A hippocampus segmentation method and device based on deep learning and a storage medium

Abstract

The embodiment of the application discloses a hippocampus segmentation method and device based on deep learning and a storage medium, wherein the hippocampus segmentation method based on deep learning comprises the following steps: S1, a first image generated by nuclear magnetic resonance imaging is acquired, a target shape in the first image is inferred by using a trained deep belief network, and a second image containing the target shape is obtained; S2, an energy function is constructed according to the second image, a deep belief network driven lattice Boltzmann model is obtained based on the energy function, curve evolution is performed on the first image by using the deep belief network driven lattice Boltzmann model, and a third image and a fourth image after segmentation are obtained, and the fourth image contains the target shape; S3, steps S1 to S2 are executed for multiple times, multiple groups of third images and fourth images obtained are fed back to an error correction convolutional neural network model, multiple fifth images are obtained, the multiple fifth images are averaged and fused, and a segmentation result is obtained.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to a hippocampus segmentation method, apparatus, and storage medium based on deep learning. Background Technology

[0002] Alzheimer's disease (AD) is a chronic neurological disorder characterized by memory loss and cognitive impairment. Currently, there is no effective drug treatment to cure AD; treatments can only alleviate symptoms or slow its progression. Early diagnosis of AD helps determine its progression and can improve the quality of life for AD patients. Clinically, it has been found that the hippocampus (HC) is one of the affected areas in Alzheimer's disease, and its shape and volume change.

[0003] Currently, the diagnosis of Alzheimer's disease (AD) involves clinical, neuropsychological, and neuroimaging assessments, with magnetic resonance imaging (MRI) considered the preferred neuroimaging examination for AD. MRI allows for high tissue contrast and accurate measurement of the size of three-dimensional (3D) volumetric brain structures, particularly the hippocampus and related regions. Therefore, MRI-based hippocampal segmentation is a crucial step in the diagnosis and detection of Alzheimer's disease. Furthermore, MRI offers excellent spatial resolution, high accessibility, and good contrast. Thus, MRI-based hippocampal segmentation is a powerful biomarker for detecting AD. However, automated segmentation of hippocampal structures is challenging due to their volume, complex shape, low contrast, and discontinuous boundaries. Therefore, it is difficult to establish a sufficiently general model to cover all possible shapes of the hippocampus.

[0004] For hippocampal segmentation to diagnose Alzheimer's disease (AD), medical institutions have traditionally relied on manual segmentation. This requires specialized training for imaging physicians, which is not only time-consuming and labor-intensive but also extremely costly to train. Furthermore, manual segmentation is susceptible to the influence of uncertainties among imaging physicians. With the development of artificial intelligence in recent years, various hippocampal segmentation methods have emerged, including but not limited to image-based methods, active contour models (ACM), active appearance and shape models, atlas (Atlas) models, and deep learning. However, these models have several drawbacks, including low robustness and accuracy, the need for large training sets to build AI models resulting in time-consuming, labor-intensive, and costly hippocampal segmentation with high uncertainty, the inability of the constructed models to cover all hippocampal shapes, and high computational costs. Summary of the Invention

[0005] The purpose of this application is to provide a hippocampal segmentation method, device, and storage medium based on deep learning, in order to solve the problems of low robustness and accuracy of existing models, the need for a large number of training sets to build artificial intelligence models, resulting in time-consuming, labor-intensive, costly, and uncertain hippocampal segmentation, the inability of the constructed models to cover all hippocampal shapes, and high computational cost.

[0006] To achieve the above objectives, this application provides a deep learning-based hippocampal segmentation method, comprising the following steps: S1, acquiring a first image generated by magnetic resonance imaging, inferring the target shape in the first image using a trained deep belief network, and obtaining a second image containing the target shape; S2, constructing an energy function based on the second image, obtaining a lattice Boltzmann model driven by the deep belief network based on the energy function, and using the lattice Boltzmann model driven by the deep belief network to perform curve evolution on the first image to obtain a segmented third image and a fourth image, wherein the fourth image contains the target shape; S3, executing steps S1 to S2 multiple times, feeding the obtained multiple sets of the third and fourth images into an error-correcting convolutional neural network model to obtain multiple fifth images, and averaging and fusing the multiple fifth images to obtain a segmentation result.

[0007] Optionally, the method for inferring the target shape in the first image using the trained deep belief network includes:

[0008] The deep belief network performs bottom-up reasoning on the input first image until it reaches the top hidden and visible layers. The inference model of the deep belief network is as follows:

[0009]

[0010] Where v is the vector representation of the input image, The label represents v, and h represents a hidden variable.

[0011] as well as ,in

[0012]

[0013] Represents the energy function of RBM. It is the bias vector. and It is a weight matrix;

[0014] Initialize the L layer and perform Gibbs sampling on the visible layers h2, h1 and L until convergence, thereby obtaining the target shape.

[0015] Optionally, before obtaining the lattice Boltzmann model driven by the deep belief network based on the energy function, the method further includes: constructing the lattice Boltzmann model, wherein the evolution equation of the lattice Boltzmann model is:

[0016]

[0017] in This represents the gray level of pixel r=(x,y) at time t. yes Gray-level distribution function on the directional sub-pixel, This represents the equilibrium distribution function describing the redistributed predicted values. It is the grid size of time. It is the relaxation time. It is along External force in direction, each pixel r contains its nine nearest neighbors, lattice vector The position of each sub-pixel is defined as:

[0018]

[0019] The lattice Boltzmann model consists of two steps, including a collision step and a flow step.

[0020] Optionally, the method for obtaining the lattice Boltzmann model driven by the deep belief network based on the energy function includes:

[0021] The energy function is minimized using the gradient descent flow method, resulting in a partial differential equation.

[0022] By comparing the evolution equation of the lattice Boltzmann model with the partial differential equation, the lattice Boltzmann model driven by the deep belief network is obtained.

[0023] Optionally, the error-correcting convolutional neural network model includes a replacement module, an optimization module, and convolutional layers;

[0024] After feeding the obtained sets of third and fourth images into the error-correcting convolutional neural network model, the process further includes: the third and fourth images are first processed by the replacement module, then processed by the optimization module, and then enter the convolutional layer to obtain the binarized fifth image.

[0025] To achieve the above objectives, this application also provides a deep learning-based hippocampal segmentation device, comprising: a memory; and

[0026] A processor connected to the memory, the processor being configured to:

[0027] S1. Obtain a first image generated by nuclear magnetic resonance imaging, and use a trained deep belief network to infer the target shape in the first image to obtain a second image containing the target shape;

[0028] S2. Based on the second image, construct an energy function, obtain the lattice Boltzmann model driven by the deep belief network based on the energy function, and use the lattice Boltzmann model driven by the deep belief network to perform curve evolution on the first image to obtain the segmented third image and fourth image, wherein the fourth image contains the target shape;

[0029] S3. Repeat steps S1 to S2 multiple times, feed the obtained sets of third and fourth images into the error correction convolutional neural network model to obtain multiple fifth images, and perform average fusion on the multiple fifth images to obtain the segmentation result.

[0030] To achieve the above objectives, this application also provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a machine, implements the steps of the method described above.

[0031] The embodiments of this application have the following advantages:

[0032] 1. This application provides a deep learning-based hippocampal segmentation method, comprising the following steps: S1, acquiring a first image generated by magnetic resonance imaging, inferring the target shape in the first image using a trained deep belief network, and obtaining a second image containing the target shape; S2, constructing an energy function based on the second image, obtaining a lattice Boltzmann model driven by the deep belief network based on the energy function, and using the lattice Boltzmann model driven by the deep belief network to perform curve evolution on the first image to obtain a segmented third image and a fourth image, wherein the fourth image contains the target shape; S3, executing steps S1 to S2 multiple times, feeding the obtained multiple sets of the third and fourth images into an error-correcting convolutional neural network model to obtain multiple fifth images, and averaging and fusing the multiple fifth images to obtain a segmentation result.

[0033] Using the above method, the shape of the object to be segmented can be modeled with a small number of training samples. Accurate segmentation results can be obtained without complex network structures. Sequence label alignment, shape priority, or target object registration are not required. The addition of shape prior technology makes the model closer to the standard case for segmentation, resulting in more accurate segmentation results. It can correctly segment partially occluded and missing targets and is more sensitive to changes in shape, size, and volume. Error correction CNN technology is used, which not only eliminates random errors but also eliminates systematic errors caused by the use of machine systems. Attached Figure Description

[0034] To more clearly illustrate the embodiments of this application or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0035] Figure 1 A flowchart illustrating a deep learning-based hippocampal segmentation method provided in this application embodiment;

[0036] Figure 2 A flowchart illustrating a deep learning-based hippocampal segmentation method provided in this application embodiment;

[0037] Figure 3 A flowchart of a deep belief network model for a hippocampal segmentation method based on deep learning, provided for an embodiment of this application;

[0038] Figure 4 A flowchart of a deep belief network and lattice Boltzmann model segmentation method for hippocampal segmentation based on deep learning provided in this application embodiment;

[0039] Figure 5 A schematic diagram of the collision process of a lattice Boltzmann model for a deep learning-based hippocampal segmentation method provided in this application embodiment;

[0040] Figure 6 A schematic diagram of the flow process of a lattice Boltzmann model based on a deep learning-based hippocampal segmentation method provided in an embodiment of this application;

[0041] Figure 7 A schematic diagram of the D2Q5 lattice structure of a lattice Boltzmann model for a deep learning-based hippocampal segmentation method provided in this application embodiment;

[0042] Figure 8 A flowchart of a lattice Boltzmann model for a hippocampal segmentation method based on deep learning, provided for embodiments of this application;

[0043] Figure 9 A flowchart of an error-correcting convolutional neural network model for a deep learning-based hippocampal segmentation method provided in this application embodiment;

[0044] Figure 10 A schematic diagram of the segmentation result of a deep learning-based hippocampal segmentation method provided in an embodiment of this application;

[0045] Figure 11A block diagram of a deep learning-based hippocampal segmentation device provided in an embodiment of this application. Detailed Implementation

[0046] The following specific embodiments illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] Furthermore, the technical features involved in the different embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0048] One embodiment of this application provides a deep learning-based hippocampal segmentation method, referencing... Figure 1 and Figure 2 , Figure 1 The flowchart provided in one embodiment of this application illustrates a deep learning-based hippocampal segmentation method. It should be understood that the method may also include additional boxes not shown and / or the boxes shown may be omitted, and the scope of this application is not limited in this respect.

[0049] Step S1 involves acquiring a first image generated by magnetic resonance imaging (MRI), and using a trained deep belief network to infer the target shape in the first image, thereby obtaining a second image containing the target shape. The first image is an image acquired through MRI.

[0050] In some embodiments, a method for inferring the target shape in the first image using the trained deep belief network includes:

[0051] The deep belief network performs bottom-up reasoning on the input first image until it reaches the top hidden and visible layers. The inference model of the deep belief network is as follows:

[0052]

[0053] Where v is the vector representation of the input image, The label represents v, and h represents a hidden variable.

[0054] as well as ,in

[0055]

[0056] Represents the energy function of RBM. It is the bias vector. and It is a weight matrix;

[0057] Initialize the L layer and perform Gibbs sampling on the visible layers h2, h1 and L until convergence, thereby obtaining the target shape.

[0058] Specifically, refer to Figure 3 The shape of the HC (i.e., the target shape) is inferred by utilizing and training a DBN. A DBN inference model was developed using joint probabilities, as follows:

[0059]

[0060] Where v is the vector representation of the input image, The label represents v, and h represents a hidden variable.

[0061] as well as ,in

[0062]

[0063] This represents the energy function of the RBM.

[0064] It is the bias vector; and It is a weight matrix.

[0065] Furthermore, this application has

[0066]

[0067] ,in , operator Return the j-th vector value. Return the j-th column of the matrix.

[0068] The DBN inference model is trained layer-by-layer in an unsupervised manner using stacked Restricted Boltzmann Machines (RBMs). Each RBM has three layers: a hidden layer, a visible layer, and a bias layer. In this unsupervised training process, the error to be minimized is the reconstruction error of the visible input. The result obtained from the previous RBM is used as the "visible" input for the next RBM. When the segmentation label L is provided as the visible input, supervised training only begins with the top RBM. The inference process involves acquiring the first input image (obtained via MRI) and performing bottom-up inference until the top two layers (the hidden and visible layers) are reached. Then, layer L is initialized to 0, and Gibbs sampling is performed on the visible layers h2, h1, and L until convergence, resulting in shape inference.

[0069] Execute step S2: Based on the second image, construct an energy function, obtain the lattice Boltzmann model driven by the deep belief network based on the energy function, and use the lattice Boltzmann model driven by the deep belief network to perform curve evolution on the first image to obtain the segmented third and fourth images. The fourth image contains the target shape. (Refer to...) Figure 4 .

[0070] In some embodiments, before obtaining the deep belief network-driven lattice Boltzmann model based on the energy function, the method further includes: constructing the lattice Boltzmann model, wherein the evolution equation of the lattice Boltzmann model is:

[0071]

[0072] in This represents the gray level of pixel r=(x,y) at time t. yes Gray-level distribution function on the directional sub-pixel, This represents the equilibrium distribution function describing the redistributed predicted values. It is the grid size of time. It is the relaxation time. It is along External force in direction, each pixel r contains its nine nearest neighbors, lattice vector The position of each sub-pixel is defined as:

[0073]

[0074] The lattice Boltzmann model consists of two steps, including a collision step and a flow step, respectively referring to... Figure 5 and Figure 6 .

[0075] Specifically, the LB model typically consists of two steps: a collision step and a flow step. The collision step involves particles colliding, which changes their velocities, thus redistributing the particle density function in each grid cell. The flow step refers to the movement of particles after the collision step.

[0076] The LB evolution equation in image processing is defined as follows:

[0077]

[0078] in This represents the gray level of pixel r=(x,y) at time t; it is considered as the quality of the subpixel. yes Gray-level distribution function on the directional sub-pixel. This represents the equilibrium distribution function that describes the redistributed predicted values. It is the grid size of time. It is the relaxation time. It is along External force in direction. Each pixel r has nine nearest neighbors (including itself). Lattice vector. The position of each sub-pixel is defined as:

[0079]

[0080] During the collision process, It has become The trend is as follows. In this step, the grayscale value of each sub-pixel is divided into two parts: one part is redistributed, and the other part remains on the original sub-pixel. The collision equation can be expressed as:

[0081]

[0082] In the flow processing step, the grayscale of each sub-pixel is updated based on the grayscale of its neighboring sub-pixels. Because Therefore, the grayscale of the pixels will change.

[0083] The flow equation is:

[0084]

[0085] To improve computational efficiency, this embodiment selects a D2Q5 lattice and applies the Chapman–Enskog expansion method to the LB evolution equation in image processing, referencing... Figure 7 ,get:

[0086]

[0087] The above is the constructed LB model.

[0088] In some embodiments, a method for obtaining the lattice Boltzmann model driven by the deep belief network based on the energy function includes:

[0089] The energy function is minimized using the gradient descent flow method, resulting in a partial differential equation.

[0090] By comparing the evolution equation of the lattice Boltzmann model with the partial differential equation, the lattice Boltzmann model driven by the deep belief network is obtained.

[0091] Specifically, refer to Figure 8 The shape inferred from DBN is used not only as part of the external force term of the LB segmentation model, but also to determine the position of the initial contour. Therefore, this embodiment constructs an energy function that includes a weighted gradient term, a weighted region term, and a shape energy term, as follows:

[0092]

[0093]

[0094]

[0095] in It is the gradient operator. Therefore is the Gaussian kernel, and I is the edge indicator for image grayscale. The shape is inferred from the trained DBN, and its definition is as follows:

[0096]

[0097] Where s represents the shape mask inferred from the DBN. H is the Heaviside function. It is the Dirac delta function:

[0098]

[0099]

[0100] The energy function is: Where λ>0, μ>0, α Minimizing the energy function using the gradient descent flow method yields the following partial differential equation (PDE):

[0101]

[0102] Due to the distance function The above equation is in the following case:

[0103]

[0104] in

[0105] Comparing the LB evolution equation and the partial differential equation (PDE) in the above image processing, we obtain:

[0106]

[0107] The LB method in image segmentation can be interpreted as the curve evolution process of isodense lines (i.e., initial contours) under the action of internal forces (i.e., diffusion) and external forces (i.e., gradient, region, shape priors).

[0108] Execute step S3 and execute steps S1 to S2 multiple times to feed the obtained sets of third and fourth images back into the error-correcting convolutional neural network model to obtain multiple fifth images. Then, perform average fusion on the multiple fifth images to obtain the segmentation result.

[0109] In some embodiments, the error-correcting convolutional neural network model includes a replacement module, an optimization module, and convolutional layers;

[0110] After feeding the obtained sets of third and fourth images into the error-correcting convolutional neural network model, the process further includes: the third and fourth images are first processed by the replacement module, then processed by the optimization module, and then enter the convolutional layer to obtain the binarized fifth image.

[0111] Specifically, refer to Figure 9 In automatic segmentation, random or systematic errors may occur due to noise, anatomical differences, or segmentation methods. Therefore, after the initial segmentation is completed, this application adds error correction CNN technology to correct errors caused by randomness in the system.

[0112] The segmented MRI image (third image) and the segmented HC extract image (fourth image) are fed into the error correction CNN, and then the resulting orthogonal slices (fifth image) are averaged and fused in the later stage. The error correction CNN is a combination of a replacement CNN and an optimization CNN. First, the replacement CNN is used, followed by the corresponding optimization CNN.

[0113] First, input 2D images (d1×d2 pixels) of the third and fourth images. The input images have 5 channels, 2 of which are used to replace the CNN, and 3 are used to optimize the CNN. The input images are then passed through 3×3 filters with additional bias parameters, and finally activated by the ReLU activation function. ,Right now:

[0114] First, downsample the feature maps of the same layer. Performing a 1×1×1 convolution operation yields The feature map of the previous layer upsampling. Performing a 1×1×1 convolution operation yields... The feature maps obtained in the previous two steps and After addition, the ReLU activation function is used to obtain... .Right now

[0115]

[0116] Where b represents the bias term, This represents the ReLU activation function.

[0117] The image re-enters the CNN convolutional layer and undergoes multiple convolutions, ultimately yielding a binary image (the fifth binarized image). After averaging and fusion, the HC segmentation result, after error correction and calibration, is output, representing the final target shape. (Refer to...) Figure 10 The underlined area in the diagram represents the final target shape.

[0118] In summary, the following model calculation process for HC segmentation was obtained:

[0119] 1. Set the initial position of the evolution curve C, and set the level set function. Defined as a signed distance function, such as:

[0120]

[0121] Where r is the position of a pixel in the image, and c>0 is a constant. and These represent the inner and outer regions of evolution curve c, respectively.

[0122] 2. Initialize the local equilibrium distribution function Using equations Calculate relaxation parameters .

[0123] 3. Calculate the external force terms and use the equations. Discrete.

[0124] 4. In the equations respectively and Following the collisions and flows described in the text, the evolution curves are updated and .

[0125] 5. If the segmentation is not completed, return to step 2 and continue segmenting.

[0126] 6. Input the preliminary segmentation results into the error correction CNN to correct random errors or machine system errors and ensure accurate segmentation.

[0127] 7. Finally, output the segmentation results.

[0128] Using the above method, the shape of the object to be segmented can be modeled with a small number of training samples. Accurate segmentation results can be obtained without complex network structures. Sequence label alignment, shape priority, or target object registration are not required. The addition of shape prior technology makes the model closer to the standard case for segmentation, resulting in more accurate segmentation results. It can correctly segment partially occluded and missing targets and is more sensitive to changes in shape, size, and volume. Error correction CNN technology is used, which not only eliminates random errors but also eliminates systematic errors caused by the use of machine systems.

[0129] Figure 11A block diagram of a deep learning-based hippocampal segmentation device provided in this application embodiment. The device includes:

[0130] Memory 201; and processor 202 connected to the memory 201, the processor 202 being configured to: S1, acquire a first image generated by magnetic resonance imaging, infer the target shape in the first image using a trained deep belief network, and obtain a second image containing the target shape;

[0131] S2. Based on the second image, construct an energy function, obtain the lattice Boltzmann model driven by the deep belief network based on the energy function, and use the lattice Boltzmann model driven by the deep belief network to perform curve evolution on the first image to obtain the segmented third image and fourth image, wherein the fourth image contains the target shape;

[0132] S3. Repeat steps S1 to S2 multiple times, feed the obtained sets of third and fourth images into the error correction convolutional neural network model to obtain multiple fifth images, and perform average fusion on the multiple fifth images to obtain the segmentation result.

[0133] In some embodiments, the processor 202 is further configured to: infer the target shape in the first image using the trained deep belief network, a method comprising:

[0134] The deep belief network performs bottom-up reasoning on the input first image until it reaches the top hidden and visible layers. The inference model of the deep belief network is as follows:

[0135]

[0136] Where v is the vector representation of the input image, The label represents v, and h represents a hidden variable.

[0137] as well as ,in

[0138]

[0139] Represents the energy function of RBM. It is the bias vector. and It is a weight matrix;

[0140] Initialize the L layer and perform Gibbs sampling on the visible layers h2, h1 and L until convergence, thereby obtaining the target shape.

[0141] In some embodiments, the processor 202 is further configured to: before obtaining the deep belief network-driven lattice Boltzmann model based on the energy function, further comprising: constructing the lattice Boltzmann model, wherein the evolution equation of the lattice Boltzmann model is:

[0142]

[0143] in This represents the gray level of pixel r=(x,y) at time t. yes Gray-level distribution function on the directional sub-pixel, This represents the equilibrium distribution function describing the redistributed predicted values. It is the grid size of time. It is the relaxation time. It is along External force in direction, each pixel r contains its nine nearest neighbors, lattice vector The position of each sub-pixel is defined as:

[0144]

[0145] The lattice Boltzmann model consists of two steps, including a collision step and a flow step.

[0146] In some embodiments, the processor 202 is further configured to: the method for obtaining the lattice Boltzmann model driven by the deep belief network based on the energy function includes:

[0147] The energy function is minimized using the gradient descent flow method, resulting in a partial differential equation.

[0148] By comparing the evolution equation of the lattice Boltzmann model with the partial differential equation, the lattice Boltzmann model driven by the deep belief network is obtained.

[0149] In some embodiments, the processor 202 is further configured such that the error-correcting convolutional neural network model includes a replacement module, an optimization module, and convolutional layers;

[0150] After feeding the obtained sets of third and fourth images into the error-correcting convolutional neural network model, the process further includes: the third and fourth images are first processed by the replacement module, then processed by the optimization module, and then enter the convolutional layer to obtain the binarized fifth image.

[0151] For specific implementation methods, please refer to the aforementioned method embodiments, which will not be repeated here.

[0152] This application may be a method, apparatus, system, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of this application.

[0153] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0154] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0155] The computer program instructions used to perform the operations of this application may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuits, such as programmable logic circuits, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), are personalized by utilizing the status information of the computer-readable program instructions. These electronic circuits can execute the computer-readable program instructions to implement various aspects of this application.

[0156] Various aspects of this application are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0157] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processing unit of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0158] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions that execute on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0159] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0160] Note that, unless otherwise explicitly stated, all features disclosed in this specification (including any appended claims, abstract, and drawings) may be replaced by alternative features for achieving the same, equivalent, or similar purpose. Therefore, unless explicitly stated otherwise, each disclosed feature is merely one example of a set of equivalent or similar features. Where used, "further," "preferably," "even further," and "more preferably" are simple starting points for describing another embodiment based on the foregoing embodiments, the combination of which with the foregoing embodiments constitutes the complete configuration of another embodiment. Any combination of several "further," "preferably," "even further," or "more preferably" settings following the same embodiment constitutes yet another embodiment.

[0161] Although this application has been described in detail above with general descriptions and specific embodiments, some modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of this application fall within the scope of protection claimed in this application.

Claims

1. A deep learning-based hippocampal segmentation method, characterized in that, Includes the following steps: S1. Obtain a first image generated by nuclear magnetic resonance imaging, and use a trained deep belief network to infer the target shape in the first image to obtain a second image containing the target shape; S2. Based on the second image, construct an energy function; construct a lattice Boltzmann model, the evolution equation of which is: ,in This represents the gray level of pixel r=(x,y) at time t. yes Gray-level distribution function on the directional sub-pixel, This represents the equilibrium distribution function describing the redistributed predicted values. It is the grid size of time. It is the relaxation time. It is along External force in direction, each pixel r contains its nine nearest neighbors, lattice vector The position of each sub-pixel is defined as: The lattice Boltzmann model consists of two steps, including a collision step and a flow step; the deep belief network-driven lattice Boltzmann model is obtained based on the energy function, and the deep belief network-driven lattice Boltzmann model is used to perform curve evolution on the first image to obtain a segmented third image and a fourth image, wherein the fourth image contains the target shape, wherein the third image is a magnetic resonance image and the fourth image is a hippocampal image; S3. Repeat steps S1 to S2 multiple times to feed the obtained sets of third and fourth images into the error correction convolutional neural network model. The error correction convolutional neural network model includes a replacement module, an optimization module, and a convolutional layer. The third and fourth images are first processed by the replacement module, then processed by the optimization module, and then enter the convolutional layer to obtain multiple binarized fifth images. The multiple fifth images are averaged and fused to obtain the segmentation result.

2. The hippocampal segmentation method based on deep learning according to claim 1, characterized in that, The method for inferring the target shape in the first image using the trained deep belief network includes: The deep belief network performs bottom-up reasoning on the input first image until it reaches the top hidden and visible layers. The inference model of the deep belief network is as follows: Where v is the vector representation of the input image, The label represents v, and h represents a hidden variable. as well as ,in Represents the energy function of RBM. It is the bias vector. and It is a weight matrix; Initialize layer L, and perform Gibbs sampling on visible layers h2, h1, and L until convergence is achieved, thereby obtaining the target shape.

3. The hippocampal segmentation method based on deep learning according to claim 1, characterized in that, The method for obtaining the lattice Boltzmann model driven by the deep belief network based on the energy function includes: The energy function is minimized using the gradient descent flow method, resulting in a partial differential equation. By comparing the evolution equation of the lattice Boltzmann model with the partial differential equation, the lattice Boltzmann model driven by the deep belief network is obtained.

4. A hippocampal segmentation device based on deep learning, characterized in that, include: Memory; as well as A processor connected to the memory, the processor being configured to: S1. Obtain a first image generated by nuclear magnetic resonance imaging, and use a trained deep belief network to infer the target shape in the first image to obtain a second image containing the target shape; S2. Based on the second image, construct an energy function and a lattice Boltzmann model. The evolution equation of the lattice Boltzmann model is as follows: ,in This represents the gray level of pixel r=(x,y) at time t. yes Gray-level distribution function on the directional sub-pixel, This represents the equilibrium distribution function describing the redistributed predicted values. It is the grid size of time. It is the relaxation time. It is along External force in direction, each pixel r contains its nine nearest neighbors, lattice vector The position of each sub-pixel is defined as: The lattice Boltzmann model consists of two steps, including a collision step and a flow step; the deep belief network-driven lattice Boltzmann model is obtained based on the energy function, and the deep belief network-driven lattice Boltzmann model is used to perform curve evolution on the first image to obtain a segmented third image and a fourth image, wherein the fourth image contains the target shape, wherein the third image is a magnetic resonance image and the fourth image is a hippocampal image; S3. Repeat steps S1 to S2 multiple times to feed the obtained sets of third and fourth images into the error correction convolutional neural network model. The error correction convolutional neural network model includes a replacement module, an optimization module, and a convolutional layer. The third and fourth images are first processed by the replacement module, then processed by the optimization module, and then enter the convolutional layer to obtain multiple binarized fifth images. The multiple fifth images are averaged and fused to obtain the segmentation result.

5. The deep learning-based hippocampal segmentation device according to claim 4, characterized in that, The processor is also configured to: The method for inferring the target shape in the first image using the trained deep belief network includes: The deep belief network performs bottom-up reasoning on the input first image until it reaches the top hidden and visible layers. The inference model of the deep belief network is as follows: Where v is the vector representation of the input image, The label represents v, and h represents a hidden variable. as well as ,in Represents the energy function of RBM. It is the bias vector. and It is a weight matrix; Initialize layer L, and perform Gibbs sampling on visible layers h2, h1, and L until convergence is achieved, thereby obtaining the target shape.

6. The deep learning-based hippocampal segmentation device according to claim 5, characterized in that, The processor is also configured to: The method for obtaining the lattice Boltzmann model driven by the deep belief network based on the energy function includes: The energy function is minimized using the gradient descent flow method, resulting in a partial differential equation. By comparing the evolution equation of the lattice Boltzmann model with the partial differential equation, the lattice Boltzmann model driven by the deep belief network is obtained.

7. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a machine, it implements the steps of the method as described in any one of claims 1 to 3.

Images