An image processing method, device and storage medium

Abstract

Embodiments of the present application provide an image processing method, device and storage medium. In the method, the obtained brain image is input to a feature extraction layer in a brain image processing model, the brain image is subjected to multi-dimensional down-sampling processing to obtain low-level feature maps of multiple dimensions of the brain image, and the low-level feature maps represent low-level semantic information. The low-level feature maps of multiple dimensions are input to a brain region segmentation layer, multi-dimensional up-sampling processing is performed based on the low-level feature maps of the maximum dimension, in the up-sampling process, the low-level feature maps of multiple dimensions obtained by up-sampling are fused with the low-level feature maps based on a skip connection to obtain a global feature map; and the brain image is segmented and labeled according to the global feature map. In this way, the brain region segmentation and labeling can be performed based on the global feature of the entire brain image, the brain image does not need to be segmented into image blocks, and thus the loss of local information between the image blocks can be avoided, and the accuracy of the segmentation and labeling is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image segmentation, and in particular to an image processing method, device and storage medium. BACKGROUND

[0002] Head CT (Computed Tomography, CT for short) image is one of the most important and most commonly used examinations in clinical practice, and is of great significance for disease diagnosis. Based on the head CT image, image segmentation is performed to obtain a plurality of anatomically labeled regions of interest, which can be used for image analysis process and clinical auxiliary diagnosis.

[0003] In the prior art, the head CT image can be divided into a plurality of image blocks, and the plurality of image blocks are input into the processing unit of the deep learning model to perform a segmentation task. However, in this way, the deep learning model cannot perceive the local spatial information between the image blocks, that is, the local spatial information between the image blocks is lost, resulting in poor accuracy of the segmentation result. Therefore, a solution is urgently needed. SUMMARY

[0004] Aspects of the present application provide an image processing method, device and storage medium to improve the accuracy of segmentation labeling.

[0005] The embodiment of the present application provides an image processing method, comprising: acquiring a brain image; inputting the brain image into a feature extraction layer in a pre-trained brain image processing model, performing multi-dimensional down-sampling processing on the brain image to obtain low-level feature maps of multiple dimensions of the brain image, the low-level feature maps representing low-level semantic information; inputting the low-level feature maps of multiple dimensions into a brain region segmentation layer in the brain image processing model, performing multi-dimensional up-sampling processing based on the low-level feature maps of the maximum dimension, and in the up-sampling processing, fusing the high-level feature maps of multiple dimensions obtained by up-sampling and the low-level feature maps based on the skip connection technology to obtain a first global feature map; and segmenting and labeling the brain regions of the brain image according to the first global feature map.

[0006] Further optionally, the method further comprises: inputting the low-level feature maps of multiple dimensions into an image reconstruction layer in the brain image processing model, performing multi-dimensional up-sampling processing based on the low-level feature maps of the maximum dimension, and in the up-sampling processing, fusing the high-level feature maps of multiple dimensions obtained by up-sampling and the low-level feature maps based on the skip connection technology to obtain a second global feature map, and reconstructing the image according to the second global feature map to obtain the brain image labeled with each brain region.

[0007] Further optionally, the brain image is subjected to multi-dimension down-sampling processing to obtain multi-dimension low-level feature maps of the brain image, including: in the feature extraction layer, the brain image is subjected to multi-dimension down-sampling processing by using stacked convolution layers, non-linear activation functions and max-pooling layers to obtain multi-dimension low-level feature maps.

[0008] Further optionally, the brain image is subjected to multi-dimension down-sampling processing to obtain multi-dimension low-level feature maps in the feature extraction layer by using stacked convolution layers, non-linear activation functions and max-pooling layers, including: in the feature extraction layer, the brain image is subjected to convolution processing by using multiple stacked convolution layers in sequence with the number of channels from small to large; in each stacked convolution layer, each convolution result is subjected to non-linear processing by using a non-linear activation function to obtain a low-level feature map of a corresponding dimension; and the low-level feature map output by a previous stacked convolution layer is subjected to pooling processing by using a max-pooling layer between adjacent stacked convolution layers to obtain the input of a next stacked convolution layer.

[0009] Further optionally, in each stacked convolution layer, each convolution result is subjected to non-linear processing by using a non-linear activation function to obtain a low-level feature map of a corresponding dimension, including: a convolution kernel window with a size of n*n*n is used to slide from the top-left corner of input data, and each sliding step is 1 voxel distance, to extract abstract features in the input data, n is a positive integer, and the input data is the brain image or a low-level feature map after pooling processing; for each position of the convolution kernel window, the overlapping part of the convolution kernel and the input data is multiplied element by element, and the sum of the products is obtained to obtain a scalar value; the scalar value is subjected to max-pooling operation with a step of m and non-linear processing based on a non-linear activation function to obtain a feature value corresponding to the convolution kernel; and the feature values corresponding to all convolution kernels form a low-level feature map with the same number of convolution kernels.

[0010] Further optionally, the multi-dimension up-sampling processing is performed based on the low-level feature map with the largest dimension, and in the up-sampling processing, the multi-dimension high-level feature maps obtained by up-sampling are fused with the low-level feature maps based on the skip connection technology to obtain a first global feature map, including: in the brain region segmentation layer, the multi-dimension up-sampling processing is performed based on the low-level feature map with the largest dimension by using a deconvolution layer, a non-linear activation function and a convolution layer, and in the up-sampling processing, the multi-dimension high-level feature maps obtained by up-sampling are fused with the low-level feature maps based on the skip connection technology to obtain a first global feature map.

[0011] Further optionally, in the brain region segmentation layer, a deconvolution layer, a nonlinear activation function and a convolution layer are adopted to perform multi-dimensional up-sampling processing based on a low-layer feature map with a maximum dimension, and during the up-sampling processing, the multi-dimensional high-layer feature maps obtained through the up-sampling are fused with the low-layer feature map based on a skip connection technology to obtain a first global feature map, including: in the brain region segmentation layer, a convolution layer is adopted to perform convolution processing on the low-layer feature map with the maximum dimension, and a nonlinear activation function is adopted to perform nonlinear processing on the convolution result to obtain a reference feature map; a deconvolution layer is adopted to perform multiple deconvolution processing on the reference feature map to obtain an output feature map of each deconvolution processing, and a residual connection is used to add an input feature map of each deconvolution processing to an output feature map to obtain a high-layer feature map of a corresponding dimension, the input feature map being the reference feature map or a target feature map obtained through last deconvolution processing; the high-layer feature map obtained through each deconvolution processing is connected with the low-layer feature map of the corresponding dimension through a skip connection to obtain an intermediate state feature map, and a convolution layer is used to down-sample the intermediate state feature map to obtain a target feature map with the same number of channels as the input feature map, the target feature map being an input feature map of next deconvolution processing or being the first global feature map.

[0012] Further optionally, according to the first global feature map, the brain image is segmented and labeled in brain regions, including: inputting the first global feature map into a fully connected layer to map a plurality of voxels in the first global feature map to a preset class label; according to the class label of the plurality of voxels, at least one voxel with the same class label in the plurality of voxels is divided into a same brain region; and the plurality of brain regions obtained through the division are respectively labeled with corresponding class labels.

[0013] The embodiment of the present application also provides an electronic device, including: a memory, a processor; wherein the memory is used to store one or more computer instructions; the processor is used to execute the one or more computer instructions to perform the steps in the image processing method.

[0014] The embodiment of the present application also provides a computer readable storage medium storing a computer program, when the computer program is executed by a processor, the processor can implement the steps in the image processing method.

[0015] In this embodiment, the obtained brain image can be input to the feature extraction layer in the brain image processing model, and the brain image is subjected to multi-dimensional down-sampling processing to obtain low-layer feature maps of the brain image in multiple dimensions, which represent low-layer semantic information. The low-layer feature maps in multiple dimensions are input to the brain region segmentation layer, and multi-dimensional up-sampling processing is performed based on the low-layer feature maps with the maximum dimension. In the up-sampling process, the high-layer feature maps in multiple dimensions obtained by up-sampling are fused with the low-layer feature maps based on the skip connection to obtain a global feature map. According to the global feature map, the brain image is segmented and labeled in the brain region. In this way, the brain region segmentation and labeling can be performed based on the global features of the entire brain image, without the need to segment the brain image into image blocks, thereby avoiding the loss of local information between image blocks and improving the accuracy of segmentation and labeling. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and serve to explain the principles of the application, and do not limit the application. In the drawings:

[0017] Figure 1 A flowchart of an image processing method provided for an exemplary embodiment of the application is shown;

[0018] Figure 2 A schematic diagram of adaptive learning provided for an exemplary embodiment of the application is shown;

[0019] Figure 3 A variation diagram of a loss function provided for an exemplary embodiment of the application is shown;

[0020] Figure 4 A schematic diagram of an electronic device provided for an exemplary embodiment of the application is shown. DETAILED DESCRIPTION

[0021] To make the objectives, technical solutions, and advantages of the application clearer, the technical solutions of the application will be described below with reference to the embodiments of the application and the corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the application, but not all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the application.

[0022] In the prior art, a head CT image can be usually divided into multiple image blocks, and the multiple image blocks are input to the processing unit of a deep learning model to perform a segmentation task. However, in this way, the deep learning model cannot perceive the local spatial information between the image blocks, that is, the local spatial information between the image blocks is lost, thereby resulting in poor accuracy of the segmentation result.

[0023] In some embodiments of the present application, a solution is provided to solve the above technical problems. The technical solutions provided by the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0024] Figure 1 A flowchart of an image processing method provided by an exemplary embodiment of the present application is shown in FIG. 1. As shown in FIG. 1, the method comprises the following steps. Figure 1

[0025] Step 11, obtaining a brain image.

[0026] Step 12, inputting the brain image into a feature extraction layer in a pre-trained brain image processing model to perform multi-dimensional down-sampling processing on the brain image to obtain low-level feature maps of multiple dimensions of the brain image, the low-level feature maps representing low-level semantic information.

[0027] Step 13, inputting the low-level feature maps of multiple dimensions into a brain region segmentation layer in the brain image processing model to perform multi-dimensional up-sampling processing based on the low-level feature maps of the maximum dimension, and in the up-sampling processing, fusing the high-level feature maps of multiple dimensions obtained by up-sampling with the low-level feature maps based on a skip connection technology to obtain a first global feature map.

[0028] Step 14, segmenting and labeling the brain regions of the brain image according to the first global feature map.

[0029] The embodiment can be executed by an electronic device. The electronic device can be a medical device, a computer, a mobile phone, a server device, or a tablet computer, etc., and the present embodiment is not limited thereto. The following will be exemplarily described taking the electronic device as the execution subject.

[0030] In the present embodiment, the brain image is a brain image obtained by performing CT on a human brain. The brain image can be a two-dimensional image or a three-dimensional image, and the present embodiment is not limited thereto.

[0031] The electronic device can input the brain image into a pre-trained brain image processing model for corresponding image processing. The brain image processing model comprises a feature extraction layer, a brain region segmentation layer, and an image reconstruction layer. The feature extraction layer is used to extract image features in the brain image, the brain region segmentation layer is used to segment and label different brain regions in the brain image based on the feature maps, and the image reconstruction layer is used to reconstruct the image based on the feature maps. The feature extraction layer and the brain region segmentation layer are connected by a skip connection, and the structure based on the skip connection can jump between different levels in the brain image processing model to transmit information between different levels, i.e., the output of the feature extraction layer can be used as the input of the brain region segmentation layer. ​

[0032] The brain image processing model can utilize the feature extraction layer to perform multi-dimensional down-sampling processing on the brain image to obtain low-level feature maps of the brain image in multiple dimensions. The low-level feature maps are used to represent low-level semantic information, which includes at least one of the following: contour features, edge features, color features, brightness features, and shape features.

[0033] The down-sampling processing refers to reducing the brain image and merging the pixel points, reducing the dimension of the features and reducing the feature extraction amount of the brain image, so as to obtain low-level feature maps with smaller resolution. The dimension can correspond to different channel numbers in the feature extraction layer, for example, the first dimension can correspond to a channel number of 32, the second dimension can correspond to a channel number of 64, and the third dimension can correspond to a channel number of 128. The larger the channel number, the higher the dimension of the down-sampling processing, and the larger the feature amount in the obtained low-level feature map.

[0034] The down-sampling processing of the brain image in any dimension can obtain the low-level feature map of the brain image in the dimension. For example, the feature extraction layer can utilize 32 channels to perform first-dimensional down-sampling processing on the brain image to obtain a first-dimensional low-level feature map of the brain image, utilize 64 channels to perform second-dimensional down-sampling processing on the brain image to obtain a second-dimensional low-level feature map of the brain image, and utilize 128 channels to perform third-dimensional down-sampling processing on the brain image to obtain a third-dimensional low-level feature map of the brain image.

[0035] After the feature extraction layer performs multi-dimensional down-sampling processing on the brain image, the brain region segmentation layer can perform multi-dimensional up-sampling processing based on the low-level feature map with the largest dimension. The up-sampling processing refers to enlarging the brain image and inserting pixel values between the pixel points, increasing the dimension of the features and increasing the feature extraction amount of the brain image, so as to obtain low-level feature maps with larger resolution. Alternatively, the up-sampling processing can be bilinear interpolation processing or deconvolution processing, which is not limited in the embodiment.

[0036] The brain region segmentation layer can select the low-level feature map with the largest dimension from the low-level feature maps of the brain image in multiple dimensions, and perform multi-dimensional up-sampling processing based on the low-level feature map. During the up-sampling processing, the brain region segmentation layer can fuse the high-level feature maps of multiple dimensions obtained by up-sampling with the low-level feature map based on the skip connection technology to obtain a first global feature map.

[0037] The skip connection technology is used to connect different layers in the brain image processing model. Based on the skip connection technology, the high-level feature maps of multiple dimensions obtained by upsampling can be fused with the low-level feature maps. The high-level feature maps of multiple dimensions obtained by upsampling can be fused with the low-level feature maps in a feature summation, feature splicing, or corresponding element multiplication manner, and the present embodiment is not limited.

[0038] Taking the foregoing example as an example, the low-level feature maps of multiple dimensions are respectively a low-level feature map L1 corresponding to a 32-channel number, a low-level feature map L2 corresponding to a 64-channel number, and a low-level feature map L3 corresponding to a 128-channel number, and the low-level feature map L3 is a low-level feature map of the largest dimension. The low-level feature map L3 is subjected to multiple-dimensional upsampling processing to obtain a high-level feature map H1, a high-level feature map H2, and a high-level feature map H3. Based on this, the high-level feature map H1, the high-level feature map H2, the high-level feature map H3, and the low-level feature map L3 are fused to obtain a first global feature map.

[0039] It should be noted that the high-level feature maps of multiple dimensions obtained by upsampling and the low-level feature maps are images of different resolutions and different levels (low-level features and high-level features). Because the low-resolution picture is relatively blurred, the approximate position of the object can be obtained, but many details are ignored; while the high-resolution picture can describe more details and has more semantic information. Through feature fusion, the features with different characteristics in these images of different resolutions and different levels can be utilized, so that the first global feature map obtained can more accurately and completely describe the features possessed by the brain image itself.

[0040] Based on the above steps, the electronic device can segment and label the brain regions of the brain image according to the first global feature map. The electronic device can input the first global feature map to a fully connected layer, and the fully connected layer can map the first global feature map to a specific class label, i.e., match the different voxels in the first global feature map to the corresponding labels, and divide the voxels with the same label into the same brain region, thereby completing the segmentation and labeling of the brain regions of the brain image.

[0041] In this embodiment, the obtained brain image can be input into the feature extraction layer in the brain image processing model, and the brain image can be subjected to multi-dimensional down-sampling processing to obtain low-layer feature maps of multiple dimensions of the brain image, which represent low-layer semantic information. The low-layer feature maps of multiple dimensions are input into the brain region segmentation layer, and multi-dimensional up-sampling processing is performed based on the low-layer feature maps of the maximum dimension. In the up-sampling process, the high-layer feature maps of multiple dimensions obtained by up-sampling are fused with the low-layer feature maps based on the skip connection to obtain a global feature map. According to the global feature map, the brain image is segmented and labeled in the brain region. In this way, the brain region segmentation and labeling can be performed based on the global feature of the entire brain image, without the need to segment the brain image into image blocks, thereby avoiding the loss of local information between image blocks and improving the accuracy of segmentation and labeling.

[0042] In some optional embodiments, the electronic device can also input the low-layer feature maps of multiple dimensions into the image reconstruction layer in the brain image processing model, and perform multi-dimensional up-sampling processing based on the low-layer feature maps of the maximum dimension. In the up-sampling processing, the high-layer feature maps of multiple dimensions obtained by up-sampling are fused with the low-layer feature maps based on the skip connection technology to obtain a second global feature map. Since the up-sampling processing can increase the resolution of the image, the size and resolution of the second global feature map are consistent with those of the brain image. The use of “second” to limit the global feature map is only used to distinguish different global feature maps (i.e., the first global feature map in the foregoing). This step has the same principle as step 13 in the foregoing, and will not be described here.

[0043] After obtaining the second global feature map, the image reconstruction layer can perform image reconstruction according to the second global feature map to generate an image similar to the initially obtained brain image. The image reconstruction layer can input the second global feature map into the fully connected layer to map the second global feature map to an image of the same size as the brain image, thereby completing the image reconstruction process.

[0044] The above process will be further described below. The image reconstruction layer can gradually restore the low-resolution feature maps (i.e., the low-layer feature maps of multiple dimensions) output by the feature extraction layer to a second feature map of the same size as the brain image through the deconvolution layer and the up-sampling (or interpolation) operation matched with each corresponding layer of the feature extraction layer. Finally, the second global feature can be converted into an image of the same size as the brain image through the fully connected layer, thereby completing the image reconstruction process.

[0045] Specifically, the image reconstruction layer can employ a convolution kernel window with a size of 2 to start from the top-left corner of the compressed second feature map, and slide 2 voxels each time; for each position slid by the deconvolution kernel, multiply the overlapping part of the deconvolution kernel and the input data element by element, and then accumulate the obtained product to obtain the pixel value obtained by the current deconvolution kernel; finally, the shape of the obtained output data is [batch_size, height x 2, width x 2, channels / 2], where height x 2 and width x 2 represent the height and width after upsampling, and channels / 2 represents the number of channels of the output data, which is half of the number of channels of the input data. After each upsampling, the resolution of each dimension of the image is doubled, so that the brain image finally obtained by marking each brain region can maintain the same size and resolution as the brain image initially obtained in step 11 in the foregoing.

[0046] In this way, the feature fusion of the low-level feature map and the high-level feature map can be performed based on the upsampling processing, and the size and resolution of the finally obtained second global feature map can be maintained consistent with the initially obtained brain image, so that the image reconstruction can be performed based on the second global feature map, and the brain image marked with each brain region can be more accurately obtained.

[0047] In some optional embodiments, the step 11 "obtaining a brain image" in the foregoing embodiments can be implemented based on the following steps:

[0048] Step 111: obtaining a head image.

[0049] Step 112: performing tilt correction on the head image to adjust the tilt angle of the head image to a preset range. For example, the electronic device can determine the rotation axis in the head image, and rotate the rotation axis by an angle corresponding to the preset range to adjust the tilt angle of the head image to the preset range. For example, the electronic device can detect the scan gantry tilt angle in the CT scanning device, which is usually 30 degrees in the Z-axis direction, and perform an affine transformation on the head image opposite to the scan gantry tilt angle based on the detected scan gantry tilt angle, for example, perform an affine transformation of the head image by an oblique cut of -30 degrees.

[0050] Step 113, extract a head region image from the tilt-corrected head image, the head region image including an intracranial region and a skull region. The skull region refers to the region occupied by the skull, and the intracranial region refers to the region occupied by the brain inside the skull. The electronic device can extract the head region image based on a target recognition algorithm, which can take the tilt-corrected head image as input and output the class label of the object existing in the head image. Based on the target recognition algorithm, the part of the image belonging to the "intracranial region and skull region" class label in the head image can be identified as the head region image.

[0051] Step 114, extract the intracranial region image from the head region image as the brain image by taking the skull region as a mask. The mask is a shield or cover, which usually exists as a separate layer, and is used to shield part of the image content, equivalent to a window. The electronic device can apply the mask to the head region image to remove the skull region and retain the intracranial region, so that the part of the image of the intracranial region can be taken as the brain image.

[0052] Optionally, after step 114, the electronic device can also window the brain image obtained in step 114, such as transforming the input image pixel value to 0-255 by using a window width of 80 and a window level of 40. After that, the windowed brain image can also be cropped to remove the objects around the brain region and the redundant blank background, thereby further reducing the image size.

[0053] In this way, the electronic device can ensure that the input image size is reduced while retaining the complete brain region in the intracranial region, thereby obtaining the brain image more accurately.

[0054] In some optional embodiments, when the electronic device performs multi-dimensional down-sampling processing on the brain image to obtain low-level feature maps of multiple dimensions of the brain image, the following implementation can be used:

[0055] Implementation R1, a stacked convolution layer, a nonlinear activation function, and a max-pooling layer are used in the feature extraction layer to perform multi-dimensional down-sampling processing on the brain image to obtain low-level feature maps of multiple dimensions.

[0056] The stacked convolutional layer can be used for convolution of the brain image, can include a plurality of stacked convolutional layers stacked with each other, and each stacked convolutional layer corresponds to a respective dimension. The nonlinear activation function is a function running on a neuron of an artificial neural network in the brain image processing model, and is responsible for mapping the input of the neuron to the output. The nonlinear activation function can be a Sigmoid function (a commonly used S-shaped function in biology), a Tanh function (a hyperbolic function), or a Relu function (The Rectified Linear Unit), and the embodiment is not limited. The max pooling layer can be arranged between any two adjacent stacked convolutional layers, and can be used for dimension reduction processing of the low-level feature map output by the input stacked convolutional layer to reduce the calculation amount.

[0057] The above embodiment R1 can include the following steps R11 to R13. The following will be further described.

[0058] In step R11, a plurality of stacked convolutional layers with a channel number from small to large are sequentially used in the feature extraction layer to perform convolution processing on the brain image. Since the channel number of each stacked convolutional layer is different, the plurality of stacked convolutional layers with a channel number from small to large are sequentially used to perform convolution processing on the brain image with different dimensions, so as to obtain convolution results with different dimensions. Through the above manner, the brain image can be subjected to multi-dimensional down-sampling processing.

[0059] In step R12, a nonlinear activation function is used in each stacked convolutional layer to perform nonlinear processing on each convolution result to obtain a low-level feature map corresponding to the dimension. When the nonlinear activation function is used for nonlinear processing, the nonlinear activation function can be used to perform corresponding function operation on the convolution result to obtain a corresponding function value, thereby introducing a nonlinear factor for the convolution result. For example, the ReLU nonlinear activation function can perform the following nonlinear processing on the convolution result: the value greater than 0 is retained, and the value less than 0 is taken as 0; the Sigmoid can constrain the convolution result between 0 and 1.

[0060] In step R13, a max pooling layer is used between adjacent stacked convolutional layers to perform pooling processing on the low-level feature map output by the previous stacked convolutional layer to obtain the input of the next stacked convolutional layer.

[0061] The following will be described for any two adjacent stacked convolutional layers and the max-pooling layer arranged between the stacked convolutional layers. After the previous stacked convolutional layer generates a low-level feature map, the low-level feature map can be output to the max-pooling layer. The max-pooling layer can divide the low-level feature map into a plurality of regions of a preset size, and take the maximum value of the pixel values of a plurality of pixels in any region as the pixel value of the region, thereby performing pooling processing on the low-level feature map. Then, the max-pooling layer can input the low-level feature map after the pooling processing to the next stacked convolutional layer.

[0062] In this way, the electronic device can perform multi-dimensional down-sampling processing on the brain image in the feature extraction layer by using the stacked convolutional layer, the nonlinear activation function and the max-pooling layer, and more accurately obtain a plurality of low-level feature maps of multiple dimensions.

[0063] Optionally, the step R12 in the foregoing embodiments can be implemented based on the following steps R121-step R123:

[0064] Step R121, sliding a convolution kernel window with a size of n*n*n from the top left corner of the input data, and each time the sliding step is 1 voxel distance, to extract abstract features in the input data. n is a positive integer, and the input data is a brain image or a low-level feature map after the pooling processing.

[0065] Step R122, for each position of the convolution kernel window that the convolution kernel slides through, performing element-by-element multiplication on the overlapping part of the convolution kernel and the input data, and adding the obtained products to obtain a scalar value.

[0066] Step R123, performing max-pooling operation with a step of m and nonlinear processing based on the nonlinear activation function on the scalar value to obtain a feature value corresponding to the convolution kernel.

[0067] In this way, the feature value corresponding to each convolution kernel can be obtained, and thus the feature values corresponding to all the convolution kernels can form a low-level feature map with the same number of convolution kernels.

[0068] In some optional embodiments, the step 13 in the foregoing embodiments, "performing multi-dimensional up-sampling processing based on the low-level feature map with the maximum dimension, and during the up-sampling processing, fusing the high-level feature map of multiple dimensions obtained by up-sampling with the low-level feature map based on the skip connection technology to obtain the first global feature map", can be implemented based on the following implementation K1:

[0069] In the brain region segmentation layer, a deconvolution layer, a nonlinear activation function and a convolution layer are used to perform multi-dimensional up-sampling processing based on a low-layer feature map with a maximum dimension, and during the up-sampling processing, the multi-dimensional high-layer feature maps obtained through the up-sampling are fused with the low-layer feature map based on a skip connection technology to obtain a first global feature map.

[0070] The embodiment K1 can be implemented based on steps K11-K13, which will be further described below.

[0071] In the brain region segmentation layer, a convolution layer is used to perform convolution processing on a low-layer feature map with a maximum dimension, and a nonlinear activation function is used to perform nonlinear processing on the convolution result to obtain a reference feature map. The related content of performing nonlinear processing on the convolution result using the nonlinear activation function has been described in the foregoing, and will not be described here again.

[0072] Step K12, the reference feature map is subjected to multiple deconvolution processing using a deconvolution layer to obtain an output feature map of each deconvolution processing. The deconvolution can be understood as an inverse operation of convolution. The dimension of each deconvolution processing of the reference feature map is different, that is, the number of channels used is different, so that the output feature map of multiple deconvolution processing can be obtained, and the dimension of the output feature map obtained by each deconvolution processing is different. Through the above manner, the brain image can be subjected to multi-dimensional up-sampling processing.

[0073] Specifically, step K12 can be implemented based on the following manner: for each deconvolution processing, a deconvolution kernel window with a size of k is used to start from the top left corner of the input feature map, and to slide k voxels each time, k being a positive integer; for each position slid by the deconvolution kernel, the overlapping part of the deconvolution kernel and the input feature map is multiplied element by element, and the obtained product is accumulated to obtain a feature value corresponding to the deconvolution kernel, and all the feature values corresponding to the deconvolution kernels generate the output feature map of this deconvolution processing.

[0074] Step K13, using a residual connection to add the input feature map and the output feature map of each deconvolution processing to obtain a high-layer feature map of a corresponding dimension, the input feature map being the reference feature map or the target feature map obtained by the last deconvolution processing. The residual connection is used to establish a connection between the input and the output of each deconvolution processing, so that the input feature map and the output feature map of each deconvolution processing can be added based on the residual connection to obtain a high-layer feature map of a corresponding dimension.

[0075] For example, the first deconvolution processing corresponds to the first dimension, and the input feature map of the first deconvolution processing is the reference feature map. Therefore, the reference feature map of the input of the first deconvolution processing can be added to the output feature map obtained by performing deconvolution processing on the reference feature map to obtain a high-level feature map of the first dimension. Similarly, the second deconvolution processing corresponds to the second dimension, and the input feature map of the second deconvolution processing is the target feature map obtained by the first deconvolution processing. Therefore, the target feature map obtained by the first deconvolution processing can be added to the output feature map obtained by performing deconvolution processing on the target feature map to obtain a high-level feature map of the second dimension.

[0076] In step K14, the high-level feature map obtained by each deconvolution processing is connected to the low-level feature map of the corresponding dimension by jump connection to obtain an intermediate state feature map. The high-level feature map obtained by each deconvolution processing can be fused with the low-level feature map of the corresponding dimension based on the jump connection, so that the intermediate state feature map can be obtained.

[0077] In step K15, the intermediate state feature map is down-sampled using a convolution layer to obtain a target feature map with the same number of channels as the input feature map, and the target feature map is the input feature map of the next deconvolution processing or the first global feature map. For example, the intermediate state feature map P1 can be down-sampled using a convolution layer to obtain a target feature map with the same number of channels as the intermediate state feature map P1, and the target feature map is the input feature map of the next deconvolution processing. For another example, in the last deconvolution processing, the intermediate state feature map P2 can be down-sampled using a convolution layer to obtain a first global feature map with the same number of channels as the intermediate state feature map P2.

[0078] In this way, the electronic device can obtain the first global feature map more accurately based on the brain image processing model.

[0079] In some optional embodiments, the aforementioned step 14 “segmenting and labeling the brain region of the brain image according to the first global feature map” can be implemented based on the following steps:

[0080] The first global feature map is input into a fully connected layer to map the plurality of voxels in the first global feature map to the preset class labels. Specifically, the fully connected layer can flatten the first global feature map into a one-dimensional vector, and connect the one-dimensional vector with the predefined number of classes. Through a softmax activation function, the original output of the fully connected layer is mapped to a probability (a floating point number between 0 and 1) representing each class, so as to map the plurality of voxels in the first global feature map to the preset class labels.

[0081] Afterwards, at least one voxel with the same class label among the plurality of voxels can be divided into the same brain region according to the class label of the plurality of voxels; and the plurality of brain regions obtained by the division are respectively marked with the corresponding class label.

[0082] In some optional embodiments, the training process of the brain image processing model in the foregoing embodiments comprises:

[0083] Step G1, obtaining a brain image sample and a real brain region label corresponding to the brain image sample.

[0084] Step G2, under the supervision of the real brain region label corresponding to the brain image sample, converging to a specified range with a preset loss function of a neural network as the target, training the preset neural network by using the brain image sample, and obtaining a brain image processing model. The loss function is used to calculate the error between the brain region label of the brain image predicted by the neural network and the real brain region label.

[0085] Optionally, obtaining the brain image sample of the human brain and the brain region label corresponding to the brain image sample comprises the following steps: obtaining the brain image sample and the brain region label corresponding to the reference brain image sample and the human brain standard sample; determining a transformation matrix from the reference brain image sample to the brain image sample by using a rigid registration algorithm; and applying the transformation matrix to the brain image sample to obtain the brain region label corresponding to the brain image sample. The reference brain image sample can be a CT reference image with a brain region label defined by a medical expert. In this way, the brain image sample can be quickly labeled based on the rigid registration algorithm and a reference brain image sample, without the need for manual labeling of the brain image sample, thereby saving labor costs.

[0086] Optionally, the electronic device can also obtain target real labels corresponding to a plurality of target brain image samples from the plurality of target brain image samples in response to a label annotation operation on the plurality of target brain image samples. The label annotation operation can be understood as a manual correction operation. These target real labels have higher accuracy and reliability after being manually corrected.

[0087] Afterwards, the electronic device can perform adaptive learning on the brain image processing model by using the plurality of target brain image samples and the target real labels corresponding to the plurality of target brain image samples. The adaptive learning is used to optimize the parameters of the feature extraction layer and the parameters of the image reconstruction layer in the brain image processing model.

[0088] As Figure 2As shown, this adaptive learning process fixes the parameters of the image reconstruction layer and optimizes only the parameters of the feature extraction layer and the brain segmentation layer in the brain image processing model to improve the model performance. In this way, through continuous iteration, the model trained on a small amount of labeled data for existing age groups is gradually transferred to more data for more age groups, thereby improving the model's repeatability and generalization ability.

[0089] like Figure 2 As shown, x represents a single target brain image sample, μ and σ are the mean and variance of the latent semantic space, respectively, and z is a feature vector randomly sampled from a Gaussian distribution defined by the mean μ and variance σ. This feature vector is processed by an image reconstruction layer to obtain the reconstructed image of x. (That is, corresponding to the image reconstruction task in the model mentioned above), after processing by the brain region segmentation layer, the predicted brain region label for x is obtained. (That is, corresponding to the brain region segmentation task mentioned earlier), the loss function L is composed of the reconstruction loss (l vae ) and cross-entropy (l cls ) are added together to form a whole.

[0090] The following will combine Figure 2 The training process of the model will be further explained.

[0091] First, three networks can be trained simultaneously based on the input image x and its coarse label y: a feature extraction layer (i.e., ... Figure 2 encoder in Image reconstruction layer (i.e.) Figure 2 decoder p in θ ), and brain region segmentation layers (i.e. Figure 2 Classifier in The labels obtained using the rigid registration algorithm described above are used to effectively supervise feature learning, ensuring that the learned features can both effectively retain key information from the input brain image samples and accurately distinguish different intracranial regions. The optimized objective function is shown in Equation 1.

[0092]

[0093] in, θ The parameters of the feature extraction layer, image reconstruction layer, and brain region segmentation layer, respectively, {x i} i＝1，…，N ∈R p For any brain image sample, {y i} i＝1，…，N ∈R 1Let N be the real brain region label corresponding to any brain image sample, N be the number of samples, and λ be an adjustable weight used to control the supervision loss. The brain image processing model uses the empirical risk minimization algorithm to solve the expected loss in formula (1). By iteratively adjusting the network weights along the negative gradient direction of the empirical loss function through gradient descent, the model parameters that minimize the difference between the reconstructed image and the original input image are found. When λ = 0, formula (1) is equivalent to the traditional variational autoencoder.

[0094] well trained p θ ,and Then, fix the image reconstruction layer p θ (That is, the image reconstruction layer does not participate in gradient update). Multiple target brain image samples, manually corrected by experts and containing the target's true labels, are used to refine the feature extraction layer. and brain region segmentation layer Perform adaptive transfer learning, such as Figure 2 As shown. The loss function used when training the brain region segmentation layer ( Figure 2 l in cls Cross-entropy is the loss function used when training the image reconstruction layer. Figure 2 Chinese vae It is the mean squared error plus the Kullback-Leibler divergence (KL divergence, used to measure the similarity or difference between two probability distributions; the smaller it is, the more similar the two distributions are).

[0095] Figure 3 (above) is the loss function (l) used when training the brain region segmentation layer. cls The change in ) can be seen in l cls A steady downward trend. Figure 3 (Below) shows the change in the loss function when training the image reconstruction layer. Since the loss function of the image reconstruction layer is not used when updating the parameters of the brain image processing model, the change in the loss function of the image reconstruction layer is stable.

[0096] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 11 to 14 can be device A; or the execution subject of steps 11 and 12 can be device A, and the execution subject of steps 13 and 14 can be device B; and so on.

[0097] In addition, in some of the processes described in the above embodiments and accompanying drawings, a plurality of operations are included in a specific order, but it should be clear that these operations can be executed in the order in which they appear in this document or in parallel, and the serial numbers of the operations such as 11, 12, etc. are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes can include more or fewer operations, and the operations can be executed in sequence or in parallel.

[0098] It should be noted that the "first", "second" and the like in this document are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence, nor do "first" and "second" represent different types.

[0099] Figure 4 is a structural schematic diagram of an electronic device provided by an exemplary embodiment of the present application, as Figure 4 indicated, the electronic device includes a memory 401 and a processor 402.

[0100] The memory 401 is used to store computer programs and can be configured to store other various data to support operations on the terminal device. Examples of these data include instructions for operating any application or method on the terminal device, contact data, phonebook data, messages, pictures, videos, etc.

[0101] The processor 402 is coupled to the memory 401 and is used to execute the computer programs in the memory 401, to: acquire a brain image; input the brain image into a feature extraction layer in a pre-trained brain image processing model, perform multi-dimensional down-sampling processing on the brain image to obtain low-level feature maps of multiple dimensions of the brain image, the low-level feature maps representing low-level semantic information; input the low-level feature maps of multiple dimensions into a brain region segmentation layer in the brain image processing model, perform multi-dimensional up-sampling processing based on the low-level feature maps of the maximum dimension, and in the up-sampling processing, fuse the high-level feature maps of multiple dimensions obtained by up-sampling with the low-level feature maps based on the skip connection technology to obtain a first global feature map; and segment and label the brain regions of the brain image according to the first global feature map.

[0102] Further optionally, the processor 402 is further used to: input the low-level feature maps of multiple dimensions into an image reconstruction layer in the brain image processing model, perform multi-dimensional up-sampling processing based on the low-level feature maps of the maximum dimension, and in the up-sampling processing, fuse the high-level feature maps of multiple dimensions obtained by up-sampling with the low-level feature maps based on the skip connection technology to obtain a second global feature map, and perform image reconstruction according to the second global feature map to obtain a brain image labeled with each brain region.

[0103] Further optionally, when the processor 402 performs the multi-dimensional down-sampling processing on the brain image to obtain the multi-dimensional low-level feature maps of the brain image, specifically, in the feature extraction layer, the stacked convolutional layer, the nonlinear activation function and the max-pooling layer are used to perform the multi-dimensional down-sampling processing on the brain image to obtain the multi-dimensional low-level feature maps.

[0104] Further optionally, when the processor 402 performs the multi-dimensional down-sampling processing on the brain image to obtain the multi-dimensional low-level feature maps in the feature extraction layer by using the stacked convolutional layer, the nonlinear activation function and the max-pooling layer, specifically, in the feature extraction layer, the brain image is sequentially processed by using multiple stacked convolutional layers with the number of channels from small to large; in each stacked convolutional layer, the nonlinear activation function is used to perform nonlinear processing on each convolution result to obtain a low-level feature map of a corresponding dimension; and the max-pooling layer is used to perform pooling processing on the low-level feature map output by the previous stacked convolutional layer to obtain the input of the next stacked convolutional layer.

[0105] Further optionally, when the processor 402 performs the nonlinear processing on each convolution result in each stacked convolutional layer to obtain a low-level feature map of a corresponding dimension, specifically, a convolution kernel window with a size of n*n*n is used to slide from the top-left corner of the input data, and the sliding step is 1 voxel distance each time, to extract abstract features in the input data, n is a positive integer, and the input data is the brain image or the low-level feature map after the pooling processing; for each convolution kernel sliding position of the convolution kernel window, the overlapping part of the convolution kernel and the input data is multiplied element by element, and the sum of the obtained products is obtained to obtain a scalar value; the max-pooling operation with a step of m and the nonlinear processing based on the nonlinear activation function are performed on the scalar value to obtain a feature value corresponding to the convolution kernel; and the feature values corresponding to all convolution kernels form a low-level feature map with the same number of convolution kernels.

[0106] Further optionally, when the processor 402 performs the multi-dimensional up-sampling processing based on the low-level feature map with the maximum dimension and fuses the multi-dimensional high-level feature maps obtained by the up-sampling with the low-level feature maps based on the skip connection technology to obtain the first global feature map, specifically, in the brain region segmentation layer, the deconvolutional layer, the nonlinear activation function and the convolutional layer are used to perform the multi-dimensional up-sampling processing based on the low-level feature map with the maximum dimension, and in the up-sampling processing, the multi-dimensional high-level feature maps obtained by the up-sampling are fused with the low-level feature maps based on the skip connection technology to obtain the first global feature map.

[0107] Further optionally, the processor 402 is configured to, in the brain region segmentation layer, perform multi-dimensional up-sampling processing on the low-level feature map with the maximum dimension by using a de-convolution layer, a nonlinear activation function and a convolution layer, and perform fusion on the high-level feature maps obtained by up-sampling and the low-level feature map based on a skip connection technology to obtain the first global feature map.

[0108] Further optionally, the processor 402 is configured to, in the brain region segmentation layer, perform multi-dimensional up-sampling processing on the low-level feature map with the maximum dimension by using a de-convolution layer, a nonlinear activation function and a convolution layer, and perform fusion on the high-level feature maps obtained by up-sampling and the low-level feature map based on a skip connection technology to obtain the first global feature map.

[0109] Further, as shown in Figure 4 the electronic device further includes a communication component 403 and other components. Figure 4 Some components are only schematically shown in the electronic device, and it does not mean that the electronic device only includes Figure 4 the components shown.

[0110] The memory in the above Figure 4 may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0111] The above Figure 4The communication component in the electronic device is configured to facilitate wired or wireless communication between the electronic device and other devices. The electronic device can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G, or 5G, or a combination thereof. In an example embodiment, the communication component receives a broadcast signal or broadcast related information from an external broadcast management system via a broadcast channel. In an example embodiment, the communication component can be implemented based on a near field communication (NFC) technology, a radio frequency identification (RFID) technology, an infrared data association (IrDA) technology, an ultra wideband (UWB) technology, a Bluetooth (BT) technology and other technologies.

[0112] In this embodiment, the obtained brain image can be input into the feature extraction layer in the brain image processing model, and the brain image is subjected to multi-dimensional down-sampling processing to obtain low-level feature maps of the brain image in multiple dimensions, and the low-level feature maps represent low-level semantic information. The low-level feature maps in multiple dimensions are input into the brain region segmentation layer, and multi-dimensional up-sampling processing is performed based on the low-level feature maps with the maximum dimension, and in the up-sampling process, the high-level feature maps in multiple dimensions obtained by up-sampling are fused with the low-level feature maps based on the skip connection to obtain a global feature map; and the brain image is segmented and labeled according to the global feature map. In this way, the brain region segmentation and labeling can be performed based on the global features of the entire brain image, without the need to segment the brain image into image blocks, thereby avoiding the loss of local information between the image blocks, and improving the accuracy of segmentation and labeling.

[0113] Correspondingly, the embodiment of the present application also provides a computer readable storage medium storing a computer program, and the computer program is executed to implement each step that can be executed by the electronic device in the above method embodiment.

[0114] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof.

[0116] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof.

[0117] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof. Figure 1 one or more flowcharts and / or blocks in the flowcharts and / or combination thereof.

[0118] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0119] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory, non-volatile memory, such as read-only memory (ROM), EPROM, and / or flash memory. The memory is an example of computer-readable media.

[0120] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can implement information storage by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0121] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not only include those elements, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0122] The above only describes the embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various changes and modifications to the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the scope of the claims of the present application.

Claims

1. An image processing method, characterized by, The method comprises: acquiring a brain image; inputting the brain image into a feature extraction layer of a pre-trained brain image processing model, in the feature extraction layer, a plurality of stacked convolution layers with increasing number of channels are sequentially used to perform convolution processing on the brain image; a convolution kernel window with a size of n*n*n is used to slide from the top-left corner of the input data, and each time the sliding step is 1 voxel distance, so as to extract abstract features in the input data, n is a positive integer, and the input data is the brain image or a low-level feature map after pooling processing; for each convolution kernel sliding position of the convolution kernel window, the overlapping part of the convolution kernel and the input data is multiplied element by element, and the obtained product is added to obtain a scalar value; a maximum pooling operation with a step of m and a nonlinear processing based on a nonlinear activation function are performed on the scalar value to obtain a feature value corresponding to the convolution kernel; wherein the feature values corresponding to all convolution kernels form a low-level feature map with the same number of convolution kernels as the number of convolution kernels; a maximum pooling layer is used between adjacent stacked convolution layers to pool the low-level feature map output by the previous stacked convolution layer to obtain the input of the next stacked convolution layer; the low-level feature map represents low-level semantic information; a plurality of dimensional low-level feature maps are input into a brain region segmentation layer of the brain image processing model, and a plurality of dimensional up-sampling processing is performed based on the low-level feature map with the largest dimension, and in the up-sampling processing, a plurality of dimensional high-level feature maps obtained by up-sampling are fused with low-level feature maps based on a skip connection technology to obtain a first global feature map; according to the first global feature map, the brain image is segmented and labeled in the brain region.

2. The method of claim 1, wherein, Further comprising: inputting the plurality of dimensional low-level feature maps into an image reconstruction layer of the brain image processing model, and performing a plurality of dimensional up-sampling processing based on the low-level feature map with the largest dimension, and in the up-sampling processing, a plurality of dimensional high-level feature maps obtained by up-sampling are fused with low-level feature maps based on a skip connection technology to obtain a second global feature map, and according to the second global feature map, the image is reconstructed to obtain a brain image with each brain region labeled.

3. The method of claim 1, wherein, performing a plurality of dimensional up-sampling processing based on the low-level feature map with the largest dimension, and in the up-sampling processing, a plurality of dimensional high-level feature maps obtained by up-sampling are fused with low-level feature maps based on a skip connection technology to obtain a first global feature map, comprising: in the brain region segmentation layer, an up-sampling processing of a plurality of dimensions is performed based on the low-level feature map with the largest dimension by using a deconvolution layer, a nonlinear activation function and a convolution layer, and in the up-sampling processing, a plurality of dimensional high-level feature maps obtained by up-sampling are fused with low-level feature maps based on a skip connection technology to obtain a first global feature map.

4. The method of claim 3, wherein, In the brain region segmentation layer, a deconvolution layer, a nonlinear activation function and a convolution layer are used to perform multi-dimensional up-sampling processing based on a low-layer feature map with the maximum dimension, and during the up-sampling processing, a plurality of high-layer feature maps obtained through up-sampling are fused with a low-layer feature map based on a skip connection technology to obtain a first global feature map, including: In the brain region segmentation layer, a convolution layer is used to perform convolution processing on a low-layer feature map with the maximum dimension, and a nonlinear activation function is used to perform nonlinear processing on the convolution result to obtain a reference feature map. A deconvolution layer is used to perform multiple deconvolution processing on the reference feature map to obtain an output feature map of each deconvolution processing, and a residual connection is used to add an input feature map of each deconvolution processing to an output feature map to obtain a high-layer feature map of a corresponding dimension, the input feature map being the reference feature map or a target feature map obtained through last deconvolution processing. The high-layer feature map obtained through each deconvolution processing is connected to a low-layer feature map of a corresponding dimension through a skip connection to obtain an intermediate state feature map, and a convolution layer is used to down-sample the intermediate state feature map to obtain a target feature map with the same number of channels as the input feature map, the target feature map being an input feature map of next deconvolution processing or being the first global feature map.

5. The method according to any one of claims 1 to 4, characterized in that, According to the first global feature map, the brain image is segmented and labeled in brain regions, including: The first global feature map is input into a fully connected layer to map a plurality of voxels in the first global feature map to a preset class label. According to the class label of the plurality of voxels, at least one voxel with the same class label in the plurality of voxels is divided into a same brain region. The plurality of brain regions obtained through division are respectively labeled with corresponding class labels.

6. An electronic device, comprising: including: a memory, a processor; wherein the memory is configured to store one or more computer instructions; the processor is configured to execute the one or more computer instructions to perform the steps in the method of any one of claims 1-5.

7. A computer readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, the processor can implement the steps in the method of any one of claims 1-5.

Images