Method and device for identifying cerebral micro-hemorrhage

Abstract

The embodiment of the invention discloses a brain micro-hemorrhage identification method and device, and the method comprises the steps: obtaining magnetic resonance images inputted by two channels which are an SWI image input channel and a phase image input channel; target detection is carried out on the magnetic resonance image based on a preset target detection model to screen out a micro-bleeding candidate area, the candidate area is cut into a 3D patch image, and before the candidate area is cut into the 3D patch image, the magnetic resonance image containing the candidate area is segmented based on BET to segment out a brain tissue area and a non-brain tissue area; cutting a candidate region contained in the brain tissue region to obtain a 3D patch image; and inputting the 3D patch image into a preset convolutional neural network model for context learning so as to determine a micro-bleeding area and a false positive area. The automatic detection efficiency and accuracy of cerebral microhemorrhage are improved, and more reliable support is provided for medical imaging diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a method and device for identifying cerebral microbleeds. Background Art

[0002] At present, automatic detection technology in the field of medical imaging plays an increasingly important role in the diagnosis and treatment process. In particular, the automatic detection of cerebral microbleeds (CMBs) in brain magnetic resonance images is crucial for early diagnosis and disease monitoring. Traditional methods mainly rely on manual extraction, which has limitations and makes it difficult to fully explore the potential information in image data. In recent years, supervised deep learning convolutional neural networks (CNNs) have attracted much attention in the field of medical imaging. They can learn and extract effective feature hierarchies, thereby significantly improving detection performance.

[0003] The main content of the existing technology includes the use of CNNs for automatic detection and classification, usually using a two-stage framework: the detection stage is used to discover potential candidate regions, and the classification or differentiation stage is used to reduce false positives. This two-stage CNN approach has achieved remarkable results in medical imaging applications such as brain, breast, lung and skin cancer. In the detection of cerebral microbleeds, SWI images are usually used for automatic detection, in which 3D fully convolutional networks (3D-FCN) and 3D-CNN are used as screening and differentiation stages. This cascade network can reduce the number of false positives while improving sensitivity, thereby improving the accuracy and efficiency of detection.

[0004] However, existing technologies also have some problems and shortcomings. First, traditional methods rely on manually designed features. This feature extraction method may not fully capture complex image information, resulting in unstable detection performance. Second, some methods may face challenges such as high computational costs and optimization difficulties when processing large amounts of data, affecting the practicality and efficiency of the algorithm. In addition, some methods may have difficulty in identifying edges and complex situations, resulting in missed detections or false detections.

[0005] The root of these problems lies in the limitations of the feature extraction method and algorithm structure design of traditional methods, which are difficult to adapt to the complex characteristics of medical imaging data. In addition, the limitation of computing resources and the difficulty of algorithm optimization also limit the application and promotion of existing technologies on large-scale data sets. Therefore, further research and improvement of existing technologies are needed to improve the accuracy, stability and efficiency of automatic detection of cerebral microbleeds and provide more reliable support for clinical diagnosis and treatment. Summary of the invention

[0006] The main purpose of the present invention is to provide a method and device for identifying cerebral microbleeds to solve the technical problems existing in the related art.

[0007] In order to achieve the above-mentioned purpose, according to a first aspect of the present invention, a method for identifying cerebral microbleeds is provided, comprising: acquiring a magnetic resonance image inputted by two channels, wherein the two channels are a SWI image input channel and a phase image input channel; performing target detection on the magnetic resonance image based on a preset target detection model to screen out candidate areas for microbleeds, and cropping the candidate areas into 3D patch images, wherein before cropping the candidate areas into 3D patch images, segmenting the magnetic resonance image containing the candidate areas based on BET to segment brain tissue areas and non-brain tissue areas; cropping the candidate areas contained in the brain tissue areas to obtain 3D patch images; and inputting the 3D patch images into a preset convolutional neural network model for context learning to determine the microbleeding areas, and the false positive areas.

[0008] Optionally, inputting the 3D patch image into a preset convolutional neural network model for context learning includes: determining 3D context information based on average values ​​of adjacent slices of the SWI and phase images: ,in, Representative The slice The value of the channel, is the channel number, Represents the SWI channel, stands for complementary phase channel; Indicates The SWI value of each slice, Indicates The phase value of each slice, is the index of the current slice, is the index of adjacent slices, ranging from arrive .

[0009] Optionally, performing target detection on the magnetic resonance image based on a preset target detection model includes: performing target detection on the magnetic resonance image based on a YOLO detection model with a preset structure, wherein the YOLO detection model includes N convolutional layers, the kernel size of the convolutional layer is 3 × 3; m maximum pooling layers, the size of the pooling layer is 2 × 2, and the stride is 2; a 1 × 1 convolutional layer, called a bottleneck layer, is inserted between every two consecutive 3 × 3 convolutional layers; each convolutional layer is followed by batch normalization and leaky activation functions, the last convolutional layer is followed by a linear activation function, and the number of features of the last convolutional layer is: a multimodal logistic regression layer is placed on the top layer to generate a tensor containing detected bounding box information and predicted categories; wherein, during the training process, the network is optimized by minimizing a loss function, which is determined based on the ground truth value and the predicted bounding box.

[0010] Optionally, the preset convolutional neural network model includes five convolutional layers, the convolution kernel size is 3 × 3 × 3, and the output is: Y = X × W + b, where Y is the output of the convolution layer, X is the input of the convolution layer, W is the convolution kernel, and the bias term is b; each convolution layer is followed by batch normalization and a rectified linear unit activation function, and the normalized input is ,in is the normalized input, is the input mean, is the variance, is the error term, and the final output is , is the output after batch normalization, is the scaling parameter, is the offset parameter; the ReLU activation function is applied to the normalized output: , then down-sampling is performed using a pooling layer; the features extracted by the convolutional layer are flattened and passed to three fully connected layers, where the number of units in the last fully connected layer corresponds to the number of classification categories, namely, microbleeding area categories, and false positive area categories, where a random dropout layer is used after the first fully connected layer to regularize the network and prevent overfitting.

[0011] According to a second aspect of the present invention, a device for identifying cerebral microbleeds is provided, comprising an image acquisition unit for acquiring magnetic resonance images input from two channels, wherein the two channels are a SWI image input channel and a phase image input channel; a microbleed candidate region detection unit for performing target detection on the magnetic resonance image based on a preset target detection model to screen out candidate regions for microbleeds, and cropping the candidate regions into 3D patch images, wherein before cropping the candidate regions into the 3D patch images, the magnetic resonance image containing the candidate regions is segmented based on BET to segment brain tissue regions and non-brain tissue regions; the candidate regions contained in the brain tissue regions are cropped to obtain the 3D patch images; and a classification unit for inputting the 3D patch images into a preset convolutional neural network model for context learning to determine the microbleed regions, and the false positive regions.

[0012] Optionally, inputting the 3D patch image into a preset convolutional neural network model for context learning includes: determining 3D context information based on average values ​​of adjacent slices of the SWI and phase images: ,in, Representative The slice The value of the channel, is the channel number, Represents the SWI channel, stands for complementary phase channel; Indicates The SWI value of each slice, Indicates The phase value of each slice, is the index of the current slice, is the index of adjacent slices, ranging from arrive .

[0013] Optionally, performing target detection on the magnetic resonance image based on a preset target detection model includes: performing target detection on the magnetic resonance image based on a YOLO detection model with a preset structure, wherein the YOLO detection model includes N convolutional layers, the kernel size of the convolutional layer is 3 × 3; m maximum pooling layers, the size of the pooling layer is 2 × 2, and the stride is 2; a 1 × 1 convolutional layer, called a bottleneck layer, is inserted between every two consecutive 3 × 3 convolutional layers; each convolutional layer is followed by batch normalization and leaky activation functions, the last convolutional layer is followed by a linear activation function, and the number of features of the last convolutional layer is: a multimodal logistic regression layer is placed on the top layer to generate a tensor containing detected bounding box information and predicted categories; wherein, during the training process, the network is optimized by minimizing a loss function, which is determined based on the ground truth value and the predicted bounding box.

[0014] Optionally, the preset convolutional neural network model includes five convolutional layers, the convolution kernel size is 3 × 3 × 3, and the output is: Y = X × W + b, where Y is the output of the convolution layer, X is the input of the convolution layer, W is the convolution kernel, and the bias term is b; each convolution layer is followed by batch normalization and a rectified linear unit activation function, and the normalized input is ,in is the normalized input, is the input mean, is the variance, is the error term, and the final output is , is the output after batch normalization, is the scaling parameter, is the offset parameter; the ReLU activation function is applied to the normalized output: , then down-sampling is performed using a pooling layer; the features extracted by the convolutional layer are flattened and passed to three fully connected layers, where the number of units in the last fully connected layer corresponds to the number of classification categories, namely, microbleeding area categories, and false positive area categories, where a random dropout layer is used after the first fully connected layer to regularize the network and prevent overfitting.

[0015] According to a third aspect of the present invention, there is provided a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute any one of the methods described in the first aspect.

[0016] According to a fourth aspect of the present invention, there is provided an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor executes the method described in any one implementation of the first aspect.

[0017] The present embodiment provides a method and device for identifying cerebral microbleeds, wherein the method includes obtaining a magnetic resonance image inputted from two channels, wherein the two channels are a SWI image input channel and a phase image input channel; performing target detection on the magnetic resonance image based on a preset target detection model to screen out candidate regions for microbleeds, and cropping the candidate regions into 3D patch images, wherein before cropping the candidate regions into 3D patch images, the magnetic resonance image containing the candidate regions is segmented based on BET to segment brain tissue regions and non-brain tissue regions; the candidate regions contained in the brain tissue regions are cropped to obtain 3D patch images; and the 3D patch images are inputted into a preset convolutional neural network model for contextual learning to determine microbleed regions and false positive regions. By introducing deep learning technology, the problems of insufficient feature extraction, high computational cost, and difficulty in identification in existing research technologies are solved, thereby improving the efficiency and accuracy of automatic detection of cerebral microbleeds and providing more reliable support for clinical medical imaging diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 is a flow chart of a method for identifying cerebral microbleeds according to an embodiment of the present invention; Figure 2 is an application diagram of the method for identifying cerebral microbleeds according to an embodiment of the present invention; Figure 3 is another application diagram of the method for identifying cerebral microbleeds according to an embodiment of the present invention; Figure 4 is another application diagram of the method for identifying cerebral microbleeds according to an embodiment of the present invention; Figure 5 is a schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first", "second", etc. in the specification of the present invention are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0022] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0023] According to an embodiment of the present invention, a flow chart of a method for identifying cerebral microbleeds is provided, such as Figure 1 As shown, it includes the following steps 101 to 103: Step 101: Acquire magnetic resonance images inputted from two channels, wherein the two channels are a SWI image input channel and a phase image input channel.

[0024] In this step, after inputting the magnetic resonance images of two channels, the input composite MR image is divided into 7×7 non-overlapping grid cells. For each microbleed appearing in the input image, only one grid cell is responsible for detecting it.

[0025] Step 102: Perform target detection on the magnetic resonance image based on a preset target detection model to screen out candidate areas for microbleeding, and crop the candidate areas into 3D patch images, wherein before cropping the candidate areas into 3D patch images, the magnetic resonance image containing the candidate areas is segmented based on BET to segment brain tissue areas and non-brain tissue areas; the candidate areas contained in the brain tissue areas are cropped to obtain the 3D patch images.

[0026] In this step, each grid cell is represented by B bounding boxes (called anchor boxes), which contain five components: x, y, width, height, and confidence. Among them, (x, y) represents the coordinates of the center of the microbleed in the MR image, and the confidence reflects the probability of the presence of microbleeds within a certain width and height range of the center. In other words, the confidence can be defined as the intersection ratio between the bounding box predicted by YOLO and the true label annotated by the neuroradiologist. If there is no microbleed in a grid cell, the confidence of the bounding box of the cell should be zero. The (x, y) coordinates, width, and height are normalized to the range [0, 1] relative to the size of the MR image, as follows: ,in, , , and They represent the original pixel intensity, normalized pixel intensity, minimum pixel intensity, and maximum pixel intensity, respectively.

[0027] Furthermore, the brain extraction tool (BET) was applied to segment the head into brain tissue and non-brain tissue. The brain tissue binary mask generated by BET was applied to SWI and phase images. This process enabled the computer-aided detection algorithm to avoid detecting more erroneous CMBs candidate regions outside the brain tissue area (such as the skull area). Figure 2 Some examples of applying BET processing on high-resolution and low-resolution SWI images are shown.

[0028] Step 103: Input the 3D patch image into a preset convolutional neural network model for context learning to determine the microbleeding area and the false positive area.

[0029] In this step, detected candidate regions are obtained from those bounding boxes with high confidence scores, and then all these regions (including CMBs and non-CMBs) are combined with small 3D patches from SWI and phase images and passed to the 3D-CNN discriminative stage to further reduce false positives.

[0030] As an optional implementation of this embodiment, the 3D patch image is input into a preset convolutional neural network model for context learning, including determining the 3D context information based on the average value of adjacent slices of the SWI and phase image: ,in, Representative The slice The value of the channel, is the channel number, Represents the SWI channel, stands for complementary phase channel; Indicates The SWI value of each slice, Indicates The phase value of each slice, is the index of the current slice, is the index of adjacent slices, ranging from arrive .

[0031] In this alternative implementation, the proposed 3D-CNN is trained using susceptibility weighted imaging (SWI) and phase images, each with a size of 16×16×16 voxels. Therefore, the 3D-CNN stage is able to learn richer contextual spatial information of CMBs from 3D training samples with a size of 16×32×16 voxels. The 3D context information is obtained using the average of adjacent slices of SWI and complementary phase images:

[0032] in, Representative The slice The value of the channel, is the channel number, Represents the SWI channel, stands for complementary phase channel; Indicates The SWI value of each slice, Indicates The phase value of each slice, is the index of the current slice, is the index of adjacent slices, ranging from arrive By averaging adjacent slices, two channels of input are generated, one is the SWI channel and the other is the complementary phase channel. This can increase the 3D context information of the image and help detect CMBs more accurately.

[0033] As an optional implementation method of this embodiment, performing target detection on a magnetic resonance image based on a preset target detection model includes: performing target detection on a magnetic resonance image based on a YOLO detection model with a preset structure, wherein the YOLO detection model includes N convolutional layers, and the kernel size of the convolutional layer is 3 × 3; m maximum pooling layers, the size of the pooling layer is 2 × 2, and the stride is 2; a 1 × 1 convolutional layer is inserted between every two consecutive 3 × 3 convolutional layers, called a bottleneck layer; each convolutional layer is followed by batch normalization and leaky activation functions, the last convolutional layer is followed by a linear activation function, and the number of features of the last convolutional layer is: a multimodal logistic regression layer is placed on the top layer to generate a tensor containing detected bounding box information and predicted categories; wherein, during the training process, the network is optimized by minimizing the loss function, and the loss function is determined based on the ground truth value and the predicted bounding box.

[0034] In this optional implementation, the YOLO algorithm is used to screen candidate regions for microbleeds in brain images. The region-based YOLO configuration used contains 23 convolutional layers with a kernel size of 3×3 and various feature maps, and 5 maximum pooling layers with a size of 2×2 and a stride of 2. A 1×1 convolutional layer, called a bottleneck layer, is inserted between every two consecutive 3×3 convolutional layers to speed up the calculation by reducing the feature maps. The subsampling layer reduces the input MR brain image from 448×448 pixels to 14×14 pixels. In addition, each convolutional layer is followed by batch normalization and leaky activation functions, except for the last convolutional layer, which is followed by a linear activation function. Taking advantage of the residual network, the proposed region-based YOLO adds a pass layer that connects the high-resolution features of the early layers with the low-resolution features. The number of features of the last convolutional layer is calculated as follows: Number of features = B×(bounding box components + C), where B and C represent the number of anchor boxes and the number of categories, respectively. Where B = 5 to obtain most possible bounding boxes associated with different aspect ratios that may contain microbleeds. Since only one category (i.e., CMBs) was used during the training phase, C = 1 was set. Therefore, the number of features in the last convolutional layer was 30. Finally, a multimodal logistic regression layer (softmax classifier) ​​was placed on the top layer of the network to generate a tensor containing the detected bounding box information and the predicted category. During training, region-based YOLO optimizes the network by minimizing the loss function, which is calculated based on the ground truth and the predicted bounding box. The designed loss function contains the sum of four parts related to the coordinates (x, y), width and height, confidence score, and classification loss of the detected bounding box and the ground truth. The algorithm uses a batch size of 64, a learning rate and decay of 0.0005, and a momentum of 0.9.

[0035] As an optional implementation of this embodiment, the preset convolutional neural network model includes five convolutional layers, the convolution kernel size is 3 × 3 × 3, and the output is: Y = X × W + b, where Y is the output of the convolution layer, X is the input of the convolution layer, W is the convolution kernel, and the bias term is b; each convolution layer is followed by batch normalization and a rectified linear unit activation function, and the normalized input is ,in is the normalized input, is the input mean, is the variance, is the error term, and the final output is , is the output after batch normalization, is the scaling parameter, is the offset parameter; the ReLU activation function is applied to the normalized output: , then down-sampling is performed using a pooling layer; the features extracted by the convolutional layer are flattened and passed to three fully connected layers, where the number of units in the last fully connected layer corresponds to the number of classification categories, namely, microbleeding area categories, and false positive area categories, where a random dropout layer is used after the first fully connected layer to regularize the network and prevent overfitting.

[0036] In this optional implementation, the proposed 3D-CNN structure contains five convolutional layers, the convolution kernel size is 3×3×3, and the output is: Y=X×W+b, where Y is the output of the convolution layer, X is the input of the convolution layer, W is the convolution kernel, and the bias term is b; in order to increase the stability of the network and speed up the learning process, batch normalization and ReLU (rectified linear unit) activation functions are attached to each convolution layer. The normalized input is ,in is the normalized input, is the input mean, is the variance, is the error term, and the final output is , is the output after batch normalization, is the scaling parameter, is the offset parameter; ReLU activation function is applied to the normalized output , then down-sampled using a 2×2×2 max pooling layer with a stride of 2.

[0037] The features extracted by these convolutional layers are flattened and passed to three fully-connected (FC) layers. The number of units in the last FC layer corresponds to the number of categories in the task, i.e., CMBs and non-CMBs. In addition, a random dropout layer is used after the first FC layer to regularize the network and prevent overfitting. The 3D-CNN network uses a learning rate of 0.001, a batch size of 50, and 200 training epochs.

[0038] refer to Figure 3 The schematic diagram of the network architecture in the method of this embodiment is illustrated, which is a deep learning two-stage method for detecting CMBs, which combines the region-based YOLO algorithm and the 3D convolutional neural network (3D-CNN). First, the region-based YOLO algorithm is used to detect potential candidate regions, and the input magnetic resonance (MR) image is processed to generate multiple bounding boxes associated with confidence scores to identify all candidate regions that may contain microbleeds. Then, these candidate regions representing microbleeds and false positives are cropped into small 3D patches and input into the 3D-CNN stage to distinguish actual CMBs from false positives, thereby reducing false positives.

[0039] Figure 4 Some examples of true positive, false positive, and false negative results generated by the algorithm at two resolutions are shown. Ultimately, for the high-resolution data, the algorithm accurately detected 176 of the 188 microbleeds and missed only 12. For the low-resolution data, the overall detection performance dropped, correctly detecting 451 of the 472 microbleeds and missing 21. The first two rows are high-resolution images, and the last three rows are low-resolution images. In each resolution, the first three rows are high-resolution images, and the last three rows are low-resolution images. In each resolution, the first row represents true positive results, the second row represents false positive results, and the third row represents false negative results.

[0040] This embodiment firstly introduces a deep learning network structure, and the algorithm can automatically learn and extract effective features in the image, thereby improving the detection quality and accuracy of cerebral microbleeds. Compared with the traditional manual feature extraction, the deep learning method can better capture the complex information in the image, reduce missed detection and false detection, and improve the accuracy of detection.

[0041] Secondly, the algorithm uses a two-stage deep learning framework, combining the detection stage and the classification stage, which effectively reduces the number of false positives and improves the yield and quality of detection. By screening potential candidate areas and accurately classifying them, the algorithm can improve the detection efficiency of cerebral microbleeds, reduce the workload of doctors, and improve the speed and accuracy of diagnosis.

[0042] In addition, the algorithm optimizes the deep learning network structure and algorithm design, improves the computing efficiency and speed on large data sets, and saves computing resources and time costs. This not only improves the practicality and efficiency of the algorithm, but also reduces the cost of medical imaging diagnosis, providing a more convenient and economical solution for clinical medicine.

[0043] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0044] According to an embodiment of the present invention, a device for identifying cerebral microbleeds is also provided, including an image acquisition unit for acquiring magnetic resonance images input from two channels, wherein the two channels are a SWI image input channel and a phase image input channel; a microbleed candidate region detection unit for performing target detection on the magnetic resonance image based on a preset target detection model to screen out candidate regions for microbleeds, and cropping the candidate regions into 3D patch images, wherein before cropping the candidate regions into 3D patch images, the magnetic resonance image containing the candidate regions is segmented based on BET to segment brain tissue regions and non-brain tissue regions; the candidate regions contained in the brain tissue regions are cropped to obtain a 3D patch image; and a classification unit for inputting the 3D patch image into a preset convolutional neural network model for context learning to determine the microbleed region and the false positive region.

[0045] As an optional implementation of this embodiment, inputting the 3D patch image into a preset convolutional neural network model for context learning includes: determining the 3D context information based on the average value of adjacent slices of the SWI and phase image: ,in, Representative The slice The value of the channel, is the channel number, Represents the SWI channel, stands for complementary phase channel; Indicates The SWI value of each slice, Indicates The phase value of each slice, is the index of the current slice, is the index of adjacent slices, ranging from arrive .

[0046] As an optional implementation method of this embodiment, performing target detection on a magnetic resonance image based on a preset target detection model includes: performing target detection on a magnetic resonance image based on a YOLO detection model with a preset structure, wherein the YOLO detection model includes N convolutional layers, and the kernel size of the convolutional layer is 3 × 3; m maximum pooling layers, the size of the pooling layer is 2 × 2, and the stride is 2; a 1 × 1 convolutional layer is inserted between every two consecutive 3 × 3 convolutional layers, called a bottleneck layer; each convolutional layer is followed by batch normalization and leaky activation functions, the last convolutional layer is followed by a linear activation function, and the number of features of the last convolutional layer is: a multimodal logistic regression layer is placed on the top layer to generate a tensor containing detected bounding box information and predicted categories; wherein, during the training process, the network is optimized by minimizing the loss function, and the loss function is determined based on the ground truth value and the predicted bounding box.

[0047] As an optional implementation of this embodiment, the preset convolutional neural network model includes five convolutional layers, the convolution kernel size is 3 × 3 × 3, and the output is: Y = X × W + b, where Y is the output of the convolution layer, X is the input of the convolution layer, W is the convolution kernel, and the bias term is b; each convolution layer is followed by batch normalization and a rectified linear unit activation function, and the normalized input is ,in is the normalized input, is the input mean, is the variance, is the error term, and the final output is , is the output after batch normalization, is the scaling parameter, is the offset parameter; the ReLU activation function is applied to the normalized output: , then down-sampling is performed using a pooling layer; the features extracted by the convolutional layer are flattened and passed to three fully connected layers, where the number of units in the last fully connected layer corresponds to the number of classification categories, namely, microbleeding area categories, and false positive area categories, where a random dropout layer is used after the first fully connected layer to regularize the network and prevent overfitting.

[0048] According to an embodiment of the present invention, the present invention also provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can implement the method described in any of the above embodiments when executing.

[0049] According to an embodiment of the present invention, the present invention further provides a readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to implement the method described in any of the above embodiments when executed.

[0050] Figure 5 A schematic block diagram of an example electronic device 300 that can be used to implement an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices.

[0051] like Figure 5 As shown, the electronic device 300 includes a computing unit 301, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. In the RAM 303, various programs and data required for the operation of the electronic device 300 can also be stored. The computing unit 301, the ROM 302, and the RAM 303 are connected to each other via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0052] A number of components in the electronic device 300 are connected to the I / O interface 305, including: an input unit 306, such as a keyboard, a mouse, etc.; an output unit 307, such as various types of displays, speakers, etc.; a storage unit 308, such as a disk, an optical disk, etc.; and a communication unit 309, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 309 allows the electronic device 300 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0053] The computing unit 301 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 301 performs the various methods and processes described above, such as the object matching method. For example, in some embodiments, the object matching method may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 300 via the ROM 302 and / or the communication unit 309. When the computer program is loaded into the RAM 303 and executed by the computing unit 301, one or more steps of the method described above may be performed.

[0054] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0055] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0056] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

Claims

1. A method for identifying cerebral microbleeds, characterized in that: include: Acquire a magnetic resonance image inputted from two channels, wherein the two channels are a SWI image input channel and a phase image input channel; Performing target detection on the magnetic resonance image based on a preset target detection model to screen out candidate areas of microbleeding, and cropping the candidate areas into 3D patch images, wherein before cropping the candidate areas into 3D patch images, segmenting the magnetic resonance image containing the candidate areas based on BET to segment brain tissue areas and non-brain tissue areas; cropping the candidate areas contained in the brain tissue area to obtain a 3D patch image; The 3D patch image is input into the preset convolutional neural network model for contextual learning to determine the microbleeding area and the false positive area.

2. The method for identifying cerebral microbleeds according to claim 1, characterized in that: Inputting the 3D patch image into the preset convolutional neural network model for context learning includes: Determine 3D context information based on the average of adjacent slices of SWI and phase images: ， in, Representative The slice The value of the channel, is the channel number, Represents the SWI channel, stands for complementary phase channel; Indicates The SWI value of each slice, Indicates The phase value of each slice, is the index of the current slice, is the index of adjacent slices, ranging from arrive .

3. The method for identifying cerebral microbleeds according to claim 1, characterized in that: Target detection on magnetic resonance images based on a preset target detection model includes: The YOLO detection model based on the preset structure performs target detection on the magnetic resonance image, wherein the YOLO detection model includes N convolutional layers with a kernel size of 3 × 3; m maximum pooling layers with a pooling size of 2 × 2 and a stride of 2; a 1 × 1 convolutional layer is inserted between every two consecutive 3 × 3 convolutional layers, called a bottleneck layer; each convolutional layer is followed by batch normalization and leaky activation functions, and the last convolutional layer is followed by a linear activation function. The number of features of the last convolutional layer: a multimodal logistic regression layer is placed on the top layer to generate a tensor containing the detected bounding box information and the predicted category; During the training process, the network is optimized by minimizing a loss function, which is determined based on the ground truth value and the predicted bounding box.

4. The method for identifying cerebral microbleeds according to claim 1, characterized in that: The preset convolutional neural network model includes five convolutional layers, the convolution kernel size is 3 × 3 × 3, and the output is: Y=X×W+b, where Y is the output of the convolution layer, X is the input of the convolution layer, W is the convolution kernel, and the bias term is b; Each convolutional layer is followed by batch normalization and rectified linear unit activation functions. The normalized input is ,in is the normalized input, is the input mean, is the variance, is the error term, and the final output is , is the output after batch normalization, is the scaling parameter, is the offset parameter; the ReLU activation function is applied to the normalized output: ,Then down-sampling is performed using a pooling layer; The features extracted by the convolutional layer are flattened and passed to three fully connected layers. The number of units in the last fully connected layer corresponds to the number of classification categories, namely, microbleeding area categories and false positive area categories. A random dropout layer is used after the first fully connected layer to regularize the network and prevent overfitting.

5. A device for identifying cerebral microbleeds, characterized in that: include: An image acquisition unit, used for acquiring magnetic resonance images inputted from two channels, wherein the two channels are respectively a SWI image input channel and a phase image input channel; A microbleed candidate region detection unit is used to perform target detection on the magnetic resonance image based on a preset target detection model to screen out the microbleed candidate region, and crop the candidate region into a 3D patch image, wherein before cropping the candidate region into the 3D patch image, the magnetic resonance image containing the candidate region is segmented based on BET to segment the brain tissue region and the non-brain tissue region; the candidate region contained in the brain tissue region is cropped to obtain the 3D patch image; The classification unit is used to input the 3D patch image into a preset convolutional neural network model for context learning to determine the microbleeding area and the false positive area.

6. The device for identifying cerebral microbleeds according to claim 5, characterized in that: Inputting the 3D patch image into the preset convolutional neural network model for context learning includes: Determine 3D context information based on the average of adjacent slices of SWI and phase images: ， in, Representative The slice The value of the channel, is the channel number, Represents the SWI channel, stands for complementary phase channel; Indicates The SWI value of each slice, Indicates The phase value of each slice, is the index of the current slice, is the index of adjacent slices, ranging from arrive .

7. The device for identifying cerebral microbleeds according to claim 5, characterized in that: Target detection on magnetic resonance images based on a preset target detection model includes: The YOLO detection model based on the preset structure performs target detection on the magnetic resonance image, wherein the YOLO detection model includes N convolutional layers with a kernel size of 3 × 3; m maximum pooling layers with a pooling size of 2 × 2 and a stride of 2; a 1 × 1 convolutional layer is inserted between every two consecutive 3 × 3 convolutional layers, called a bottleneck layer; each convolutional layer is followed by batch normalization and leaky activation functions, and the last convolutional layer is followed by a linear activation function. The number of features of the last convolutional layer: a multimodal logistic regression layer is placed on the top layer to generate a tensor containing the detected bounding box information and the predicted category; During the training process, the network is optimized by minimizing a loss function, which is determined based on the ground truth value and the predicted bounding box.

8. The device for identifying cerebral microbleeds according to claim 5, characterized in that: The preset convolutional neural network model includes five convolutional layers, the convolution kernel size is 3 × 3 × 3, and the output is: Y=X×W+b, where Y is the output of the convolution layer, X is the input of the convolution layer, W is the convolution kernel, and the bias term is b; Each convolutional layer is followed by batch normalization and rectified linear unit activation functions. The normalized input is ,in is the normalized input, is the input mean, is the variance, is the error term, and the final output is , is the output after batch normalization, is the scaling parameter, is the offset parameter; the ReLU activation function is applied to the normalized output: ,Then down-sampling is performed using a pooling layer; The features extracted by the convolutional layer are flattened and passed to three fully connected layers. The number of units in the last fully connected layer corresponds to the number of classification categories, namely, microbleeding area categories and false positive area categories. A random dropout layer is used after the first fully connected layer to regularize the network and prevent overfitting.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the method according to any one of claims 1 to 4.

10. An electronic device, characterized in that: include: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor executes the method described in any one of claims 1 to 4.

Images