling and down - sampling features of the hippocampus in the MR image, and the hyper - neurons are used for skip and concatenation fusion between the encoder and the decoder.

[0144] It should be noted here that each module in this embodiment corresponds to each step in Embodiment 1 one by one, and the specific implementation process is the same, so it will not be repeated here.

[0145] Embodiment 3

[0146] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the steps in the MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system as described in Embodiment 1 above.

[0147] Embodiment 4

[0148] This embodiment provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps in the MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system as described in Embodiment 1 above.

[0149] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories and optical memories, etc.) containing computer-usable program code.

[0150] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0151] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0152] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions for implementing the process Figure 1 in one process or multiple processes and / or boxes Figure 1 steps for the functions specified in one box or multiple boxes.

[0153] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), or the like.

[0154] The foregoing is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. MRI hippocampus segmentation method based on hypergraph numerical neural membrane system, characterized in that, it includes: Obtain the MR image to be segmented; Input the MR image to be segmented into the hypergraph numerical neural membrane system to obtain the hippocampus segmentation result; Among them, the hypergraph numerical neural membrane system includes hyperneurons, as well as an encoder and a decoder composed of multiple basic neurons. The encoder and decoder are used to learn the upsampling and downsampling features of the hippocampus in the MR image, and the hyperneurons are used to perform jump and serial fusion between the encoder and the decoder; Among them, the encoder includes a first basic neuron, a second basic neuron, a third basic neuron, a fourth basic neuron, and a fifth basic neuron connected in sequence; Among them, the decoder includes a sixth basic neuron, a seventh basic neuron, an eighth basic neuron, and a ninth basic neuron connected in sequence; Among them, the first basic neuron and the ninth basic neuron, the second basic neuron and the eighth basic neuron, the third basic neuron and the eighth basic neuron, and the fourth basic neuron and the ninth basic neuron are respectively included in the first hyperneuron, the second hyperneuron, the third hyperneuron, and the fourth hyperneuron; Among them, the learning information of all basic neurons is transmitted to the fifth hyperneuron; the fifth hyperneuron is also the input neuron and the output neuron; Among them, after the MR image to be segmented is converted into an initial variable, the fifth hyperneuron uses the H rule to send the initial variable to the first basic neuron; Among them, the first basic neuron performs two convolution operations on the initial variable through the V rule to obtain two post-convolution features, and connects the two post-convolution features through a residual connection to obtain the output feature of the first basic neuron; Among them, according to the three relationships in the hypergraph numerical neural membrane, namely the hyper-sub relationship (H) between v and e and v i , v j and e i , e j the adjacent relationships (V) and (E) between them, three new rule types are defined: H rule: Cause the basic neuron v to generate and send P v (δ v ) to its hyper neuron e. At the same time, the hyper neuron e generates and sends P e (δ e ) to its sub neuron v: [v,e]:[(δ v ,P v ,up);(P e (δ e ),in)] [((δ e , P e , down); (P v (δ v ), in)] where, δ v and δ e are variables in neurons v and e respectively, P e and P v are correlation functions; V rule: When the basic neurons v i and v j contain variables δ i and δ j , use the V rule; the functions P i and P j convert δ i and δ j into P i and P j ; v i , v j simultaneously exchange their outputs; if i = j, the variable δ i will be changed by the P i in the neuron v i : [v i , v j : [(δ i , P i , forward); (P j (δ j ), in)] [((δ j , P j , backward); (P i (δ i ), in)] where δ i and δ j are variables of v i and v j in the neuron, and P i and P j are correlation functions; E rule: When hyper neuron e i contains variable δ i the E rule is activated, and function P i converts δ i into P i (δ i ) and passes P i (δ i ) to the connected neuron e j . If i = j, variable δ i will be changed by P i in neuron e i : [e i , e j : [δ i , P i → [{P i (δ i ), δ j}] Among them, δ i and δ j are variables in neurons e i and e j respectively, and P i is the correlation function.

2. The MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system according to claim 1, characterized in that, The second basic neuron, the third basic neuron, the fourth basic neuron, and the fifth basic neuron all perform max pooling on the input through the V rule to obtain the pooled feature, and then execute two or three convolution operation functions to obtain the post-convolution feature, and connect the pooled feature and the post-convolution feature through a residual connection to obtain the output feature.

3. The MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system according to claim 1, characterized in that, The sixth basic neuron, the seventh basic neuron, the eighth basic neuron, and the ninth basic neuron respectively perform multi-residual connection on the output feature of its adjacent basic neuron and the multi-level fusion attention feature output by its hyperneuron, and then use the V rule to execute an anti-convolution function to obtain the anti-convolution feature, and perform two or three convolution functions on the anti-convolution feature to obtain the post-convolution feature, and connect the anti-convolution feature and the post-convolution feature through a residual connection to obtain the output feature.

4. The MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system according to claim 1, characterized in that, The first hyperneuron uses the E rule and the H rule for the output feature of the first basic neuron to obtain the multi-level fusion attention feature output by the first hyperneuron.

5. The MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system according to claim 1, characterized in that, the second hyperneuron, the third hyperneuron, and the fourth hyperneuron respectively use the E rule and the H rule for the output features of the basic neurons of the encoders they contain and the multi-level fusion attention features output by their adjacent hyperneurons to obtain the multi-level fusion attention features they output.

6. An MRI hippocampus segmentation system based on the hypergraph numerical neural membrane system, characterized in that, comprising: an image acquisition module configured to: acquire an MR image to be segmented; a hippocampus segmentation module configured to: input the MR image to be segmented into the hypergraph numerical neural membrane system to obtain a hippocampus segmentation result; wherein, the hypergraph numerical neural membrane system includes hyperneurons, and an encoder and a decoder composed of multiple basic neurons, the encoder and the decoder are used to learn the upsampling and downsampling features of the hippocampus in the MR image, and the hyperneurons are used to perform skip and concatenation fusion between the encoder and the decoder; wherein, the encoder includes a first basic neuron, a second basic neuron, a third basic neuron, a fourth basic neuron, and a fifth basic neuron connected in sequence; wherein, the decoder includes a sixth basic neuron, a seventh basic neuron, an eighth basic neuron, and a ninth basic neuron connected in sequence; wherein, the first basic neuron and the ninth basic neuron, the second basic neuron and the eighth basic neuron, the third basic neuron and the eighth basic neuron, and the fourth basic neuron and the ninth basic neuron are respectively included in the first hyperneuron, the second hyperneuron, the third hyperneuron, and the fourth hyperneuron; wherein, the learning information of all basic neurons is transmitted to the fifth hyperneuron; the fifth hyperneuron is also an input neuron and an output neuron; wherein, after the MR image to be segmented is converted into an initial variable, the fifth hyperneuron uses the H rule to send the initial variable to the first basic neuron; wherein, the first basic neuron performs two convolution operations on the initial variable through the V rule to obtain two post-convolution features, and connects the two post-convolution features through a residual connection to obtain the output feature of the first basic neuron; Among them, according to the three relationships in the hypergraph numerical neural membrane, namely the hyperon relationship (H) between v and e and v i , v j and e i , e j the adjacent relationships (V) and (E) between them, three new rule types are defined: H rule: Make the basic neuron v generate and send P v (δ v ) to its hyper neuron e. At the same time, the hyper neuron e generates and sends P e (δ e ) to its sub neuron v: [v,e]:[(δ v ,P v ,up);(P e (δ e ),in)] [((δ e , P e , down); (P v (δ v ), in)] Among them, δ v and δ e are variables in neurons v and e respectively, P e and P v are correlation functions; V rule: When the basic neurons v i and v j contain variables δ i and δ j , use the V rule; the functions P i and P j convert δ i and δ j into P i and P j ; v i , v j simultaneously exchange their outputs; if i = j, the variable δ i will be changed by the P i in the neuron v i : [v i , v j : [(δ i , P i , forward); (P j (δ j ), in)] [((δ j , P j , backward); (P i (δ i ), in)] Among them, δ i and δ j are respectively the variables of v i and v j in the neuron, and P i and P j are related functions; E rule: When the hyper neuron e i contains the variable δ i , the E rule is activated, and the function P i converts δ i into P i (δ i ), and passes P i (δ i ) to the connected neuron e j . If i = j, the variable δ i will be changed by the P i in the neuron e i : [e i , e j : [δ i , P i → [{P i (δ i ), δ j}] where δ i and δ j are variables in neurons e i and e j respectively, and P i is a correlation function.

7. A computer-readable storage medium, on which a computer program is stored, characterized in that, when the program is executed by a processor, it implements the steps in the MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system according to any one of claims 1-5.

8. A computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the steps in the MRI hippocampus segmentation method based on the hypergraph numerical neural membrane system according to any one of claims 1-5.

Images