Reconstruction method of organ three-dimensional image, intelligent equipment and storage medium

Abstract

The invention relates to the technical field of medical detection, in particular to an organ three-dimensional image reconstruction method, intelligent equipment and a storage medium. The method comprises the following steps: acquiring a target tomography image group of a target organ, and generating a target image matrix based on the target tomography image group; creating a standard human body finite element coordinate system, performing conductivity assignment on the target image matrix, and importing the assigned target image matrix into the standard human body finite element coordinate system to generate a conductivity matrix; obtaining a one-dimensional boundary vector of the conductance matrix, and converting the one-dimensional boundary vector into a three-dimensional voltage tensor; generating a target sample according to the three-dimensional voltage tensor, and inputting the target sample into a pre-trained U2-Net deep convolutional network to obtain a target three-dimensional pixel matrix; and generating a target three-dimensional image according to the target three-dimensional pixel matrix. According to the invention, the three-dimensional image of the target organ can be accurately presented in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of medical detection, and particularly to a method for reconstructing three-dimensional images of organs, an intelligent device, and a storage medium. Background Art

[0002] EIT (Electrical Impedance Tomography) is a medical imaging technology that generates images by measuring the impedance distribution inside an object, thereby providing information about the internal structure and tissue characteristics of the object.

[0003] Traditional EIT technology mainly performs image reconstruction based on a two-dimensional plane field. Although this method can ensure the imaging speed, it has limitations in capturing complex three-dimensional structures. Human organs have three-dimensional characteristics, and two-dimensional images cannot accurately reflect the three-dimensional structure of the organs. Since the positions of organ lesions are random, this can lead to insufficient understanding by doctors of the positions and sizes of organ lesions. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide a method for reconstructing three-dimensional images of organs, an intelligent device, and a storage medium to solve the problem in the prior art that the restored two-dimensional images cannot accurately reflect the three-dimensional structure of human organs, resulting in insufficient understanding by doctors of the positions and sizes of organ lesions.

[0005] The present invention discloses a method for reconstructing three-dimensional images of organs, including:

[0006] Obtaining a target tomographic image group of a target organ, and generating a target image matrix based on the target tomographic image group;

[0007] Creating a standard human finite element coordinate system, assigning conductivity values to the target image matrix, and importing the assigned target image matrix into the standard human finite element coordinate system to generate a conductivity matrix;

[0008] Obtaining a one-dimensional boundary vector of the conductivity matrix, and converting the one-dimensional boundary vector into a three-dimensional voltage tensor;

[0009] Generating a target sample according to the three-dimensional voltage tensor, and inputting the target sample into a pre-trained U 2 -Net deep convolutional network to obtain a target three-dimensional pixel matrix;

[0010] Generating a target three-dimensional image according to the target three-dimensional pixel matrix.

[0011] Optionally, the step of generating a target sample according to the three-dimensional voltage tensor includes:

[0012] Obtain the first length, width, and height information and the first number of channels information of the three-dimensional voltage tensor, and generate a four-dimensional tensor based on the first length, width, and height information and the first number of channels information, where the four-dimensional tensor is the target sample.

[0013] Optionally, before the step of inputting the target sample into the pre-trained U 2 -Net deep convolutional network, it includes:

[0014] Obtain multiple groups of historical tomographic image groups, and obtain multiple groups of historical image matrices based on the multiple groups of historical tomographic image groups;

[0015] Perform element elimination and / or scaling operations on each group of the historical image matrices to obtain multiple groups of augmented image matrices;

[0016] Obtain the boundary voltage of each group of the augmented image matrices, and generate training samples based on the boundary voltage and the augmented image matrices;

[0017] Input the training samples into the U 2 -Net deep convolutional network for training to obtain the pre-trained U 2 -Net deep convolutional network.

[0018] Optionally, the step of generating training samples based on the boundary voltage and the augmented image matrices includes:

[0019] Convert the boundary voltage into a three-dimensional boundary tensor, and obtain the second length, width, and height information and the second number of channels information of the three-dimensional boundary tensor;

[0020] Obtain the number of samples of the augmented image matrix, and generate a five-dimensional tensor based on the number of samples, the second number of channels information, and the second length, width, and height information;

[0021] Use the five-dimensional tensor and the corresponding augmented image matrix as a pair of training samples.

[0022] Optionally, before the step of inputting the training samples into the U 2 -Net deep convolutional network for training, it includes:

[0023] Reduce the number of channels of the residual U-block in the U 2 -Net deep convolutional network by a preset value.

[0024] Optionally, the step of generating the target image matrix based on the target tomographic image group includes:

[0025] Extract the points belonging to the target organ from the tomographic image group by the pixel threshold method to form an initial pixel matrix;

[0026] The sampling nearest neighbor interpolation method scales the initial pixel matrix to a preset size to obtain the image matrix.

[0027] Optionally, the step of importing the assigned target image matrix into the standard human body finite element coordinate system includes:

[0028] Assign conductivity to the target image matrix to obtain an assigned matrix;

[0029] Create an initial matrix in the standard human body finite element coordinate system according to preset coordinate information, where the preset coordinate information includes the central coordinate and / or boundary coordinate of the initial matrix;

[0030] Obtain the centroid of each tetrahedral mesh in the standard human body finite element coordinate system, take the centroid located in the initial matrix as the target centroid, and take the matrix composed of the target centroids as the matrix to be transformed;

[0031] Perform coordinate transformation on the matrix to be transformed to obtain a transformation matrix, and assign values to the centroids of each tetrahedral mesh in the transformation matrix according to the assigned matrix.

[0032] Optionally, the 2 U-Net deep convolutional network includes a convolutional layer, a normalization layer, and an activation layer. The convolutional kernel in the convolutional layer has a size of 3×3, a dilation coefficient of 1, and a padding of 1.

[0033] The present invention also discloses a computer-readable storage medium storing a computer program, which when executed by a processor causes the processor to execute the steps of the above-mentioned method.

[0034] The present invention also discloses an intelligent device including a memory and a processor. The memory stores a computer program, which when executed by the processor causes the processor to execute the steps of the above-mentioned method.

[0035] Compared with the prior art, the beneficial effect of the organ three-dimensional image reconstruction method provided by the embodiments of the present invention is that it combines the finite element method and deep learning technology, and can make full use of the advantages of both. It can not only perform accurate conductivity assignment based on a real physical model, but also efficiently extract features and reconstruct three-dimensional images through a deep learning network. This combination can improve the calculation efficiency while ensuring accuracy, meet the requirement of real-time imaging, and thus accurately present the three-dimensional image of the target organ in real time to assist doctors in accurate diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The following will further elaborate on the solution of the present invention in conjunction with the drawings and embodiments. In the drawings:

[0037] Figure 1 It is a schematic flowchart of an embodiment of the method for reconstructing a three-dimensional image of an organ provided by the present invention;

[0038] Figure 2 It is a schematic structural diagram of the U2-Net deep convolutional network provided by the present invention. Detailed implementation manners

[0039] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. Now, with reference to the accompanying drawings, the preferred embodiments of the present invention will be described in detail.

[0040] Please refer to Figure 1 , Figure 1 It is a schematic flowchart of an embodiment of the method for reconstructing a three-dimensional image of an organ provided by the present invention. The method for reconstructing a three-dimensional image of an organ provided by the present invention includes the following steps:

[0041] S101: Obtain a target tomographic image group of a target organ, and generate a target image matrix based on the target tomographic image group.

[0042] In a specific implementation scenario, to obtain a target tomographic image group of a target organ, CT (Computed Tomography) examination can be performed on the target organ of a patient (for example, the lung, pancreas, etc.), so as to obtain the target tomographic image group of the target organ, which includes a group of target tomographic images corresponding to different sections, and each target tomographic image is a two-dimensional cross-sectional image of the target organ.

[0043] Perform preprocessing on the obtained group of target tomographic images, such as removing noise and adjusting the size and resolution of the images. Perform image registration on the preprocessed target tomographic images, and ensure the alignment between different images based on the common feature points or feature structures of adjacent target tomographic images. Stack all the target tomographic images in an orderly manner according to the spatial position relationship to form a three-dimensional data set. This stacking can be performed along the vertical direction to form a stack along the Z axis, so as to form a three-dimensional volume data. In this three-dimensional volume data, mark the pixels corresponding to the target organ in each target tomographic image as 1, and mark the pixels in the remaining areas as 0, so as to generate a target image matrix.

[0044] When stacking images, interpolation operations may be required to fill the gaps between different layers to ensure a smooth transition in the generated three-dimensional image. Common interpolation methods include bilinear interpolation, trilinear interpolation, etc. In other implementation scenarios, specific reconstruction algorithms can also be used to process two-dimensional target tomographic images and generate three-dimensional images. Then, according to the three-dimensional image, the pixels corresponding to the target organ are labeled as 1, while the pixels in the remaining areas are labeled as 0, thus generating the target image matrix.

[0045] In one implementation scenario, the pixel threshold method can be used for each target tomographic image to identify and extract the pixels belonging to the target organ. Set a threshold and compare the gray value of the pixels in the image with this threshold to determine which pixels belong to the target organ. The pixels of the target organ are segmented from the entire three-dimensional image to form the initial pixel matrix of the target organ. The nearest neighbor interpolation method can be used to adjust this matrix to a preset size. The nearest neighbor interpolation method adjusts the size of the image by selecting the closest pixel value, which can preserve the shape of the target organ and will not change with the change of the matrix size. In this embodiment, the preset size is 360×360×360. In other implementation scenarios, other methods such as bilinear interpolation, bicubic interpolation, cubic interpolation, and spline interpolation can also be used to adjust this matrix to the preset size.

[0046] S102: Create a standard human finite element coordinate system, assign conductivity values to the target image matrix, and import the assigned target image matrix into the standard human finite element coordinate system to generate a conductivity matrix.

[0047] In a specific implementation scenario, to create a standard human finite element coordinate system, a corresponding standard human finite element coordinate system can be created according to the location of the target organ. For example, if the target organ is the lung, the standard human finite element coordinate system is the thoracic cavity finite element coordinate system; another example is that if the target organ is the stomach, the standard human finite element coordinate system is the abdominal cavity finite element coordinate system.

[0048] After determining the human body area corresponding to the standard human finite element coordinate system to be created, obtain the shape and size information of this human body area according to human anatomy knowledge or existing medical image data to construct a three-dimensional human body area. Create a coordinate system in the three-dimensional human body area, and make the X-axis, Y-axis, and Z-axis of the coordinate system correspond to the front-back, left-right, and up-down directions of the human body respectively to conform to the standards of human anatomy. Divide the three-dimensional human body area into small grids according to the coordinate system, such as tetrahedral grids or hexahedral grids.

[0049] Conductivity assignment is performed on the target image matrix. As described above, the target image matrix is a three-dimensional matrix composed of "0" and "1". The elements of "1" are replaced with conductivity values to achieve conductivity assignment and obtain the assigned matrix. The specific value of the conductivity can be randomly selected from the interval [0.05, 0.15] S / m.

[0050] In the standard human finite element coordinate system, an initial matrix is created according to the preset coordinate information and denoted as M1. The preset coordinate information includes the central coordinates and / or boundary coordinates of the initial matrix. For example, taking (0, 0, 180) as the center point, the coordinate range is within ((-179~180), (-179~180), (21~380)). The size of such an initial matrix is 360×360×360, and the size of the assigned matrix is also 360×360×360, and the two match. The corresponding relationship between each element of the initial matrix M1 and the assigned matrix can be obtained through the interpolation algorithm. By applying the interpolation idea, the processing and matching of complex spatial data can be realized in the finite element model, thereby improving the accuracy and reliability of the model.

[0051] In the finite element model, a tetrahedral mesh is a solid geometric unit composed of four nodes. For each tetrahedral mesh, its centroid is calculated and denoted as V points (x i ,y i ,z i ,v i ), v i is the serial number of the tetrahedral mesh corresponding to the centroid. The centroids of the tetrahedral meshes located in the initial matrix M1 are extracted, rounded, and used as the target centroids, denoted as M points (x i ,y i ,z i ,m i ), where m i is the serial number of the target centroid M points .

[0052] The matrix composed of the target centroids is used as the matrix to be transformed. Coordinate transformation is performed on each M points in the matrix to be transformed, x i +180, y i +180, z i -20, thereby obtaining the transformation matrix. It can be understood that n iIt will not change, only the position of the matrix changes. According to the assignment matrix, the element numbers of each tetrahedral mesh in the transformation matrix are assigned to obtain the conductance matrix. The centroid of each tetrahedral mesh in the transformation matrix is corresponded to each element in the assignment matrix. If the element corresponding to a centroid is "0", the parameters of this centroid are not modified. If the element corresponding to a centroid is the conductivity and not "0", then the value of this centroid is modified to the corresponding conductivity value. In this way, the conductivity in the assignment matrix is imported into the standard human finite element coordinate system.

[0053] S103: Obtain the one-dimensional boundary vector of the conductance matrix and convert the one-dimensional boundary vector into a three-dimensional voltage tensor.

[0054] EIDORS (Electrical Impedance and Diffuse Optical Reconstruction Software) is used to simulate and reconstruct the impedance and optical properties inside biological tissues. In a specific implementation scenario, the conductance matrix is imported into EIDORS. Based on the conductance matrix, the positions, shapes, and conductive parameters of the electrodes are defined. The excitation and measurement modes are defined through stim_pattern and meas_pattern. The excitation electrodes for applying current and the measurement electrodes for measuring voltage are defined. In the adjacent excitation and adjacent measurement mode, the excitation electrodes and the measurement electrodes are adjacent. Using the EIT forward model in EIDORS, according to the electrode configuration and the geometry of the object, the boundary voltage response caused by the current applied by the excitation electrodes is calculated. According to the boundary voltage response calculated by the forward model, the actual data acquisition process is simulated to obtain a complete boundary voltage data set, which is the one-dimensional boundary vector.

[0055] In electrical impedance tomography (EIT), the boundary voltage data is usually represented as a one-dimensional vector (one-dimensional boundary vector), where each element corresponds to the voltage response on a measurement electrode. If this one-dimensional boundary voltage vector needs to be converted into a three-dimensional boundary voltage tensor, a concatenation layer can be used to achieve this conversion. Determine the dimensions of the three-dimensional tensor to be created. In this case, the tensor can be regarded as a matrix, where the number of rows represents the measurement time points and the number of columns represents different measurement electrodes.

[0056] Reshape the one-dimensional boundary voltage vector into a matrix such that each row corresponds to the voltage response at a measurement time point. In this matrix, each row represents the measurement result at a time point. To convert it into a three-dimensional tensor, an additional dimension, such as the time dimension, can be added to represent the measurement results at different time points. Use the concatenation layer in a deep learning framework (such as TensorFlow, PyTorch, etc.) to concatenate these matrices along the time dimension, thus forming a three-dimensional tensor. (Three-dimensional voltage tensor) In this way, the measurement result at each time point will become a slice in the tensor.

[0057] S104: Generate a target sample based on the three-dimensional voltage tensor, and input the target sample into the pre-trained U 2 -Net deep convolutional network to obtain the target three-dimensional pixel matrix.

[0058] In a specific implementation scenario, generate a target sample based on the three-dimensional voltage tensor. The target sample includes the information of the three-dimensional voltage tensor. Input the target sample into the pre-trained U 2 -Net deep convolutional network. The pre-trained U 2 -Net deep convolutional network is the trained U 2 -Net deep convolutional network. When training, the input is the pairwise corresponding three-dimensional voltage tensor and three-dimensional pixel matrix. Therefore, the pre-trained U 2 -Net deep convolutional network can output the target three-dimensional pixel matrix corresponding to the three-dimensional voltage tensor. The target three-dimensional pixel matrix includes two elements, "0" and "1". Among them, "0" indicates that the position corresponds to the area where the non-target organ is located, while "1" indicates that the position corresponds to the area where the target organ is located.

[0059] In an implementation scenario, the target sample is a four-dimensional tensor. Its four dimensions include the length, width, and height information of the three-dimensional voltage tensor and the number of channels of its U 2 -Net deep convolutional network, which is the first channel number information. Each element represents the voltage response at a time point or measurement time. In this embodiment, the first channel number information is 1. In other implementation scenarios, the first channel number information can also be other positive integers.

[0060] In this embodiment, U 2- The -Net network includes a convolutional layer, a normalization layer, and an activation layer. The convolutional kernel in the convolutional layer has a size of 3×3, a dilation coefficient of 1, and a padding of 1. A dilation coefficient of 1 means that each element in the convolutional kernel is continuous on the input feature map. The dilation coefficient is used to control the spacing between elements in the convolutional kernel. When the dilation coefficient is 1, the convolution operation is a standard convolution operation. A padding of 1 means adding a layer of zero padding to the edges of the input feature map. This can keep the size of the feature map unchanged after the convolution operation. Specifically, if the size of the input feature map is H×W, after the convolution operation with a 3×3 convolutional kernel, a dilation coefficient of 1, and a padding of 1, the size of the output feature map is still H×W.

[0061] Target sample Where C is the first channel number information, H is the height information of the three-dimensional voltage tensor, W is the width information of the three-dimensional voltage tensor, and L is the length information of the three-dimensional voltage tensor.

[0062] After passing the target sample X through the convolutional layer, we get: x = W(X).

[0063] The convolutional layer includes two channels, and the mean and variance of each channel are expressed as:

[0064]

[0065]

[0066] After x passes through the normalization layer, we get:

[0067]

[0068] Where ∈ is a constant.

[0069] Performing scaling and translation operations on Y, we get:

[0070] Z[c,h,w,l] = γ c ·Y[c,h,w,l] + β c

[0071] Where γ c and β c are learnable scaling and translation reconstruction parameters.

[0072] Then, passing through the activation layer, we get

[0073] Z′ = ReLU(Z) = max(0,Z)

[0074] Where ReLU is the activation function.

[0075] U 2- The -Net deep convolutional network connects the corresponding layers between the encoder (downsampling path) and the decoder (upsampling path) to help information transfer better in the network and maintain gradient stability. For example, in this embodiment, there are five downsampling layers. The downsampling operation causes the size of the feature map to decrease, the number of channels to increase, and the number of features to increase accordingly. After 5 downsamplings, the size of the feature map has been reduced to the minimum, while the number of features reaches the maximum, enabling better feature learning and multi-layer feature fusion to be achieved.

[0076] Through multi-layer feature fusion, the network can capture richer semantic and spatial information, thereby improving the understanding and expression capabilities. A four-dimensional tensor with two channels is obtained through the convolutional layer. Using an independent one-hot encoding, the largest number at the corresponding position can be extracted on the two channels. The channel represents the result corresponding to the target organ area at that position. At the last step of the network, using the ReLU activation function helps introduce non-linearity, limits the output of the network to the non-negative range, and at the same time maintains the sparsity of the network, which helps improve the expression and generalization capabilities of the network.

[0077] S105: Generate a target three-dimensional image according to the target three-dimensional pixel matrix.

[0078] In a specific implementation scenario, a corresponding target three-dimensional image is generated according to the obtained target three-dimensional pixel matrix. For example, a volume rendering algorithm can be used to convert the three-dimensional pixel matrix into a visualized three-dimensional image. Pixels in the target three-dimensional pixel matrix are sampled in three-dimensional space to obtain density values or other attributes. An image is generated by projecting light rays along the line-of-sight direction in three-dimensional space and calculating the transmission and color of the light rays based on the sampled pixel values. The sampled pixel values are synthesized into the final image, usually by integrating the light rays.

[0079] In other implementation scenarios, an isosurface extraction method can also be used to connect points with the same value in the target three-dimensional pixel matrix to form a surface to display the shape and structure of the data.

[0080] In other implementation scenarios, since the size of the target three-dimensional pixel matrix may be small and the information it contains is also less, the obtained target three-dimensional image may not be clear enough and lack details. The target three-dimensional pixel matrix can be first expanded by the interpolation method and then the target three-dimensional image is generated.

[0081] As can be seen from the above description, in this embodiment, a standard human body finite element coordinate system is created, the conductivity is assigned to the target image matrix, and the assigned target image matrix is imported into the standard human body finite element coordinate system to generate a conductivity matrix, which can more accurately simulate the conductivity distribution of human tissues, U 2-Net is a deep convolutional network with powerful feature extraction and image segmentation capabilities. By inputting the target samples into the pre-trained U2-Net network, the three-dimensional pixel matrix of the target can be effectively extracted from the complex voltage tensor data, realizing high-quality three-dimensional image reconstruction. Since the U 2 -Net deep convolutional network has been pre-trained on a large amount of data. After inputting the target samples, the pre-trained U 2 -Net deep convolutional network can quickly process the data and extract the three-dimensional pixel matrix of the target, thus reducing the computational time cost.

[0082] In summary, this embodiment combines the finite element method and deep learning technology, which can make full use of the advantages of both. It can not only perform accurate conductivity assignment based on a real physical model, but also efficiently extract features and reconstruct three-dimensional images through a deep learning network. This combination can improve the computational efficiency while ensuring accuracy, meet the requirement of real-time imaging, and thus present the three-dimensional image of the target organ in real time and accurately to assist doctors in accurate diagnosis.

[0083] In an implementation scenario, the U 2 -Net deep convolutional network needs to be trained before use, so as to learn rich feature representations through a large amount of data. These features can better capture image information at different scales and levels. This feature learning helps the network better understand the image content, improve the performance of the network in various image processing tasks, and accelerate the convergence speed of the model.

[0084] Collect multiple groups of historical tomographic image groups. It can be understood that these historical tomographic image groups correspond to the same target organs of different people, such as the lungs of multiple patients, or the pancreas of multiple patients, etc. Based on multiple groups of historical tomographic image groups, multiple groups of historical image matrices are obtained. The method for obtaining the historical image matrices is basically the same as the steps for generating the target image matrix based on the target tomographic image group in the above text, and will not be elaborated here.

[0085] Perform element elimination and / or scaling operations on each group of historical image matrices to obtain multiple groups of augmented image matrices. Element elimination refers to combined excision according to the distribution characteristics of the target organ, and setting the pixels in the area to be eliminated as null values. This operation can simulate the missing data or changes that may occur in actual situations, and helps the training model better handle the situation of missing information. For example, if the target organ is the lung, combined excision is performed according to the five lung lobe regions of the human body. Scaling is achieved by the nearest neighbor interpolation method, which estimates the values of unknown pixels based on the values of known pixels. In this case, scaling may change the resolution or size of the pixels, thus increasing the data diversity. This helps the model learn features at different resolutions or sizes.

[0086] Obtain the boundary voltage of each group of augmented image matrices, and generate training samples based on the boundary voltage and the augmented image matrices; the boundary voltage is a one-dimensional vector, and the steps for obtaining the boundary voltage are basically the same as the steps for obtaining the one-dimensional boundary vector of the conductance matrix in the present invention, which will not be elaborated here. Generate training samples based on the boundary voltage and the augmented image matrices. The content of the training samples includes the corresponding three-dimensional pixel matrix (the augmented image matrix in this implementation scenario) and the corresponding voltage information (which can be obtained according to the boundary voltage). Input such training samples into the U 2 -Net deep convolutional network for training, so that in actual use, an accurate target three-dimensional pixel matrix can be obtained according to the three-dimensional voltage tensor.

[0087] Furthermore, the boundary voltage obtained by the steps in the above text can be converted into a three-dimensional boundary tensor. To achieve the correspondence between the three-dimensional boundary tensor and the augmented image matrix, the sizes of the three-dimensional boundary tensor and the augmented image matrix are the same, both being 64×64×64. At this time, the augmented image matrix is smaller than the preset size of 360×360×360. The augmented image matrix of 360×360×360 can be shrunk using the interpolation method to make its size meet 64×64×64.

[0088] Obtain the second length, width, and height information and the second number of channels information of the three-dimensional boundary tensor; obtain the number of samples of the augmented image matrix, and generate a five-dimensional tensor based on the number of samples, the second number of channels information, and the second length, width, and height information. Where N is the number of samples. The reason for including the number of samples in the five-dimensional tensor is that when training a deep learning model, the appropriate batch size is usually selected according to the computing resources and the requirements of the model (including the number of samples). Through parallel computing, the GPU can process multiple samples simultaneously, accelerating the training speed. When the batch size is 32, it means that in each training iteration, the GPU will process 32 groups of data simultaneously, which helps to improve the training efficiency and utilize the parallel computing ability of the GPU. Use the five-dimensional tensor and the corresponding resized augmented image matrix in the above text as a pair of training samples.

[0089] As can be seen from the above description, adding the number of samples to the five-dimensional tensor during training can enable the GPU to process tensor data more efficiently. By training on the five-dimensional tensor, the parallel computing ability of the GPU can be utilized to accelerate the training process and improve the training efficiency.

[0090] In one implementation scenario, reduce the number of channels of the residual U-block in the U 2 -Net deep convolutional network by a preset value. Please refer to Figure 2 , Figure 2 which is the U provided by the present invention. 2-Schematic diagram of the structure of the -Net deep convolutional network. By reducing the number of channels in the RSU (Residual U-Block), the number of input and output channels of each convolutional layer can be reduced, thereby reducing the number of parameters and the computational burden. Specifically, each channel corresponds to a convolutional kernel, and reducing the number of channels will reduce the number of parameters in each convolutional layer. Due to the reduction in the number of parameters, the complexity of the model will be reduced, making the model more concise and lightweight. The feature map of each channel needs to undergo a convolutional operation, and reducing the number of channels will reduce the computational amount of each convolutional layer. This can reduce the computational burden of the model during the training and inference stages and speed up the running speed of the model. The reduction in the number of channels also helps to reduce the complexity of the model and the risk of overfitting. Reducing the number of channels can also reduce the complexity of the model and accelerate the training and inference speed of the model.

[0091] The present invention also provides an intelligent device, which includes a processor and a memory. The processor is coupled to the memory. A computer program is stored in the memory, and the processor executes the computer program during operation to implement the above method. For detailed steps, refer to the above, and details will not be elaborated here.

[0092] The present invention also provides a computer-readable storage medium. At least one computer program is stored in the computer-readable storage medium, and the computer program is used to be executed by the processor to implement the above method. For detailed steps, refer to the above, and details will not be elaborated here. In one embodiment, the computer-readable storage medium can be a storage chip in the terminal, a hard disk, a mobile hard disk, a USB flash drive, an optical disc, or other writable and readable storage tools, or it can also be a server, etc.

[0093] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above various methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0094] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

[0095] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. For those skilled in the art, the technical solutions described in the above embodiments can be modified, or some of the technical features can be equivalently replaced; and all such modifications and replacements should fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for reconstructing a three-dimensional image of an organ, characterized in that: include: Acquire a target tomographic image group of a target organ, and generate a target image matrix based on the target tomographic image group; Creating a standard human finite element coordinate system, assigning conductivity to the target image matrix, importing the assigned target image matrix into the standard human finite element coordinate system, and generating a conductivity matrix; Obtaining a one-dimensional boundary vector of the conductance matrix, and converting the one-dimensional boundary vector into a three-dimensional voltage tensor; Generate a target sample according to the three-dimensional voltage tensor, and input the target sample into the pre-trained U 2 -Net deep convolutional network to obtain the target three-dimensional pixel matrix; A target three-dimensional image is generated according to the target three-dimensional pixel matrix.

2. The method for reconstructing a three-dimensional image of an organ according to claim 1, characterized in that: The step of generating a target sample according to the three-dimensional voltage tensor comprises: First length, width, and height information and first channel number information of the three-dimensional voltage tensor are obtained, and a four-dimensional tensor is generated based on the first length, width, and height information and the first channel number information, where the four-dimensional tensor is the target sample.

3. The method for reconstructing a three-dimensional image of an organ according to claim 1, characterized in that: The target sample is input into the pre-trained U 2 -Net deep convolutional network steps, including: Acquire multiple groups of historical tomographic image groups, and acquire multiple groups of historical image matrices based on the multiple groups of historical tomographic image groups; Performing element elimination and / or scaling operations on each group of the historical image matrices to obtain multiple groups of expanded image matrices; Acquire the boundary voltage of each group of the expanded image matrix, and generate a training sample based on the boundary voltage and the expanded image matrix; Input the training samples into U 2 -Net deep convolutional network is trained to obtain the pre-trained U 2 -Net deep convolutional network.

4. The method for reconstructing a three-dimensional image of an organ according to claim 3, characterized in that: The step of generating training samples based on the boundary voltage and the expanded image matrix comprises: Convert the boundary voltage into a three-dimensional boundary tensor, and obtain second length, width, and height information and second channel number information of the three-dimensional boundary tensor; Obtaining the number of samples of the expanded image matrix, and generating a five-dimensional tensor based on the number of samples, the second channel number information, and the second length, width, and height information; The five-dimensional tensor and the corresponding expanded image matrix are used as a pair of training samples.

5. The method for reconstructing a three-dimensional image of an organ according to claim 3, characterized in that: The training samples are input into U 2 -Net deep convolutional network training steps include: The U 2 -Net The number of channels of the residual U block in the deep convolutional network is reduced by a preset value.

6. The method for reconstructing a three-dimensional image of an organ according to claim 1, characterized in that: The step of generating a target image matrix based on the target tomographic image group comprises: Extracting points belonging to the target organ from the tomographic image group by a pixel threshold method to form an initial pixel matrix; The initial pixel matrix is ​​scaled to a preset size by using a sampling nearest neighbor interpolation method to obtain the image matrix.

7. The method for reconstructing a three-dimensional image of an organ according to claim 1, characterized in that: The step of importing the assigned target image matrix into the standard human body finite element coordinate system comprises: Assigning conductivity to the target image matrix to obtain an assignment matrix; Creating an initial matrix in the standard human body finite element coordinate system according to preset coordinate information, wherein the preset coordinate information includes the center coordinates and / or boundary coordinates of the initial matrix; Obtaining the centroid of each tetrahedral mesh in the standard human body finite element coordinate system, taking the centroid located in the initial matrix as the target centroid, and taking the matrix formed by the target centroid as the matrix to be transformed; A transformation matrix is ​​obtained by performing coordinate transformation on the matrix to be transformed, and a value is assigned to the centroid of each tetrahedral mesh in the transformation matrix according to the assignment matrix.

8. The method for reconstructing a three-dimensional image of an organ according to claim 1, characterized in that: The U 2 The -Net deep convolutional network includes a convolutional layer, a normalization layer and an activation layer. The convolution kernel size in the convolutional layer is 3×3, the expansion coefficient is 1, and the padding is 1.

9. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 8.

10. A smart device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 8.

Images