Intracranial hemorrhage identification method and system based on deep transfer learning

Abstract

The invention discloses an intracranial hemorrhage identification method and system based on deep transfer learning, and the method comprises the steps: obtaining a CT slice image, and extracting a binary image containing a skull part in the CT slice image, so as to complete the extraction of a brain tissue part image; a deep residual network is constructed, and discriminant feature automatic extraction of the input brain tissue part image is completed; and constructing a full connection layer, and performing intracranial hemorrhage judgment on the extracted discrimination features. By constructing the deep residual network, deep learning can be carried out on an input brain tissue part image without the problem of performance reduction, and meanwhile, in cooperation with the constructed full-connection layer, intracranial hemorrhage judgment can be carried out on the judgment features obtained after deep learning, so that the accuracy of the intracranial hemorrhage judgment is improved. In this way, the judgment accuracy is improved on a small-scale data set through deep learning, the requirement for a large amount of marked data is avoided, the time cost of data marking and model training is saved, and the overall judgment efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a method and system for identifying intracranial hemorrhage based on deep transfer learning. Background Art

[0002] In the field of medical imaging, intracranial hemorrhage (ICH) is a serious cerebrovascular disease, usually caused by bleeding in the brain tissue due to ruptured blood vessels. Intracranial hemorrhage is a type of stroke and one of the main causes of high mortality and disability rates in stroke. After intracranial hemorrhage, patients often need emergency intervention and treatment to reduce brain tissue damage and complications. Traditionally, the diagnosis of intracranial hemorrhage mainly relies on doctors' observation and analysis of CT images. This method has problems such as strong subjectivity, long time consumption, and easy errors. Therefore, the development of automated intracranial hemorrhage diagnosis technology is of great significance to improve the speed, accuracy and treatment effect of diagnosis. Under the background of existing technology, the technology of automatic detection of intracranial hemorrhage in the field of medical imaging has made certain progress. Traditional image analysis techniques such as fuzzy C-means, histogram analysis, neural networks, etc. are widely used in the detection of intracranial hemorrhage. However, these methods have some limitations when processing complex CT images, such as the accuracy and efficiency of feature extraction and classification need to be improved.

[0003] In recent years, deep learning technology has emerged in the field of medical imaging, providing new possibilities for automatic detection and diagnosis. By building multi-layer neural networks, deep learning methods can learn and extract advanced features from large amounts of data, thereby achieving more accurate image analysis and diagnosis. In the automatic detection of intracranial hemorrhage, researchers have begun to try to apply deep learning methods, such as convolutional neural networks (CNNs) and transfer learning, to improve accuracy and efficiency.

[0004] However, there are still some problems and challenges in the automatic detection of intracranial hemorrhage in the existing technology. First, traditional deep learning models may require a large amount of labeled data for training in medical image analysis, and the acquisition of this data is limited. Secondly, it is also a challenge to accurately classify different types of intracranial hemorrhage, because different types of hemorrhage have similar appearance characteristics in CT images, requiring more sophisticated feature extraction and classification methods. In addition, some deep learning models in the existing technology have problems of overfitting or underfitting, resulting in insufficient generalization ability in practical applications. More efficient algorithms and technical support are also needed for the processing and analysis of large-scale CT image data. Therefore, how to effectively combine deep learning technology and the characteristics of the medical imaging field to improve the accuracy and efficiency of automatic detection of intracranial hemorrhage is a problem that needs to be solved urgently. Summary of the invention

[0005] In view of this, the present invention proposes a method and system for identifying intracranial hemorrhage based on deep transfer learning, which can better extract and learn the characteristics of intracranial hemorrhage on a small-scale data set and accurately classify it. The present invention provides the following technical solutions:

[0006] A method for identifying intracranial hemorrhage based on deep transfer learning, the method comprising: obtaining a CT slice image and extracting a binary image containing a skull portion therein to complete the extraction of a brain tissue portion image; constructing a deep residual network to automatically extract discriminant features of the input brain tissue portion image; constructing a fully connected layer, and performing intracranial hemorrhage judgment on the extracted discriminant features.

[0007] Optionally, the method of acquiring a CT slice image and extracting a binary image containing a skull portion therein to complete the extraction of a brain tissue portion image includes: extracting a CT slice image in a DICOM format and converting it into a JPG format; extracting a binary image containing a skull portion in the CT image by using an Otsu threshold method; and performing an element-by-element multiplication operation on the inverted binary image and the CT slice image to extract a brain tissue portion image.

[0008] Optionally, the method of constructing a deep residual network to automatically extract discriminant features of an input brain tissue partial image includes: adding shortcut connections between network layers of the deep residual network; inputting a feature map into an initial residual block of the deep residual network, and learning a required mapping through shortcut connections; constructing a residual function through the feature map and the required mapping; performing element-by-element addition operations on the residual function and the feature map to obtain an output vector; inputting the output vector as an input vector into a subsequent residual block to output a higher-level output vector, and outputting the extracted discriminant features after multiple residual blocks are processed.

[0009] Optionally, the residual function includes three convolutional layers, namely a 1×1 filter layer, a 3×3 filter layer and a 1×1 filter layer, and each filter layer is followed by a batch normalization and a rectified linear unit as an activation function.

[0010] Optionally, the method of constructing a fully connected layer and performing intracranial hemorrhage judgment on the extracted discriminant features includes: performing a convolution operation on the input discriminant features, and performing a maximum pooling operation on the convolution result; passing the feature map output by the maximum pooling operation to multiple stacked residual blocks to reduce the spatial size of the feature map; performing a global average pooling operation on the feature map after the size reduction, and outputting a one-dimensional feature map; inputting the one-dimensional feature map into the fully connected layer for classification; passing the classification output by the fully connected layer to an output layer with a sigmoid activation function, if the output sigmoid neuron is greater than a set threshold, the corresponding discriminant feature is an intracranial hemorrhage judgment, otherwise it is a non-intracranial hemorrhage judgment.

[0011] Optionally, the method for constructing a fully connected layer and performing intracranial hemorrhage judgment on the extracted discriminant features further includes: using a binary cross entropy loss function to measure the difference between the judgment result of the fully connected layer and the true label, and the loss function is specifically:

[0012]

[0013] Among them, y∈{0,1} represents the true label of the judgment category, Represents the predicted probability of the judgment category being intracranial hemorrhage. Adaptive moment estimation is used as the optimizer to improve the judgment convergence speed and training stability of the fully connected layer.

[0014] The present invention further discloses an intracranial hemorrhage recognition system based on deep transfer learning, comprising: an image extraction module, used to obtain CT slice images and extract binary images containing skull parts therein, so as to complete the extraction of brain tissue partial images; a feature extraction module, used to construct a deep residual network, and automatically extract discriminant features of the input brain tissue partial images; a feature judgment module, used to construct a fully connected layer, and perform intracranial hemorrhage judgment on the extracted discriminant features.

[0015] The present invention further discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned intracranial hemorrhage identification method based on deep transfer learning.

[0016] The present invention further discloses an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned intracranial hemorrhage identification method based on deep transfer learning when executing the program.

[0017] The present invention further discloses a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned intracranial hemorrhage identification method based on deep transfer learning.

[0018] According to the technical solution of the present invention, by constructing a deep residual network, deep learning can be performed on the input partial brain tissue image without the problem of performance degradation. At the same time, in conjunction with the constructed fully connected layer, intracranial hemorrhage judgment can be made based on the discriminant features obtained after deep learning, thereby improving the accuracy of judgment through deep learning on a small-scale data set, avoiding the need for a large amount of labeled data, saving the time cost of data annotation and training models, and improving the overall judgment efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] For the purpose of illustration and not limitation, the present invention is now described in conjunction with the embodiments of the present invention and the accompanying drawings, in which:

[0020] Figure 1 is a schematic flow chart of a method for identifying intracranial hemorrhage in an embodiment of the present invention;

[0021] Figure 2 Schematic diagram of the implementation architecture of intracranial hemorrhage recognition in an embodiment of the present invention;

[0022] Figure 3 is a schematic diagram of the architecture of a residual block in an embodiment of the present invention;

[0023] Figure 4 is a schematic diagram of the architecture of a fully connected layer in an embodiment of the present invention;

[0024] Figure 5 is a schematic structural diagram of an intracranial hemorrhage identification system in an embodiment of the present invention;

[0025] Figure 6 is a schematic diagram comparing an ICH sample and a non-ICH sample in an embodiment of the present invention;

[0026] Figure 7 is a schematic diagram of a feature map output by a residual block in an embodiment of the present invention;

[0027] Figure 8 It is a schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the implementation mode of the present application will be clearly and completely described below in conjunction with the drawings in the implementation mode of the present application. Obviously, the described implementation mode is only a part of the implementation mode of the present application, not all the implementation modes. Based on the implementation mode in the present application, all other implementation modes obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first", "second", etc. in the specification of the present application and the above-mentioned drawings 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 application 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.

[0030] In the present application, the terms "upper", "lower", "left", "right", "front", "back", "top", "bottom", "inner", "outer", "middle", "vertical", "horizontal", "lateral", "longitudinal" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings. These terms are mainly used to better describe the present application and its implementation methods, and are not used to limit the indicated devices, elements or components to have a specific orientation, or to be constructed and operated in a specific orientation.

[0031] In addition, some of the above terms may be used to express other meanings in addition to indicating orientation or positional relationship. For example, the term "on" may also be used to express a certain dependency or connection relationship in some cases. For those of ordinary skill in the art, the specific meanings of these terms in this application can be understood according to the specific circumstances. In addition, the term "plurality" should mean two or more than two.

[0032] It should be noted that, in the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other. The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0033] The diagnosis of intracranial hemorrhage depends on the analysis of the patient's head CT image. After obtaining the patient's CT image, data processing and analysis are performed to obtain a judgment on the classification of the current CT image, that is, whether the patient has symptoms of intracranial hemorrhage.

[0034] refer to Figure 1 and Figure 2 This embodiment discloses a method for identifying intracranial hemorrhage based on deep transfer learning, the method comprising:

[0035] S100: Acquire a CT slice image and extract a binary image containing a skull portion therein to complete the extraction of a brain tissue portion image.

[0036] A binary image refers to an image in which the grayscale value of each pixel has only two possible values, namely 0 or 1, which are usually represented as black and white. For example, in the present embodiment, in a binary image, 0 represents black and 1 represents white. The binary image can be used to more conveniently perform edge detection, shape analysis, and preliminary image preprocessing.

[0037] Specifically, when extracting the brain tissue part image, the CT slice image in DICOM format is extracted and converted into JPG format. For example, the resolution of the CT slice image in DICOM format is 512×512. After obtaining the slice image, the binary image containing the skull part in the CT image is extracted by the Otsu threshold method, and the inverted binary image and the CT image are multiplied element by element to extract the brain tissue part. Among them, by extracting the CT image containing the skull part, the brain tissue part inside the skull can be clearly converted from the CT image to a binary image, and all brain tissue images can be quickly and effectively obtained in this way. The threshold method is to set a threshold, compare the grayscale of each pixel in the CT image with the threshold, and for example, when the grayscale of the pixel is greater than the threshold, the pixel is set to 1 (white), otherwise it is set to 0 (black), and the CT image is converted into a binary image.

[0038] Since the grayscale of the skull and brain tissue images is quite different, and the skull pixel is 1 (white) and the brain tissue pixel is 0 (black) during the initial recognition, the image of the brain tissue needs to be extracted through element-by-element multiplication. Therefore, the pixels of the brain tissue cannot be set to 0, so the binary image is inverted so that the pixels of the brain tissue are 1. At the same time, each pixel of the CT image is marked with a corresponding value through its own grayscale. For example, each pixel is assigned a value between 0 and 255 to retain the image of the brain tissue during the element-by-element multiplication process, and the pixels with a product of 0 are removed. That is, the inverted binary image and the CT image are operated through element-by-element multiplication, and the non-0 pixel points calculated are the image of the brain tissue.

[0039] After extracting the image of the brain tissue part, it is necessary to extract the classification features in the image, specifically:

[0040] S200: Build a deep residual network (ResNet) to automatically extract discriminant features of the input brain tissue images.

[0041] Residual learning is introduced into the deep residual network to solve the problems of small gradients and gradient explosion in the training process of deep neural networks, and to improve the trainability of deep networks. Specifically, shortcut connections are added between network layers to allow information to skip several layers and propagate directly in the network, thereby more effectively retaining key information and avoiding the gradient vanishing problem. In the residual block, the network learns the residual mapping instead of directly learning the mapping between input and output. This makes it easier to optimize the model, especially when processing complex CT images, to better extract the features of intracranial hemorrhage. The deep residual network contains multiple residual blocks, each of which contains an identity mapping and a residual mapping, so that the learning of the deep residual network is to learn the residual mapping instead of directly learning the mapping between input and output, further avoiding the problem of performance degradation of the constructed deep residual network.

[0042] refer to Figure 3 Specifically, the acquired brain tissue image is input into the stacked layers of the deep residual network for training to fit the residual mapping. For example, assuming that the input is x and the mapping to be learned is H(x), then the residual function F(x) is defined as:

[0043] F(x)=H(x)-x,

[0044] That is, the output minus the input. At this time, the residual function F(x) is learned instead of H(x) itself, where: H(x) = F(x)-x is learned by adding shortcut connections in the feedforward neural network. In this embodiment, the residual function F(x) includes three convolutional layers, specifically a 3×3 filter layer and a 1×1 filter layer, and each of the filter layers is followed by a batch normalization and a rectified linear unit as an activation function (ReLU). Therefore, the architecture of the residual block is defined as:

[0045] y=F(x,{W i})+x,

[0046] Among them, x and y are the input and output vectors of the layer under consideration, and the function F(x,{W i}) represents the residual function mapping to be learned, which has two layers:

[0047] F(x)=W2σ(W1 x),

[0048] Where σ represents the ReLU activation function.

[0049] Thus, the residual block F(x)+x is completed by shortcut connection and element-by-element addition, and then the sum is passed as the new input to the next ReLU activation function. The dimensions of F(x) and x must be equal. If the dimensions are not equal when changing the input and output channels, a linear projection W is required through a shortcut connection. s To match the dimensions, specifically: y = F(x, {W i})+W s x. Exemplary, reference Figure 3 and Figure 4 , the shortcut connections of residual block 1, residual block 2, residual block 3 and residual block 4 perform the same mapping, and Figure 4 In , a residual block is set to increase the dimension when crossing two feature maps of different sizes, and a mapping shortcut connection with stride 2 is used to match the dimensions.

[0050] After extracting the discriminant features of the brain tissue image, it is necessary to determine the result of intracranial hemorrhage (ICH) based on the discriminant features, as follows:

[0051] refer to Figure 4 , S300: Construct a fully connected layer, and judge intracranial hemorrhage for the extracted discriminant features. The extracted discriminant features are used as input and input into the input layer of the fully connected layer. The figure exemplifies the structure of a fully connected layer, which includes 64 convolution kernels of 7×7 size and stride 2. After the discriminant features (224×224×3) are convolved with the above convolution kernels, 112×112×64 is output, and it is input into a maximum pooling operation of 3×3 with stride 2, and 56×56×64 is output. It is input into the residual block 1 in the figure, and 56×56×64 is output. The output of residual block 1 is input into residual block 2, and 28×28×512 is output. And so on, until all residual blocks are processed. Exemplarily, four stacked residual blocks are set in this embodiment, namely residual block 1, residual block 2, residual block 3 and residual block 4. By convolving the input with stacked residual blocks, the computational complexity is reduced and redundant information in the input is removed, thereby enhancing deep learning capabilities and integrating multi-level features to effectively avoid overfitting, thereby improving the generalization ability of training. Figure 7The figure shows the feature maps output by the discriminant features after they are convolved by four residual blocks, a 7×7 convolution layer, and a maximum pooling layer. Among them, a is the output of the 7×7 convolution layer, b is the output of the maximum pooling layer, c is the output of residual block 1, d is the output of residual block 2, e is the output of residual block 3, and f is the output of residual block 4. When spanning two feature maps of different sizes, a mapping shortcut connection is required to match the dimensions. For example, as shown in the figure, residual block 1 to residual block 2, residual block 2 to residual block 3, and residual block 3 to residual block 4 all span different sizes. Therefore, the discriminant features (224×224×3) initially input are output as 7×7×2048 after being processed by the residual block, thereby reducing the spatial size of the feature map. The reduced-size feature map output by the last residual block is subjected to a global average pooling operation to reduce the feature map to 1×1×2048. These 2048 feature maps are input into a fully connected layer with 64 neurons for classification, and then the classification output by the fully connected layer is passed to the output layer with a sigmod activation function. If the output sigmod neuron is greater than the set threshold, the corresponding discriminant feature is an intracranial hemorrhage judgment, otherwise it is a non-intracranial hemorrhage judgment. In order to improve the accuracy of the output results of the fully connected layer, the binary cross entropy loss function is used to measure the difference between the judgment results of the fully connected layer and the true label. The specific loss function is:

[0052]

[0053] Among them, y∈{0,1} represents the true label of the judgment category, Represents the predicted probability of the judgment category being intracranial hemorrhage. Adaptive moment estimation is used as the optimizer to improve the judgment convergence speed and training stability of the fully connected layer.

[0054] In summary, by propagating the brain CT image forward through each layer, a series of feature maps are obtained, discriminant features are extracted at different levels, and these features are used to train the neurons in the fully connected layer, and finally ICH is accurately identified. The discriminant model constructed by the intracranial hemorrhage judgment method proposed in this embodiment can accurately distinguish normal images from ICH lesions, and achieves a higher classification accuracy rate (99.6%). The judgment of ICH by the fully connected layer of this embodiment is as follows Figure 6 As shown, a and b are ICH samples, and c is a non-ICH sample.

[0055] refer to Figure 5 This embodiment further discloses an intracranial hemorrhage recognition system based on deep transfer learning, including:

[0056] The image extraction module 21 is used to obtain the CT slice image and extract the binary image containing the skull part therein to complete the extraction of the brain tissue part image; specifically: extract the CT slice image in DICOM format and convert it into JPG format; extract the binary image containing the skull part in the CT image by Otsu threshold method; perform element-by-element multiplication operation on the inverted binary image and the CT slice image to extract the brain tissue part image.

[0057] The feature extraction module 22 is used to construct a deep residual network to automatically extract discriminant features of the input brain tissue partial image; specifically: adding shortcut connections between the network layers of the deep residual network; inputting feature maps into the initial residual block of the deep residual network, and learning the required mapping through shortcut connections; constructing a residual function through the feature map and the required mapping; performing element-by-element addition operations on the residual function and the feature map to obtain an output vector; the residual function includes three convolutional layers, which are a 1×1 filter layer, a 3×3 filter layer and a 1×1 filter layer, respectively, and each of the filter layers is followed by a batch normalization and a rectified linear unit as an activation function; the output vector is input as an input vector into a subsequent residual block to output a higher-level output vector, and the extracted discriminant features are output after the processing of multiple residual blocks is completed.

[0058] The feature judgment module 23 is used to construct a fully connected layer and perform intracranial hemorrhage judgment on the extracted discriminant features; specifically: perform a convolution operation on the input discriminant features, and perform a maximum pooling operation on the convolution result; pass the feature map output by the maximum pooling operation to multiple stacked residual blocks to reduce the spatial size of the feature map; perform a global average pooling operation on the feature map after the size reduction, and output a one-dimensional feature map; input the one-dimensional feature map into the fully connected layer for classification; pass the classification output by the fully connected layer to the output layer with a sigmoid activation function, if the output sigmoid neuron is greater than the set threshold, the corresponding discriminant feature is an intracranial hemorrhage judgment, otherwise it is a non-intracranial hemorrhage judgment; use a binary cross entropy loss function to measure the difference between the judgment result of the fully connected layer and the true label, and the loss function is specifically:

[0059]

[0060] Among them, y∈{0,1} represents the true label of the judgment category, Represents the predicted probability of the judgment category being intracranial hemorrhage. Adaptive moment estimation is used as the optimizer to improve the judgment convergence speed and training stability of the fully connected layer.

[0061] Figure 8 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as Figure 8As shown, the electronic device 50 includes: a processor 501 (processor), a memory 502 (memory) and a bus 503;

[0062] The processor 501 and the memory 502 communicate with each other via the bus 503 ; the processor 501 is used to call program instructions in the memory 502 to execute the methods provided by the above-mentioned method implementation methods.

[0063] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the methods provided by the above-mentioned method embodiments.

[0064] A person skilled in the art can understand that all or part of the steps for implementing the above-mentioned method implementation method can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium, which, when executed, executes the steps of the above-mentioned method implementation method; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various storage media that can store program codes.

[0065] The device implementation described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present implementation scheme. Those of ordinary skill in the art may understand and implement it without creative effort.

[0066] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each implementation method or some parts of the implementation method.

[0067] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying intracranial hemorrhage based on deep transfer learning, characterized in that: Methods include: Acquire the CT slice image and extract the binary image containing the skull part, so as to complete the extraction of the brain tissue part image; Construct a deep residual network to automatically extract discriminant features from the input brain tissue images; A fully connected layer is constructed, and the extracted discriminant features are used to determine intracranial hemorrhage.

2. The intracranial hemorrhage recognition method based on deep transfer learning according to claim 1, characterized in that: The method of acquiring a CT slice image and extracting a binary image containing a skull portion therein to complete the extraction of a brain tissue portion image comprises: Extract CT slice images in DICOM format and convert them into JPG format; The binary image containing the skull part in the CT image is extracted by Otsu threshold method; The inverted binary image and the CT slice image are multiplied element by element to extract the brain tissue image.

3. The intracranial hemorrhage recognition method based on deep transfer learning according to claim 1, characterized in that: The method of constructing a deep residual network to automatically extract discriminant features of an input brain tissue partial image comprises: Adding shortcut connections between layers of deep residual networks; Input the feature map to the initial residual block of the deep residual network and learn the required mapping through shortcut connections; Construct a residual function through the feature map and the desired map; Perform element-wise addition of the residual function and the feature map to obtain an output vector; The output vector is input as an input vector to the subsequent residual block to output a higher-level output vector, and the extracted discriminative features are output after multiple residual blocks are processed.

4. The intracranial hemorrhage recognition method based on deep transfer learning as claimed in claim 3, characterized in that: The residual function includes three convolutional layers, which are a 1×1 filter layer, a 3×3 filter layer and a 1×1 filter layer, respectively. Each filter layer is followed by a batch normalization and a rectified linear unit as an activation function.

5. The intracranial hemorrhage recognition method based on deep transfer learning according to claim 1, characterized in that: The method of constructing a fully connected layer and judging intracranial hemorrhage based on the extracted discriminant features includes: Perform a convolution operation on the input discriminant features and perform a maximum pooling operation on the convolution result; Pass the feature map output by the maximum pooling operation to multiple stacked residual blocks to reduce the spatial size of the feature map; Perform global average pooling on the reduced-size feature map and output a one-dimensional feature map; Input the one-dimensional feature map into the fully connected layer for classification; The classification output by the fully connected layer is passed to the output layer with the sigmod activation function. If the output sigmod neuron is greater than the set threshold, the corresponding discriminant feature is intracranial hemorrhage judgment, otherwise it is non-intracranial hemorrhage judgment.

6. The intracranial hemorrhage recognition method based on deep transfer learning according to claim 5, characterized in that: The method of constructing a fully connected layer and performing intracranial hemorrhage judgment on the extracted discriminant features also includes: The binary cross entropy loss function is used to measure the difference between the judgment result of the fully connected layer and the true label. The specific loss function is: Among them, y∈{0,1} represents the true label of the judgment category, It indicates the predicted probability of intracranial hemorrhage; Adaptive moment estimation is used as the optimizer to improve the judgment convergence speed and training stability of the fully connected layer.

7. An intracranial hemorrhage recognition system based on deep transfer learning, characterized in that: include: An image extraction module is used to obtain the CT slice image and extract the binary image containing the skull part, so as to complete the extraction of the brain tissue part image; The feature extraction module is used to construct a deep residual network to automatically extract discriminant features from the input brain tissue images; The feature judgment module is used to construct a fully connected layer and perform intracranial hemorrhage judgment on the extracted discriminant features.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, 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