Brain image detection method, computer device and storage medium

Abstract

The present application relates to a brain image detection method, computer equipment and storage medium. The method comprises: obtaining a brain image of an object to be detected; dividing the brain image into a plurality of image blocks, and inputting each of the image blocks into a corresponding image block hierarchical network for feature extraction to obtain a feature map of each of the image blocks; performing a first weight screening process on the feature map of each of the image blocks, and determining the feature map of each candidate image block from the feature map of each of the image blocks, wherein the number of each candidate image block is less than the number of the plurality of image blocks; inputting the feature map of each of the candidate image blocks into a neural network model to obtain the target area corresponding to the brain image and the category of the brain image. The use of this method can improve detection efficiency and detection accuracy.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a brain image detection method, a computer device and a storage medium. Background Art

[0002] Brain atrophy refers to a phenomenon in which organic lesions of the brain tissue itself occur due to various reasons, resulting in atrophy. Its pathological manifestations are a reduction in the volume of brain tissue, a decrease in the number of cells, and an expansion of the ventricles and subarachnoid space. When brain atrophy occurs in the human brain, it will bring a lot of unnecessary trouble to the normal life of the human body. Therefore, it is particularly important to detect whether brain atrophy occurs in the human brain.

[0003] In the related art, when detecting brain atrophy, the human brain image is generally manually divided into multiple regions of interest. Then, the doctor manually marks the lesion area in the divided regions of interest based on experience, and extracts features in the marked lesion area. The extracted features are then input into a trained classifier to obtain the category of the lesion area.

[0004] However, the above technology manually marks the lesion area based on experience, which has the problems of low detection efficiency and high false detection rate. Summary of the invention

[0005] Based on this, it is necessary to provide a brain image detection method, device, computer equipment and storage medium that can improve detection efficiency and accuracy in response to the above technical problems.

[0006] A brain image detection method, the method comprising:

[0007] Acquiring a brain image of a subject to be detected;

[0008] Divide the brain image into multiple image blocks, and input each image block into the corresponding image block hierarchical network for feature extraction to obtain a feature map of each image block;

[0009] Performing a first weighted screening process on the feature map of each image block, determining the feature maps of each candidate image block from the feature map of each image block, wherein the number of each candidate image block is less than the number of the plurality of image blocks;

[0010] The feature map of each candidate image block is input into the neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0011] In one of the embodiments, after performing the first weight screening process on the feature maps of each image block and determining the feature maps of each candidate image block from the feature maps of each image block, the method further includes:

[0012] Performing regional combination on the feature maps of each candidate image block to obtain feature maps of each combined region;

[0013] Each combined regional feature map is input into the corresponding regional hierarchical network for feature extraction to obtain a regional level feature map of each combined regional feature map.

[0014] In one embodiment, the method further comprises:

[0015] Each region-level feature map is subjected to a second weight screening process, and a feature map of each candidate region is determined from each region-level feature map; the number of each candidate region is less than the number of each combined region.

[0016] In one embodiment, the feature map of each candidate image block is input into a neural network model to obtain a target region corresponding to the brain image and a category of the brain image, including:

[0017] The feature maps of each candidate area are concatenated and input into the individual hierarchical network of the neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0018] In one of the embodiments, the first weight screening process is used to calculate the weight corresponding to the feature map of each image block, and obtain the feature map of each candidate image block based on the feature map of each image block and the corresponding weight; the feature map of each candidate image block is the feature map of each image block whose weight is greater than a preset threshold.

[0019] In one of the embodiments, the second weight screening process is used to calculate the weights corresponding to each region-level feature map, and obtain the feature map of each candidate region based on each region-level feature map and the corresponding weight; the feature map of each candidate region is the region-level feature map whose weight is greater than a preset threshold in the region-level feature map.

[0020] In one embodiment, the feature maps of each candidate region are concatenated and input into the individual hierarchical network of the neural network model to obtain the target region corresponding to the brain image and the category of the brain image, including:

[0021] The feature maps of each candidate area are concatenated and input into the individual hierarchical network of the neural network model for convolution processing to obtain the target area corresponding to the brain image;

[0022] A classification operation is performed on the target area corresponding to the brain image to obtain the category of the brain image.

[0023] In one embodiment, the training method of the above-mentioned image block level network, region level network, and individual level network includes:

[0024] Obtain a sample brain image, where the sample brain image has been labeled with an image category;

[0025] Training an initial image block hierarchical network based on sample brain images to obtain an image block hierarchical network;

[0026] The initial region-level network is trained based on the image block-level network and the labeled image categories to obtain a region-level network;

[0027] Based on the image block level network, the region level network and the labeled image categories, the initial individual level network is trained to obtain the individual level network.

[0028] A brain image detection device, comprising:

[0029] An acquisition module, used for acquiring a brain image of a subject to be detected;

[0030] An extraction module, used for dividing the brain image into a plurality of image blocks, and inputting each of the image blocks into a corresponding image block hierarchical network for feature extraction to obtain a feature map of each of the image blocks;

[0031] A first screening processing module, configured to perform a first weighted screening process on the feature graphs of each of the image blocks, and determine the feature graphs of each candidate image block from the feature graphs of each of the image blocks, wherein the number of the candidate image blocks is less than the number of the plurality of image blocks;

[0032] The determination module is used to input the feature map of each candidate image block into a neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0033] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0034] Acquiring a brain image of a subject to be detected;

[0035] Divide the brain image into multiple image blocks, and input each image block into the corresponding image block hierarchical network for feature extraction to obtain a feature map of each image block;

[0036] Performing a first weighted screening process on the feature map of each image block, determining the feature maps of each candidate image block from the feature map of each image block, wherein the number of each candidate image block is less than the number of the plurality of image blocks;

[0037] The feature map of each candidate image block is input into the neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0038] A readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0039] Acquiring a brain image of a subject to be detected;

[0040] Divide the brain image into multiple image blocks, and input each image block into the corresponding image block hierarchical network for feature extraction to obtain a feature map of each image block;

[0041] Performing a first weighted screening process on the feature map of each image block, determining the feature maps of each candidate image block from the feature map of each image block, wherein the number of each candidate image block is less than the number of the plurality of image blocks;

[0042] The feature map of each candidate image block is input into the neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0043] The above-mentioned brain image detection method, device, computer equipment and storage medium obtain a brain image of the object to be detected, divide the brain image into multiple image blocks, input the brain image into the corresponding image block hierarchical network for feature extraction, obtain the feature map of each image block, and perform a first weight screening process on the feature map of each image block, determine the feature map of each candidate image block from the feature map of each image block, and finally input the feature map of each candidate image block into the neural network model to obtain the target area corresponding to the brain image and the category of the brain image, wherein the number of each candidate image block is less than the number of multiple image blocks. In this method, since the lesion area is obtained by detecting the brain image through a neural network instead of obtaining the lesion area through manual annotation, the detection result obtained by this method is more accurate and the detection efficiency of the brain image is higher; in addition, since the method can perform weight screening on the important features in the brain image when detecting the brain image, and obtain the important features for detection, the parameters of the network model can be reduced, the detection speed can be accelerated, and the detection efficiency can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is an internal structure diagram of a computer device in one embodiment;

[0045] Figure 2 is a schematic diagram of a flow chart of a brain image detection method in one embodiment;

[0046] Figure 3 is a schematic flow chart of a brain image detection method in another embodiment;

[0047] Figure 4a is a schematic flow chart of a brain image detection method in another embodiment;

[0048] Figure 4b is a schematic flow chart of a brain image detection method in another embodiment;

[0049] Figure 5 is a schematic flow chart of a brain image detection method in another embodiment;

[0050] Figure 6 FIG. 4 is a structural block diagram of a brain image detection device in one embodiment. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0052] Brain atrophy refers to a phenomenon in which the brain tissue itself undergoes organic lesions due to various reasons, resulting in atrophy. Its pathological manifestations include a reduction in brain tissue volume, a decrease in the number of cells, and an expansion of the ventricles and subarachnoid space. It can last for several years to decades. CT or MR examinations can show atrophy of the cortex, cerebellum, brainstem, etc. It is one of the neuroimaging manifestations of many diseases such as Alzheimer's disease (AD), mild cognitive impairment (MCI), Parkinson's disease (PD), etc. Taking Alzheimer's disease (AD) as an example, it is one of the most common neurodegenerative diseases, so the detection of brain atrophy is particularly important. Magnetic resonance imaging (MRI) is increasingly widely used in the diagnosis of AD and its related diseases due to its non-radioactive and high-quality imaging of brain structures. This has also made the demand for computer-aided detection (CAD, Computer AD) urgent. Computer-aided diagnosis of AD can effectively screen AD patients and their early mild cognitive impairment (MCI) stages based on medical images such as MRI, greatly reducing the workload of doctors and improving the accuracy of doctors' detection. In addition, automatic labeling of key areas related to AD is also very important: on the one hand, since the labeled area is highly related to the disease, some CAD methods can be performed directly in the labeled area, greatly reducing the amount of calculation; on the other hand, automatic labeling of key areas related to the disease provides doctors with an effective reference for the diagnosis of the disease, which is of great significance in clinical applications. At present, the main problems facing CAD are low automation, low detection accuracy and slow speed; CAD processing usually uses manually designed feature extractors to extract features, which will lose a lot of useful information in the process and greatly limit the learning ability of the model. Traditional computer-aided diagnosis of AD and its related diseases using structural MRI images mainly includes five steps, namely: 1) data acquisition, 2) preprocessing, 3) image segmentation, 4) feature extraction, and 5) training classifiers and evaluating their performance; data acquisition refers to the process of collecting medical images through imaging equipment; preprocessing refers to removing the influence of the working condition of the imaging equipment on the image quality through a series of operations such as resampling, adjusting the direction, removing the skull, and grayscale correction, so as to provide high-quality images for subsequent analysis; image segmentation refers to pre-defining some characteristic areas related to AD based on current clinical knowledge, such as hippocampus, temporal lobe, amygdala, cingulate gyrus, etc., and then manually segmenting the MR images into multiple ROIs (regions of interest) based on these characteristic areas. There are also methods to define ROIs using template image registration methods; feature extraction refers to extracting several features from each divided ROI as samples for classifier training and testing by performing a series of operations on the image (such as sparse autoencoder, wavelet transform, singular value decomposition, etc.).Next, the input samples are divided into training sets and test sets according to a certain ratio, and the training data set is sent to the classifier for training, and the test data is used to evaluate the model performance (such as accuracy, robustness, precision, recall, etc.). The traditional use of structural MRI images for AD feature area acquisition mainly relies on manual annotation, which is mainly completed by doctors with rich clinical experience, which is slow and consumes huge manpower costs; at the same time, since the pathogenesis of AD is still unclear, there is no conclusive conclusion on the key diagnostic area related to its disease in clinical practice, so there are errors in the annotation results of different annotation personnel, which is not conducive to disease diagnosis; at the same time, there are individual differences in the brain lesion area of ​​AD patients, which further increases the difficulty of manual annotation. In recent years, research has proposed to use image registration methods to locate lesion areas. This method has achieved certain results for the positioning of atrophy or lesion areas, but because registration is time-consuming, limited in accuracy, and independent of the subsequent diagnosis process, it will lead to suboptimal performance of the AD diagnosis and feature area detection method combined with registration. Therefore, the embodiment of the present application provides a brain image detection method, device, computer equipment and storage medium, aiming to solve the above technical problems.

[0053] The brain image detection method provided in the embodiment of the present application can be applied to a computer device, and the internal structure diagram of the computer device can be as follows: Figure 1 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a brain image detection method is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0054] Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0055] It should be noted that the execution subject of the embodiments of the present application may be a brain image detection device or a computer device. The following embodiments are described with the computer device as the execution subject.

[0056] In one embodiment, a brain image detection method is provided. This embodiment involves a specific process of how to weight-screen a feature map of a brain image and input the screened feature map into a neural network model to obtain a target area and category of the brain image. Figure 2 As shown, the method may include the following steps:

[0057] S202: Acquire a brain image of the subject to be detected.

[0058] Among them, the objects to be detected can be children, the elderly, adults (neither children nor the elderly), etc., and of course can also be animals. However, in this embodiment, the objects to be detected are humans, and generally are described as elderly people; the objects to be detected can be people with normal brain development or people with abnormal brain development, and the objects to be detected here can be one or more; in addition, the brain image of the object to be detected can be a one-dimensional image, a two-dimensional image, a three-dimensional image, etc., and this embodiment mainly uses the three-dimensional brain image of the object to be detected.

[0059] Specifically, the computer device can reconstruct and correct the brain data of the object to be detected collected by the MR (Magnetic Resonance) device, CT (Computed Tomography) device or PET (Positron Emission Computed Tomography) device or PET-CT device or PET-MR device, so as to obtain the brain image of the object to be detected. Alternatively, the brain image of the object to be detected can also be reconstructed and corrected in advance and stored in the computer device. When it needs to be processed, the brain image is directly read from the memory of the computer device. Alternatively, the computer device can also obtain the brain image from an external device, for example, the brain image of the object to be detected is stored in the cloud, and when a processing operation is required, the computer device obtains the brain image of the object to be detected from the cloud. This embodiment does not limit the method for obtaining the brain image of the object to be detected.

[0060] S204, dividing the brain image into a plurality of image blocks, and inputting each image block into a corresponding image block hierarchical network for feature extraction to obtain a feature map of each image block.

[0061] In this step, before dividing the brain image into image blocks, the brain image can optionally be preprocessed, such as rotation, resampling, resizing, skull removal, image non-uniformity correction, histogram matching, grayscale normalization, etc., so that the size of all brain images can be the same value, such as 256*256*256mm for a three-dimensional image. 3 Etc., the directions are all standard Cartesian LPI coordinate systems, and the grayscale ranges are all standard images in the (-1,1) interval.

[0062] Secondly, the image block level network can be a neural network model, including a deep learning model, a convolutional neural network model, etc., and of course it can also be other network models; in addition, the number of image block level networks can be the same as the number of divided image blocks, and the structure and parameters of each image block level network are the same. In this embodiment, the image block level subnetwork continues the network design of DenseNet. In a traditional convolutional neural network, if the number of network layers is L, then there will be L connections, but in DenseNet, there will be L(L+1) / 2 network connections, that is, the input of each layer is composed of the output of all previous layers. The number of feature maps output by the convolution layer in each module of the entire network is very small (less than 100), and there will not be a large width like in a traditional convolutional neural network. This connection method makes the transmission of features and gradients more effective, and the network is easier to train. The main part of Densenet is composed of several Dense blocks. The 3*3*3 convolution of each Dense block is preceded by a 1*1*1 convolution operation, called a Bottleneck layer, whose purpose is to compress the number of input feature maps, fuse the features of each channel, and reduce the amount of calculation. Bottleneck The output result of the layer is used as the input of the 3*3*3 convolution. According to the network structure design of Densenet, the output of each layer should be connected with the output of all previous layers by channel as the input of the next layer. Therefore, the number of output channels of each Denseblock is also huge. In order to reduce memory usage and fuse the features of each output channel, there is a group of 1*1*1 convolution operations between every two Denseblocks, called the Transition layer, to reduce the number of output feature maps. Of course, some improvements have been made to the image block level network here, including two specific points: first, a Dilation Block composed of a hole convolution module is added after each DenseBlock to expand the receptive field of the convolution kernel; second, a compression-activation module is added after the 3×3×3 convolution of the Denseblock to obtain the weights of different channels of the feature map, and a bypass composed of a residual attention module is added to each Denseblock to obtain the weights of different voxels of the feature map.

[0063] Of course, the image block level network here can also be a three-dimensional residual network, such as ResNet, or it can also be a modified three-dimensional densely connected network (such as DenseNet) combined with a feature weighting module (SEBlock), etc.

[0064] Specifically, after obtaining the brain image, the computer device can divide the brain image into multiple image blocks according to the preset image block size, and then input each image block into its corresponding image block hierarchical network for feature extraction. The output of each image block hierarchical network is the feature map of the corresponding image block. Among them, the preset image block size here can be determined according to the actual situation, and this embodiment does not limit its size. Of course, when dividing the image blocks here, the image blocks can be divided into uniform sizes or different sizes, and this embodiment does not make specific limitations on this. For example, the size of the image blocks divided here can be 48*48*48, 64*64*64, 128*128*128, etc.

[0065] S206, performing a first weighted screening process on the feature map of each image block, and determining the feature maps of each candidate image block from the feature map of each image block, wherein the number of the candidate image blocks is less than the number of the plurality of image blocks.

[0066] Among them, the first weight screening process can be a pruning process performed by using an image block level pruning network, and the pruning process refers to a processing method of canceling the connection between unimportant branches in the network and retaining the connection between important branches. In this embodiment, the important branches can be considered as network branches corresponding to each candidate image block. In addition, the image block level pruning network can be a neural network model, which can be composed of a fully connected layer, or of course, other structures. This embodiment does not specifically limit this. It should be noted that the pruning process in a narrow sense is a non-differentiable operation and cannot participate in the optimization process of the network. Therefore, in actual operation, the pruning process is approximated by the fully connected layer + compression function P(x). For example, when performing the pruning process (i.e., weight screening process), assuming that the fully connected layer is composed of k input neurons and k output neurons, then for the k input feature maps, each value x of its k output values ​​can be compressed to a number between 0 and 1 by the compression function P(x). The formula of the compression function P(x) can be shown in the following formula (1):

[0067]

[0068] In this step, optionally, the first weight screening process is used to calculate the weight corresponding to the feature map of each image block, and obtain the feature map of each candidate image block according to the feature map of each image block and the corresponding weight. The feature map of each candidate image block is the feature map of each image block whose weight is greater than the preset threshold value; that is, the computer device can use formula (1) to perform weight calculation on the feature map of each image block, and compare the calculated weights with the preset threshold value to obtain the comparison results of each calculated weight, and take out the target calculated weights in the comparison results whose calculated weights are greater than the preset threshold value. The feature maps of the image blocks corresponding to these target weights can be used as the feature maps of the candidate image blocks. According to the properties of the P(x) function, the input number will be approximated to 0 or 1 by the threshold value, which is equivalent to adaptively pruning the feature maps of each image block; in addition, the preset threshold value in this step can be determined according to the actual situation, and can be 0.5, 0.7, etc. Secondly, in this embodiment, the feature map of each image block can be subjected to a weight screening process once, twice, or three or more times, which is not specifically limited in this embodiment.

[0069] Specifically, after obtaining the feature map of each image block, the computer device can perform weighted screening processing on the feature map of each image block, screen out at least one important feature map from the feature map of each image block as a candidate feature map, and record these candidate feature maps as feature maps of the candidate image block. Since generally multiple candidate image blocks are obtained, feature maps of multiple candidate image blocks can be obtained.

[0070] S208, inputting the feature map of each candidate image block into the neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0071] Among them, the neural network model can be a deep learning model, a convolutional neural network model, etc., and of course it can also be other network models; the target area refers to the area where at least one candidate image block selected by the neural network model from each candidate image block is located, or the candidate image blocks are combined and then processed by the neural network model to select the area obtained from the combined image block area; for example, the categories of brain images here can include Alzheimer's disease (AD), mild cognitive impairment (MCI), Parkinson's disease (PD), and normal (NC), among which MCI can include stable MCI (sMCI) and progressive MCI (pMCI) or early MCI (EMCI) and late MCI (LMCI). The category of the brain image here can actually be the category of the target area. Optionally, the category can be obtained by performing a softmax classification operation on the brain image.

[0072] It should be noted that the final output here can be the prediction index of each category, such as NC: 5%, sMCI: 7%, pMCI: 20%, AD: 68%. Finally, the category with the largest prediction index can be used as the category corresponding to the brain image.

[0073] Specifically, the computer device can input the feature map of each candidate image block into the neural network model, perform convolution and other related processing on each feature map through the neural network model, so as to obtain a target feature map in each feature map, and record the area where the image block corresponding to the target feature map is located as the target area, and perform classification operations on the brain image to obtain the category of the brain image.

[0074] In the above-mentioned brain image detection method, by obtaining a brain image of the object to be detected, the brain image is divided into multiple image blocks and input into the corresponding image block hierarchical network for feature extraction to obtain a feature map of each image block, and the feature map of each image block is subjected to a first weight screening process, and the feature map of each candidate image block is determined from the feature map of each image block. Finally, the feature map of each candidate image block is input into the neural network model to obtain the target area corresponding to the brain image and the category of the brain image, wherein the number of each candidate image block is less than the number of multiple image blocks. In this method, since the lesion area is obtained by detecting the brain image through a neural network instead of obtaining the lesion area through manual annotation, the detection result obtained by this method is more accurate and the detection efficiency of the brain image is higher; in addition, since the method can perform weight screening on the important features in the brain image when detecting the brain image, and obtain the important features for detection, the parameters of the network model can be reduced, and the detection speed can be accelerated, thereby further improving the detection efficiency.

[0075] In another embodiment, another brain image detection method is provided. This embodiment relates to a specific process of performing a first weight screening process on the feature map of the image block, and then performing a region combination on the feature map of each candidate image block, and further obtaining a region-level feature map. Based on the above embodiment, after the above step S206, as Figure 3 As shown, the above method may further include the following steps:

[0076] S302, combining the feature maps of each candidate image block by region to obtain feature maps of each combined region.

[0077] Among them, when performing regional combination, the feature maps of each candidate image block can be combined according to certain rules. For example, the rule can be based on the principle of proximity, etc. For example, the feature map of each candidate image block is a 1*1*1 feature map, and the feature maps of 8 candidate image blocks in a 2*2*2 neighborhood can be regionally combined, that is, the feature maps of 8 candidate image blocks are combined into 1 feature map, which can also be called a feature map group.

[0078] Specifically, after obtaining the feature map of each candidate image block, the computer device can perform regional combination of the feature maps of each candidate image block according to the position of the candidate image block, so as to obtain a feature map after regional combination, which is recorded as a combined regional feature map. Since there are generally many candidate image blocks, multiple combined regions are usually obtained after regional combination, so multiple combined regional feature maps can be obtained here (of course, only one combined regional feature map can be obtained in the end, but the present embodiment uses multiple combined regional feature maps).

[0079] S304, inputting each combined regional feature map into the corresponding regional hierarchical network for feature extraction to obtain a regional level feature map of each combined regional feature map.

[0080] Among them, the regional hierarchical network here can be a neural network model, including a deep learning model, a convolutional neural network model, etc., and of course it can also be other network models; in addition, the number of regional hierarchical networks can be the same as the number of combined regions. For example, assuming that there are 5 combined regions, there are 5 regional hierarchical networks, and the structure and parameters of each regional hierarchical network are the same. In addition, for example, the regional hierarchical network of this embodiment can be composed of 3 layers of 1×1×1 convolutional layers. For 8 adjacent image blocks in a 2*2*2 neighborhood, the regional hierarchical subnetwork corresponding to each image block outputs a set of 1*1*1 feature maps, and the feature maps of the 8 image blocks are combined into a feature map group, which is the feature representation of the regional level covered by the 8 image blocks. The combined feature map group is used as the input of the 1*1*1 convolutional layer, and a skip layer connection is added between the input and output of the convolutional layer, so that the convolutional layer learns the residual between the input and output, thereby increasing the learning ability of the network.

[0081] Specifically, after obtaining each combined area feature map, the computer device can input each combined area feature map into the corresponding regional hierarchical network. In the respective corresponding regional hierarchical networks, further feature extraction can be performed on each combined area feature. Then, a feature map corresponding to each combined area feature map can be obtained, which can be recorded as the feature map of each combined area feature map, or can also be called the regional-level feature map of each combined area.

[0082] The brain image detection method provided in this embodiment can further perform regional combination on the feature maps of each candidate image block after performing the first weight screening on each image block to obtain feature maps of each combined region, and input each combined region feature map into the corresponding regional hierarchical network for further feature extraction to obtain regional-level feature maps of each combined region. In this embodiment, since the feature maps of the candidate image blocks can be further extracted, the extracted features can be made more comprehensive and richer, and features at different scales can be integrated, so that the important feature maps of the brain image obtained will be more accurate, so that when the brain image is subsequently classified, the classification results obtained will be more accurate.

[0083] In another embodiment, another brain image detection method is provided. This embodiment involves a specific process of obtaining a candidate region by performing a second weight screening process on each region-level feature map after obtaining each region-level feature map. Based on the above embodiment, after the above step S304, the above method may further include:

[0084] Each region-level feature map is subjected to a second weight screening process, and a feature map of each candidate region is determined from each region-level feature map; the number of each candidate region is less than the number of each combined region.

[0085] Among them, the second weight screening process can be a pruning process performed by using a regional level pruning network, and its structure can be similar to the structure of the above-mentioned image block level pruning network. The pruning process here can be the same as the pruning process in the above-mentioned first weight screening process. However, in this embodiment, the important branches can be considered as the network branches corresponding to each combined area. In addition, the regional level pruning network can be a neural network model, which can be composed of a fully connected layer, and of course it can also be other structures. This embodiment does not specifically limit this. Secondly, the pruning process here is also the same as the first weight screening process, which is approximated by the fully connected layer + compression function P(x). The formula of the compression function P(x) is also the above formula (1).

[0086] In addition, in this step, optionally, the second weight screening process is used to calculate the weight corresponding to each region-level feature map, and obtain the feature map of each candidate region according to each region-level feature map and the corresponding weight; the feature map of each candidate region is the region-level feature map whose weight is greater than the preset threshold in each region-level feature map. That is to say, here, as in the first weight screening process, the computer device can use formula (1) to calculate the weight of each region-level feature map, and compare the calculated weights with the preset threshold to obtain the comparison results of each calculated weight, and take out the target calculated weights in the comparison results whose calculated weights are greater than the preset threshold. The region-level feature maps corresponding to these target weights can be used as the feature maps of the candidate region. According to the properties of the P(x) function, the input number will be approximated to 0 or 1 by the threshold, which is equivalent to adaptively pruning each region-level feature map; in addition, the preset threshold in this step can be determined according to the actual situation, which can be 0.5, 0.7, etc. Secondly, in this embodiment, after performing the secondary weight screening process on each region-level feature map, three or more weight screening processes can be performed, and this embodiment does not specifically limit this.

[0087] In addition, after each combined area is processed by weight screening, the number of candidate areas obtained is usually less than the number of each combined area. Here, it can be considered that one combined area corresponds to a feature map of the combined area, that is, corresponds to a region-level feature map, and one candidate area corresponds to a feature map of the candidate area.

[0088] Specifically, after obtaining each region-level feature map, the computer device can perform weight screening processing on each region-level feature map, and screen out at least one important region-level feature map from each region-level feature map as the feature map of the candidate region. Since generally multiple candidate regions are obtained, feature maps of multiple candidate regions can be obtained.

[0089] The brain image detection method provided in this embodiment can further perform a second weight screening process on each region-level feature map after obtaining each region-level feature map, and determine the feature map of each candidate region from each region-level feature map, wherein the number of each candidate region obtained is less than the number of each combined region. In this embodiment, since the further extracted feature map can be further screened, the extracted features will be richer, so that the obtained target region can be more accurate, and the subsequent classification results can be more accurate; in addition, since the number of further screened feature maps is small, the amount of calculation can be reduced when the target region is subsequently classified, thereby speeding up the detection speed.

[0090] In another embodiment, another brain image detection method is provided, and this embodiment relates to a specific process of how to obtain the target area and category of the brain image using an individual hierarchical network. Based on the above embodiment, the above S208 may include the following steps:

[0091] The feature maps of each candidate area are concatenated and input into the individual hierarchical network of the neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0092] In this step, the individual hierarchical network can be a neural network model, including a deep learning model, a convolutional neural network model, etc., and of course it can also be other network models; in addition, the number of individual hierarchical networks here can be one, and of course it can be multiple, but this embodiment uses one individual hierarchical network.

[0093] Optionally, when processing the feature map of each candidate region, Figure 4a The method steps shown are as follows: Figure 4a As shown, this step may include the following:

[0094] S402, feature stitching is performed on the feature maps of each candidate region and the feature maps are input into the individual hierarchical network of the neural network model for convolution processing to obtain the target region corresponding to the brain image.

[0095] S404: performing a classification operation on the target area corresponding to the brain image to obtain a category of the brain image.

[0096] In this embodiment, the individual-level subnetwork can be composed of a 1×1×1 convolution layer, and the outputs at different regional levels are combined to obtain the feature representation of the entire individual level. The features at the individual level are processed using a 1*1*1 convolution layer, and through a softmax operation, the final diagnosis result of the disease can be obtained. That is to say, the computer device can perform feature splicing on the feature maps of each candidate area, and then input the spliced ​​feature maps into the individual-level network for convolution processing to obtain a final feature map. The final feature map can represent the most important features of the brain image, and the corresponding area can be called the target area. After that, the target area is classified to obtain the category of the target area, that is, the category of the brain image.

[0097] In the embodiment of the present application, if the first weight screening process is performed by using an image block level pruning network, and the second weight screening process is performed by using a region level pruning network, then the process of using the 5-type network to detect brain images can be referred to. Figure 4bAs shown, the image blocks are input into the image block level network (also called the image block level subnetwork), and the output is input into the image block level pruning network. Then, the output of the image block level pruning network is combined by regions and then input into the regional level network (also called the regional level subnetwork). The output of the regional level network is input into the regional level pruning network. The output of the regional level pruning network is combined and input into the individual level network (also called the individual level subnetwork). Finally, the output is subjected to softmax classification operation to obtain the prediction result.

[0098] It should be noted that the networks involved in the embodiments of the present application can all be three-dimensional deep fully convolutional neural networks. The three-dimensional deep fully convolutional neural networks allow / can process all information including the entire lesion area and its surrounding structures. It takes into account the correlation between different layers and automatically obtains representations of multiple abstract levels of data. Therefore, useful features can be amplified, thereby obtaining image areas that are highly correlated with the lesions, thereby efficiently and accurately diagnosing brain atrophy.

[0099] The brain image detection method provided in this embodiment can obtain the target area corresponding to the brain image and the category of the brain image by splicing the feature maps of each candidate area and inputting them into the individual hierarchical network of the neural network model. In this embodiment, since the target area and category of the brain image can be obtained by using the individual hierarchical network, the method can simultaneously complete the detection of the key area of ​​the disease and the classification of the disease, thereby improving the detection accuracy and increasing the interpretability of the model.

[0100] In another embodiment, another brain image detection method is provided. This embodiment involves the specific process of how to train the image block hierarchical network, the region hierarchical network, and the individual hierarchical network. Figure 5 As shown, the training method may include the following steps:

[0101] S502, obtaining a sample brain image, where the sample brain image has been labeled with an image category.

[0102] The individuals corresponding to the samples here can be children, elderly people, adults, etc., but they are generally elderly people. The number of sample brain images can be determined according to the actual situation, for example, it can be 100-8000, etc., and this embodiment does not make a specific limitation on this. In addition, the sample brain images here can be brain images of individuals with normal brain development, or brain images of individuals with brain atrophy, that is, abnormal brain development. However, after each sample is collected, its corresponding category label can also be obtained.

[0103] For example, this embodiment uses an MRI scanning device to obtain a brain structure image of a scanned sample, for example, a T1 or T2 weighted image is obtained, and the label corresponding to each scanned sample is recorded, such as Alzheimer's disease (AD), mild cognitive impairment (MCI), and normal sample (NC), etc. In particular, for MCI samples, according to the changes in their pathological diagnosis in multiple scans during the follow-up stage (the diagnosis result is always MCI, or the corresponding diagnosis result is changed from MCI to AD in the later scan), the MCI samples are further subdivided into two categories: stable MCI (sMCI) and progressive MCI (pMCI). So far, there are a total of 6792 brain image samples, of which 1152 samples are AD, 2086 samples are sMCI, 2043 samples are pMCI, and 1511 samples are NC. In actual use, 80% of these samples can be randomly selected as training samples, and the remaining 20% ​​can be used as test samples. That is to say, this embodiment will select 80% as training samples, and the rest as test samples of S202, that is, samples to be tested.

[0104] Specifically, the computer device can reconstruct and correct the sample brain data collected by the CT device or PET device or MR device or CT-PET, PET-MR device to obtain the sample brain image. Alternatively, the sample brain image can also be reconstructed and corrected in advance and stored in the computer device. When it needs to be processed, the sample brain image is directly read from the memory of the computer device. Alternatively, the computer device can also obtain the sample brain image from an external device, for example, the sample brain image is stored in the cloud, and when the processing operation is required, the computer device obtains the sample brain image from the cloud. This embodiment does not limit the method for obtaining the sample brain image. In addition, the sample brain images here are all labeled with categories, and the labeled categories are the real categories corresponding to the samples.

[0105] S504, training the initial image block hierarchical network based on the sample brain image to obtain an image block hierarchical network.

[0106] Among them, after obtaining the sample brain image, the sample brain image can be first preprocessed. The image preprocessing can include at least one of the following processes: rotation processing, resampling processing, non-uniform field deviation field correction processing, skull removal processing, etc. After the image preprocessing, the image data input to the network can be standardized to facilitate subsequent unified processing.

[0107] Specifically, after obtaining the sample brain image, the computer device can first divide it into multiple sample image blocks, where the size of the divided image blocks can be determined according to actual conditions, for example, it can be 48*48*48, 64*64*64, 128*128*128, etc. After the division, the sample image blocks can be optionally input into the corresponding initial image block hierarchical network to obtain the predicted category corresponding to each sample image block, and based on the above-mentioned labeled category and predicted category, calculate the loss between the labeled category and the predicted category, and use the loss as the value of the loss function, and use the value of the loss function to train each initial image block hierarchical network, and finally obtain the trained image block hierarchical network. Here, the loss can be the error, variance, norm, etc. between the predicted category and the labeled category; when training the image block level network, when the sum of the loss functions of all the image block level networks is less than a preset threshold, or when the sum of the loss functions is basically stable (that is, it no longer changes), it can be determined that all the image block level networks have been trained, otherwise the training must continue. When the training is completed, the parameters of all the image block level networks can be fixed to facilitate the next step of training the regional level network.

[0108] S506, training the initial region-level network based on the image block-level network and the labeled image categories to obtain a region-level network.

[0109] Specifically, when training the regional hierarchical network here, the output of the last layer of each image block hierarchical network can be combined and input into the corresponding regional hierarchical network to obtain the predicted category, and the loss between the labeled category and the predicted category is calculated based on the above-mentioned labeled category and the predicted category, and the loss is used as the value of the loss function, and the value of the loss function is used to train each initial regional hierarchical network, and finally the trained regional hierarchical networks are obtained. When the training is completed, the parameters of all regional hierarchical networks can be fixed, which is convenient for the next step of training the individual hierarchical networks.

[0110] S508, based on the image block hierarchical network, the region hierarchical network and the labeled image categories, the initial individual hierarchical network is trained to obtain an individual hierarchical network.

[0111] Specifically, when training the individual hierarchical network here, the output of the last layer of each regional hierarchical network can be combined and input into the individual hierarchical network to obtain the predicted category, and the loss between the labeled category and the predicted category is calculated based on the above-mentioned labeled category and the predicted category, and the loss is used as the value of the loss function, and the value of the loss function is used to train each initial individual hierarchical network, and finally a trained individual hierarchical network is obtained, and the parameters of the individual hierarchical network are fixed.

[0112] It should be noted that if the above-mentioned first weight screening process is performed using an image block level pruning network, and the above-mentioned second weight screening process is performed using a region level pruning network, then after the image block level network, the region level network, and the individual level network are trained, the trained image block level network, the region level network, the individual level network, the image block level pruning network, and the region level pruning network can be combined for joint training. During joint training, the loss between the category output by the last network and the labeled category can also be used for training. After the joint training is completed, the final network parameters can be obtained. After the final training converges, the network branches with weights close to 0 are manually pruned (i.e., the connection is canceled) according to the fully connected layer parameters of the two pruned networks, thereby reducing the computational overhead in practical applications.

[0113] The brain image detection method provided in the present embodiment obtains a sample brain image, and the sample brain image has been labeled with an image category; based on the sample brain image, an initial image block hierarchical network is trained to obtain an image block hierarchical network; based on the image block hierarchical network and the labeled image category, an initial regional hierarchical network is trained to obtain a regional hierarchical network; based on the image block hierarchical network, the regional hierarchical network and the labeled image category, an initial individual hierarchical network is trained to obtain an individual hierarchical network. In the present embodiment, since each network is trained using sample brain images of labeled categories, each network obtained is relatively accurate, and when the accurate network is used for detection, the target area and category obtained are also relatively accurate.

[0114] It should be understood that although Figure 2 , 3 The steps in the flowcharts of , 4a, and 5 are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2 , 3 At least part of the steps in 4a, 4b, and 5 may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily have to be executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages does not necessarily have to be sequentially, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0115] In one embodiment, Figure 6 As shown, a brain image detection device is provided, comprising: an acquisition module 10, an extraction module 11, a first screening processing module 12 and a determination module 13, wherein:

[0116] An acquisition module 10 is used to acquire a brain image of a subject to be detected;

[0117] An extraction module 11 is used to divide the brain image into a plurality of image blocks, and input each of the image blocks into a corresponding image block hierarchical network for feature extraction to obtain a feature map of each of the image blocks;

[0118] A first screening processing module 12 is used to perform a first weight screening process on the feature graphs of each of the image blocks, and determine the feature graphs of each candidate image block from the feature graphs of each of the image blocks, wherein the number of the candidate image blocks is less than the number of the plurality of image blocks;

[0119] The determination module 13 is used to input the feature map of each candidate image block into a neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0120] For the specific limitations on the brain image detection device, please refer to the limitations on the brain image detection method above, which will not be repeated here.

[0121] In another embodiment, another brain image detection device is provided. Based on the above embodiment, the above device may further include: a combination determination module, wherein:

[0122] The combination determination module is used to perform regional combination on the feature maps of each candidate image block to obtain each combined regional feature map; each combined regional feature map is input into the corresponding regional hierarchical network for feature extraction to obtain a regional level feature map of each combined regional feature map.

[0123] In another embodiment, another brain image detection device is provided. Based on the above embodiment, the device may further include: a second screening processing module, wherein:

[0124] The second screening processing module is used to perform a second weight screening process on each of the region-level feature maps to determine the feature map of each candidate region from each of the region-level feature maps; the number of the candidate regions is less than the number of the combined regions.

[0125] In another embodiment, the determination module 13 is further used to perform feature splicing on the feature maps of each candidate area and input the feature maps into the individual hierarchical network of the neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0126] In another embodiment, the first weight screening process is used to calculate the weight corresponding to the feature map of each of the image blocks, and obtain the feature map of each candidate image block based on the feature map of each of the image blocks and the corresponding weight; the feature map of each candidate image block is a feature map in the feature map of each of the image blocks whose weight is greater than a preset threshold.

[0127] In another embodiment, the second weight screening process is used to calculate the weights corresponding to each of the region-level feature maps, and obtain the feature maps of each candidate region based on each of the region-level feature maps and the corresponding weights; the feature maps of each candidate region are the region-level feature maps whose weights are greater than a preset threshold in the region-level feature maps.

[0128] In another embodiment, the determination module 13 is further used to perform feature splicing on the feature maps of each candidate area and input the convolution processing into the individual hierarchical network of the neural network model to obtain the target area corresponding to the brain image; and perform classification operation on the target area corresponding to the brain image to obtain the category of the brain image.

[0129] In another embodiment, another brain image detection device is provided. Based on the above embodiment, the device may further include: a training module, wherein:

[0130] A training module is used to obtain a sample brain image, wherein the sample brain image has annotated image categories; train an initial image block hierarchical network based on the sample brain image to obtain the image block hierarchical network; train an initial regional hierarchical network based on the image block hierarchical network and the annotated image categories to obtain the regional hierarchical network; train an initial individual hierarchical network based on the image block hierarchical network, the regional hierarchical network and the annotated image categories to obtain the individual hierarchical network.

[0131] For the specific limitations on the brain image detection device, please refer to the limitations on the brain image detection method above, which will not be repeated here.

[0132] Each module in the above-mentioned brain image detection device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0133] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0134] Acquiring a brain image of a subject to be detected;

[0135] Dividing the brain image into a plurality of image blocks, and inputting each of the image blocks into a corresponding image block hierarchical network for feature extraction to obtain a feature map of each of the image blocks;

[0136] Performing a first weighted screening process on the feature graphs of each of the image blocks, and determining feature graphs of each candidate image block from the feature graphs of each of the image blocks, wherein the number of the candidate image blocks is less than the number of the plurality of image blocks;

[0137] The feature map of each candidate image block is input into a neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0138] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0139] Performing regional combination on the feature maps of each of the candidate image blocks to obtain feature maps of each combined region;

[0140] Each of the combined regional feature maps is input into a corresponding regional hierarchical network for feature extraction to obtain a regional level feature map of each of the combined regional feature maps.

[0141] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0142] A second weight screening process is performed on each of the region-level feature maps to determine a feature map of each candidate region from each of the region-level feature maps; the number of the candidate regions is less than the number of the combined regions.

[0143] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0144] The feature maps of the candidate regions are spliced ​​and input into the individual hierarchical network of the neural network model to obtain the target region corresponding to the brain image and the category of the brain image.

[0145] In one embodiment, the first weight screening process is used to calculate the weight corresponding to the feature map of each of the image blocks, and obtain the feature map of each candidate image block based on the feature map of each of the image blocks and the corresponding weight; the feature map of each candidate image block is a feature map in the feature map of each of the image blocks whose weight is greater than a preset threshold.

[0146] In one embodiment, the second weight screening process is used to calculate the weights corresponding to each of the region-level feature maps, and obtain the feature maps of each candidate region based on each of the region-level feature maps and the corresponding weights; the feature maps of each candidate region are region-level feature maps whose weights are greater than a preset threshold in each of the region-level feature maps.

[0147] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0148] The feature maps of each candidate area are spliced ​​and then input into the individual hierarchical network of the neural network model for convolution processing to obtain the target area corresponding to the brain image; the target area corresponding to the brain image is classified to obtain the category of the brain image.

[0149] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0150] A sample brain image is obtained, wherein the sample brain image has been labeled with an image category; an initial image block hierarchical network is trained based on the sample brain image to obtain the image block hierarchical network; an initial regional hierarchical network is trained based on the image block hierarchical network and the labeled image category to obtain the regional hierarchical network; an initial individual hierarchical network is trained based on the image block hierarchical network, the regional hierarchical network and the labeled image category to obtain the individual hierarchical network.

[0151] In one embodiment, a readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0152] Acquiring a brain image of a subject to be detected;

[0153] Dividing the brain image into a plurality of image blocks, and inputting each of the image blocks into a corresponding image block hierarchical network for feature extraction to obtain a feature map of each of the image blocks;

[0154] Performing a first weighted screening process on the feature graphs of each of the image blocks, and determining feature graphs of each candidate image block from the feature graphs of each of the image blocks, wherein the number of the candidate image blocks is less than the number of the plurality of image blocks;

[0155] The feature map of each candidate image block is input into a neural network model to obtain the target area corresponding to the brain image and the category of the brain image.

[0156] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0157] The feature maps of each candidate image block are regionally combined to obtain each combined regional feature map; each combined regional feature map is input into a corresponding regional hierarchical network for feature extraction to obtain a regional level feature map of each combined regional feature map.

[0158] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0159] A second weight screening process is performed on each of the region-level feature maps to determine a feature map of each candidate region from each of the region-level feature maps; the number of the candidate regions is less than the number of the combined regions.

[0160] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0161] The feature maps of the candidate regions are spliced ​​and input into the individual hierarchical network of the neural network model to obtain the target region corresponding to the brain image and the category of the brain image.

[0162] In one embodiment, the first weight screening process is used to calculate the weight corresponding to the feature map of each of the image blocks, and obtain the feature map of each candidate image block based on the feature map of each of the image blocks and the corresponding weight; the feature map of each candidate image block is a feature map in the feature map of each of the image blocks whose weight is greater than a preset threshold.

[0163] In one embodiment, the second weight screening process is used to calculate the weights corresponding to each of the region-level feature maps, and obtain the feature maps of each candidate region based on each of the region-level feature maps and the corresponding weights; the feature maps of each candidate region are region-level feature maps whose weights are greater than a preset threshold in each of the region-level feature maps.

[0164] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0165] The feature maps of each candidate area are spliced ​​and then input into the individual hierarchical network of the neural network model for convolution processing to obtain the target area corresponding to the brain image; the target area corresponding to the brain image is classified to obtain the category of the brain image.

[0166] In one embodiment, when the computer program is executed by a processor, the following steps are also implemented:

[0167] A sample brain image is obtained, wherein the sample brain image has been labeled with an image category; an initial image block hierarchical network is trained based on the sample brain image to obtain the image block hierarchical network; an initial regional hierarchical network is trained based on the image block hierarchical network and the labeled image category to obtain the regional hierarchical network; an initial individual hierarchical network is trained based on the image block hierarchical network, the regional hierarchical network and the labeled image category to obtain the individual hierarchical network.

[0168] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. 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. As an 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 (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0169] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, 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, they should be considered to be within the scope of this specification.

[0170] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A brain image detection method, characterized in that: The method comprises: Acquiring a brain image of a subject to be detected; Dividing the brain image into a plurality of image blocks, and inputting each of the image blocks into a corresponding image block hierarchical network for feature extraction to obtain a feature map of each of the image blocks; A first weight screening process is performed on the feature map of each of the image blocks using a preset first pruning network, and a feature map of each candidate image block is determined from the feature map of each of the image blocks; the number of the candidate image blocks is less than the number of the plurality of image blocks, and the first weight screening process is used to calculate the weight corresponding to the feature map of each of the image blocks, and obtain the feature map of each candidate image block according to the feature map of each of the image blocks and the corresponding weight; the feature map of each of the candidate image blocks is input into a neural network model to obtain a target area corresponding to the brain image and a category of the brain image; The first pruning network includes a fully connected layer and a preset compression function; the first pruning network is used to perform a first weight screening process on the feature graphs of each of the image blocks, and the feature graphs of each candidate image block are determined from the feature graphs of each of the image blocks, including: The feature map of each image block is input into the fully connected layer to obtain the output value corresponding to the feature map of each image block, and the compression function is used to compress each output value to obtain the weight corresponding to the feature map of each image block; according to the weight corresponding to the feature map of each image block, the feature map of each candidate image block is screened from the feature map of each image block.

2. The method according to claim 1, characterized in that: After performing the first weight screening process on the feature graphs of each of the image blocks and determining the feature graphs of each candidate image block from the feature graphs of each of the image blocks, the method further includes: Performing regional combination on the feature maps of each of the candidate image blocks to obtain feature maps of each combined region; Inputting each of the combined regional feature maps into a corresponding regional hierarchical network for feature extraction to obtain a regional level feature map of each of the combined regional feature maps; Performing a second weight screening process on each of the region-level feature maps, and determining a feature map of each candidate region from each of the region-level feature maps; the number of the candidate regions is less than the number of the combined regions; Accordingly, the feature map of each candidate image block is input into a neural network model to obtain the target area corresponding to the brain image and the category of the brain image, including: The feature maps of the candidate regions are spliced ​​and input into the individual hierarchical network of the neural network model to obtain the target region corresponding to the brain image and the category of the brain image.

3. The method according to claim 2, characterized in that The step of performing a first weighted screening process on the feature graphs of each of the image blocks to determine the feature graphs of each candidate image block from the feature graphs of each of the image blocks includes: Using a preset first pruning network to perform a first weight screening process on the feature graphs of each of the image blocks, and determining the feature graphs of each candidate image block from the feature graphs of each of the image blocks; The performing a second weight screening process on each of the region-level feature maps to determine a feature map of each candidate region from each of the region-level feature maps includes: Using a preset second pruning network to perform a second weight screening process on each of the region-level feature maps, and determining a feature map of each candidate region from each of the region-level feature maps; Wherein, the first pruned network and the second pruned network are both neural networks.

4. The method according to claim 2, characterized in that: The feature map of each candidate image block is a feature map whose weight is greater than a preset threshold in the feature map of each image block; the second weight screening process is used to calculate the weight corresponding to each of the region-level feature maps, and obtain the feature map of each candidate region based on each of the region-level feature maps and the corresponding weights; the feature map of each candidate region is a region-level feature map whose weight is greater than a preset threshold in the region-level feature map.

5. The method according to claim 2, characterized in that: The feature splicing of the feature maps of the candidate regions is input into the individual hierarchical network of the neural network model to obtain the target region corresponding to the brain image and the category of the brain image, including: The feature maps of each candidate area are concatenated and input into the individual hierarchical network of the neural network model for convolution processing to obtain the target area corresponding to the brain image; A classification operation is performed on the target area corresponding to the brain image to obtain the category of the brain image.

6. The method according to claim 5, characterized in that The training methods of the image block level network, the region level network, and the individual level network include: Acquire a sample brain image, wherein the sample brain image has been labeled with an image category; Training an initial image block hierarchical network based on the sample brain image to obtain the image block hierarchical network; Training an initial region-level network based on the image block-level network and the labeled image categories to obtain the region-level network; Based on the image block level network, the region level network and the labeled image category, an initial individual level network is trained to obtain the individual level network.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

8. A readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images