Automatic identification method and system of cerebral hemorrhage based on deep learning

Abstract

The present invention discloses a method and system for automatic identification of cerebral hemorrhage based on deep learning, the method comprising: performing data preprocessing on the original brain CT image to obtain preprocessed data; constructing an improved UNet model with dilated convolution, performing downsampling through an encoder, performing upsampling through a decoder, separating the encoder from the decoder through a single convolution layer, and connecting each downsampling step to the corresponding upsampling step; introducing a GAN network structure into the improved UNet model, inputting the preprocessed data into the improved UNet model to train the improved UNet model to obtain an optimal improved UNet model; performing data preprocessing on the brain CT image to be identified, inputting the data into the optimal improved UNet model, and outputting the identification result. The application of the technical solution of the present invention is helpful to solve the problems of low accuracy, efficiency and reliability of hemorrhagic stroke detection in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for automatically identifying cerebral hemorrhage based on deep learning. Background Art

[0002] Stroke is a serious cerebrovascular disease that can be divided into ischemic stroke and hemorrhagic stroke. Timely diagnosis of stroke is crucial for treatment, so computed tomography (CT) technology has been widely used in the medical field. In recent years, with the rapid development of artificial intelligence technology, deep learning methods have also made significant progress in stroke detection. Deep learning methods can automatically learn complex features from large-scale data and are more robust and accurate than traditional methods. Deep learning methods can process high-dimensional image data and capture subtle information that is difficult to detect with traditional methods. With the accumulation of data and the update of algorithms, they can continuously improve and optimize detection performance.

[0003] However, the performance of deep learning methods depends largely on the quality and quantity of training data. Insufficient or low-quality data may lead to insufficient model generalization ability, thus affecting detection accuracy. The training and reasoning process of deep learning models usually requires a lot of computing resources, which may be a burden for medical institutions. The decision-making process of deep learning models is often difficult to explain, which may affect doctors' confidence in the diagnosis results.

[0004] Since hemorrhagic stroke hematomas have significant variability in location, size, and shape, it poses a challenge to model training and recognition. In addition, hematomas may interact with surrounding brain tissue and blood vessels, increasing the complexity of feature extraction.

[0005] Therefore, there is an urgent need for a method that can improve the accuracy, efficiency, and reliability of hemorrhagic stroke detection. Summary of the invention

[0006] The purpose of the present invention is to provide a method and system for automatic identification of cerebral hemorrhage based on deep learning, so as to solve the problems of low accuracy, efficiency and reliability of hemorrhagic stroke detection in the prior art.

[0007] To achieve the above object, the present invention provides a method for automatic identification of cerebral hemorrhage based on deep learning, the method specifically comprising:

[0008] Acquiring an original brain CT image, and performing data preprocessing on the original brain CT image to obtain preprocessed data, wherein the data preprocessing includes data symmetry processing and data enhancement processing;

[0009] Constructing an improved UNet model with dilated convolution; wherein the improved UNet model includes an encoder and a decoder, downsampling is performed by the encoder, upsampling is performed by the decoder, the encoder is separated from the decoder by a single convolutional layer, and each downsampling step is connected to the corresponding upsampling step;

[0010] Introducing the GAN network structure into the improved UNet model, inputting the preprocessed data into the improved UNet model to train the improved UNet model, so as to obtain the optimal improved UNet model;

[0011] After data preprocessing, the brain CT image to be identified is input into the optimal improved UNet model to output the recognition result.

[0012] On the basis of the above technical solution, the present invention can also be improved as follows:

[0013] Optionally, the data symmetric processing includes:

[0014] Generate a flipped image for each original brain CT image;

[0015] A matching pixel pair between the flipped image and the original brain CT image is obtained, and a rigid transformation is used to align the flipped image with the original brain CT image based on the matching pixels.

[0016] Optionally, after aligning the flipped image with the original brain CT image based on the matching pixels using a rigid transformation, the method further comprises:

[0017] Each of the flipped images and the original brain CT image corresponding to each of the flipped images are spliced ​​to obtain a spliced ​​image.

[0018] Optionally, the data enhancement processing includes horizontal flipping processing, random cropping processing and principal component analysis processing.

[0019] Optionally, inputting the preprocessed data into an improved UNet model to train the improved UNet model to obtain an optimal improved UNet model includes:

[0020] Calculated by formula (1) ;

[0021] Formula (1);

[0022] In the formula, It represents the function F, the expansion factor Discrete convolution operation between and function k, p is the independent variable, and the result of the convolution is taken at p. For input data, is the expansion factor, The filter or operation rule used to process or transform F, is the independent variable of the result in the convolution operation, indicating the position of the output, is the input variable of the convolution, representing the input of F, is the input variable of the convolution, representing the input of k, is the value of function F at position s, is the value of function k at position t, is the discrete convolution operator;

[0023] When training the improved UNet model, the input preprocessed data is subjected to feature extraction and transformation by formula (1), and training is continuously iterated to obtain the optimal model parameters.

[0024] Optionally, the step of inputting the preprocessed data into an improved UNet model to train the improved UNet model to obtain an optimal improved UNet model further includes:

[0025] By improving the UNet model, the target category is predicted at the pixel level to obtain the segmentation result;

[0026] Based on the segmentation results, the performance of the improved UNet model is verified based on the segmentation results.

[0027] Optionally, the step of inputting the preprocessed data into an improved UNet model to train the improved UNet model to obtain an optimal improved UNet model further includes:

[0028] The real data and the segmentation result are received through the discriminant network, the prediction probability is output, and the performance of the improved UNet model is evaluated based on the prediction probability to obtain the optimal improved UNet model.

[0029] An automatic cerebral hemorrhage recognition system based on deep learning, comprising:

[0030] An acquisition module, used for acquiring original brain CT images;

[0031] A data processing module, used for performing data preprocessing on the original brain CT image to obtain preprocessed data, wherein the data preprocessing includes data symmetry processing and data enhancement processing;

[0032] A construction module for constructing an improved UNet model with dilated convolutions; wherein the improved UNet model includes an encoder and a decoder, downsampling is performed by the encoder, upsampling is performed by the decoder, the encoder is separated from the decoder by a single convolutional layer, and each downsampling step is connected to a corresponding upsampling step;

[0033] A training module, used for introducing the GAN network structure into the improved UNet model, inputting the preprocessed data into the improved UNet model to train the improved UNet model, so as to obtain an optimal improved UNet model;

[0034] The prediction module is used to preprocess the brain CT image to be identified, input it into the optimal improved UNet model, and output the recognition result.

[0035] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the method are implemented when the processor executes the computer program.

[0036] A non-transitory computer-readable storage medium stores a computer program, which implements the steps of the method when executed by a processor.

[0037] The embodiments of the present invention have the following advantages:

[0038] The present invention is based on a deep learning-based automatic identification method for cerebral hemorrhage. The present invention significantly improves the accuracy of hemorrhage area segmentation by performing data enhancement processing and symmetry processing on brain CT images. The model effectively enhances the accuracy of the location and boundary of hemorrhage lesions by using left-right flipping and symmetry information of the image. Through automated symmetry processing and data enhancement strategies, the workload and time cost of manual annotation are significantly reduced, thereby improving the efficiency of the segmentation process. By optimizing the processing flow and reducing manual intervention, the carbon emissions and environmental load in the medical image processing process are significantly reduced, meeting the requirements of sustainable development.

[0039] The present invention not only performs well on specific data sets, but also improves the generalization ability of the model on different CT brain image types and scales through adversarial training and diversified data enhancement technology, and is applicable to a wide range of clinical practice scenarios. Through precise hemorrhage area segmentation, it provides more accurate and detailed support for medical decision-making, helps doctors quickly identify and analyze the patient's lesions, thereby improving the quality and efficiency of medical services.

[0040] In summary, the present invention is not only significantly superior in terms of technological innovation, but also brings significant efficiency improvements and environmentally friendly characteristics in practical applications, which plays an important role in promoting the existing technology and future development of medical image analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0042] The structures, proportions, sizes, etc. illustrated in this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with the technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantial technical significance. Any structural modification, change in proportion or adjustment of size shall still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0043] Figure 1 It is a first flow chart of a method for automatic identification of cerebral hemorrhage based on deep learning in an embodiment of the present invention;

[0044] Figure 2 Schematic diagram of a second flow chart of the method for automatic identification of cerebral hemorrhage based on deep learning in an embodiment of the present invention;

[0045] Figure 3 Schematic diagram of symmetry characteristic processing of CT brain images;

[0046] Figure 4 Schematic diagram of the structure of GAN of the present invention;

[0047] Figure 5 It is a schematic diagram of the main components of the automatic identification system of cerebral hemorrhage based on deep learning in an embodiment of the present invention;

[0048] Figure 6 This is a schematic diagram of the physical structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0049] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are 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 are within the scope of protection of the present invention.

[0050] Figure 1 FIG. 1 is a schematic diagram of a first flow chart of a method for automatic identification of cerebral hemorrhage based on deep learning in an embodiment of the present invention. Figure 1As shown, an automatic identification method for cerebral hemorrhage based on deep learning provided by an embodiment of the present invention includes the following steps S101 to S104.

[0051] S101, collecting a certain number of original brain CT images, performing data preprocessing on the original brain CT images, and obtaining preprocessed data.

[0052] Since the skull content is usually remarkably symmetrical in axial CT images, traditional methods do not fully consider this special symmetry feature, but instead segment by processing the content of small neighborhood blocks (i.e., unilateral descriptors), which is inconsistent with the actual gold standard annotation. In fact, lesions rarely appear in symmetrical areas at the same time, so it is a natural method to compare the symmetrical areas of the image to find differences. When stroke lesions occur, the symmetry of CT brain images will be reduced.

[0053] In the embodiment of the present invention, two symmetrical blocks are extracted to describe a voxel, and they are simultaneously input into the network. Specifically, in the embodiment of the present invention, the original image is flipped left and right and spliced ​​together as the network input, such as Figure 3 As shown in Figure 2, for each block, the convolution is based on both the original region and its symmetric part. The difference between each pair of images is expressed through an encoder-decoder structure and iterative learning.

[0054] However, CT images are not completely axisymmetric. Therefore, before flipping, rotating, and splicing, image alignment is required to ensure that symmetric regions are considered simultaneously. In this embodiment, brain symmetry information is encoded into a UNet-based deep learning model by adding data symmetry processing and data enhancement processing steps. The specific steps for calculating the symmetry map are as follows:

[0055] Generate a flipped image for each original brain CT image;

[0056] Acquire matching pixel pairs between the flipped image and the original brain CT image, and use rigid transformation (translation and rotation) to align the flipped image with the original brain CT image based on the matching pixels;

[0057] Each of the flipped images and the original brain CT image corresponding to each of the flipped images are spliced ​​to obtain a spliced ​​image (2-channel image).

[0058] Due to the irregular size and shape of the lesion area, accurately segmenting the hemorrhage area from the CT image is a complex and time-consuming task. Lack of sufficient training data is one of the key reasons for the poor network training effect. In the embodiments of the present invention, a series of enhancement strategies are proposed and verified, including horizontal flipping, random cropping and principal component analysis (PCA), which can effectively capture the important features in the image.

[0059] Data imbalance is another important issue that affects the efficiency of network training. In medical images, the lesion area and the slices containing the lesions only account for a small part of the data. Therefore, the number of positive and negative samples is extremely unbalanced, which may introduce bias towards the majority when developing a classification model. In order to address this problem, the embodiment of the present invention divides the training data into two categories: hemorrhage and normal, and then performs enhancement operations such as horizontal flipping, random cropping and rotation on these two data sets. The purpose of this is to ensure that the amount of training data for the hemorrhage and normal categories is equal, thereby improving the training effect and generalization ability of the model.

[0060] S102, build an improved UNet model with dilated convolution.

[0061] Specifically, the improved UNet model includes an encoder and a decoder, downsampling is performed by the encoder, upsampling is performed by the decoder, the encoder is separated from the decoder by a single convolutional layer, and each downsampling step is connected to the corresponding upsampling step.

[0062] S103, introducing the GAN network structure into the improved UNet model, inputting the preprocessed data into the improved UNet model to train the improved UNet model, so as to obtain the optimal improved UNet model.

[0063] Specifically, dilated convolution is an improved standard convolution that expands the effective receptive field of the convolution kernel by inserting zeros (or holes) between each pixel of the convolution kernel. This enables the dilated convolution network (dilatedCNN) to improve the accuracy of segmentation while maintaining a limited number of trainable weights and convolution layers.

[0064] Calculated by formula (1) ;

[0065] Formula (1);

[0066] In the formula, It represents the function F, the expansion factor Discrete convolution operation between and function k, p is the independent variable, and the result of the convolution is taken at p. For input data, is the expansion factor, The filter or operation rule used to process or transform F, is the independent variable of the result in the convolution operation, indicating the position of the output, is the input variable of the convolution, representing the input of F, is the input variable of the convolution, representing the input of k, is the value of function F at position s, is the value of function k at position t, is the discrete convolution operator.

[0067] When training the improved UNet model, the input preprocessed data is subjected to feature extraction and transformation using formula (1), and the optimal model parameters are obtained through continuous iterative training.

[0068] For input data , through the expansion factor and filter or operation rules Convolution operations can fuse features at different levels. For example, during the downsampling process of the encoder, this convolution can help capture key features in brain CT images at different scales. For example, hemorrhage areas of different sizes may show different features at different scales, which can be effectively extracted through this convolution.

[0069] Input Data Through Convolution is performed to process and transform the original data. In the automatic identification of cerebral hemorrhage, the data of brain CT images may be relatively complex. Through this convolution operation, the original data can be converted into a form that is more conducive to model recognition and analysis.

[0070] For example, convolution operations can highlight features related to cerebral hemorrhage and suppress irrelevant background information, so that the model can more accurately focus on meaningful feature areas in subsequent processing and improve the accuracy of cerebral hemorrhage identification.

[0071] Formula (1) is a key component of the improved UNet model. During the model training process, the model parameters are continuously adjusted through this convolution operation to make it better fit the training data. When the model is trained, Formula (1) is used to extract and transform the input preprocessed data, thereby obtaining more accurate recognition results at the output. Finally, through continuous iterative training, the optimal model parameters are found to achieve accurate and automatic recognition of cerebral hemorrhage.

[0072] In order to speed up the training process, the batch normalization technology is introduced in the embodiment of the present invention, and the ReLU activation function (ReLU, Rectified Linear Unit) is applied after each convolution layer to enhance the nonlinear expression ability of the model.

[0073] Transfer learning improves learning by transferring knowledge from learned tasks to new tasks, especially when there is insufficient training data or to prevent overfitting. Since data annotation in the medical field is difficult and time-consuming, it is difficult to obtain enough training data to train deep learning models. Transfer learning has been widely used in medical imaging problems, improving model performance by leveraging models pre-trained on large-scale datasets (such as ImageNet) and then fine-tuning them on small-scale datasets.

[0074] In the embodiment of the present invention, a pre-trained UNet++ model is used for hemorrhage segmentation of CT brain images. UNet++ has an encoder-decoder structure similar to UNet, in which the encoder and decoder subnetworks are connected by a series of nested dense skip paths. The redesigned skip path is intended to reduce the semantic gap between the feature maps of the encoder and decoder subnetworks. In order to utilize the pre-trained weights on ImageNet, the pre-trained Inception / VGG is used as an encoder in the embodiment of the present invention, replacing the original encoder. When constructing the encoder, the last fully connected layer is removed, and the encoder and decoder are separated by a single convolutional layer. The decoder part is the same as the original UNet++. By fixing the encoder weights and fine-tuning them on CT brain images, an improved UNet model specifically for hemorrhage segmentation is finally obtained.

[0075] Adversarial training was first introduced by Generative Adversarial Networks (GANs) for generative modeling. GANs consist of two core modules: a generator and a discriminator. The task of the generator is to synthesize samples that are close to the distribution of real data, while the discriminator evaluates the probability of whether the sample comes from a real dataset or synthesized data by the generator. The generator aims to maximize the probability of the discriminator making mistakes, while the discriminator seeks to minimize its error rate. The core of this adversarial training is that the discriminator can be a complex neural network that can capture high-order inconsistencies in the generator samples, thereby generating highly realistic samples (such as images).

[0076] Figure 2 FIG. 1 is a specific flow chart of a method for automatic identification of cerebral hemorrhage based on deep learning in an embodiment of the present invention. Figure 2As shown in the figure, in this process, the adversarial process is integrated into the UNet segmentation model to achieve accurate segmentation of CT images. Specifically, a convolutional semantic segmentation network and an adversarial network are trained, the latter of which is used to distinguish whether the segmentation result comes from the real data or the segmentation network. The UNet segmentation network acts as a generator in the GAN to make pixel-level predictions for the target category, while the discriminant network receives the real data and the UNet segmentation results and outputs the predicted probability.

[0077] According to the schematic diagram of the GAN UNet generator (see Figure 4 The process shown in (a) generator in ), the depth of the downsampling process is 7, and each step contains a convolution layer, an activation layer and a batch normalization layer. Compared with the traditional UNet, the implementation of the present invention adopts patch training and LeakyReLU (Leaky Rectified LinearUnit, a variant of the rectified linear unit) activation layer to retain more features in the convolution layer. The depth of the upsampling process is consistent with the downsampling encoder. In particular, each downsampling step is connected to the corresponding upsampling step, rather than the jump connection in the traditional UNet. This design provides more training parameters for the discriminator and improves its judgment accuracy. The discriminator includes 4 steps, such as Figure 4 The last three steps consist of a convolutional layer, an activation layer, and a batch normalization layer, while the first step does not have a batch normalization layer to prevent overfitting.

[0078] S104, after data preprocessing, the brain CT image to be identified is input into the optimal improved UNet model to output the identification result.

[0079] According to the technical solution of the embodiment of the present invention, a certain number of original brain CT images are collected, and data preprocessing is performed on the original brain CT images to obtain preprocessed data, wherein the data preprocessing includes data symmetry processing and data enhancement processing; an improved UNet model with dilated convolution is constructed; wherein the improved UNet model includes an encoder and a decoder, downsampling is performed by the encoder, upsampling is performed by the decoder, the encoder is separated from the decoder by a single convolutional layer, and each downsampling step is connected to the corresponding upsampling step; the GAN network structure is introduced into the improved UNet model, the preprocessed data is input into the improved UNet model to train the improved UNet model to obtain the optimal improved UNet model; after data preprocessing of the brain CT image to be identified, it is input into the optimal improved UNet model, and the identification result is output. The problem of low accuracy, efficiency and reliability of hemorrhagic stroke detection in the prior art is solved.

[0080] Figure 5 Schematic diagram of the main components of the deep learning-based automatic identification system for cerebral hemorrhage in an embodiment of the present invention. Figure 5 As shown, the deep learning-based automatic identification system 1 for cerebral hemorrhage provided in an embodiment of the present invention includes an acquisition module 10, a data processing module 20, a construction module 30, a training module 40 and a prediction module 50.

[0081] The acquisition module 10 is used to acquire a certain number of original brain CT images.

[0082] The data processing module 20 is used to perform data preprocessing on the original brain CT image to obtain preprocessed data, wherein the data preprocessing includes data symmetry processing and data enhancement processing.

[0083] The data processing module 20 is also used to generate a flipped image for each original brain CT image; obtain matching pixel pairs between the flipped image and the original brain CT image, and use rigid transformation to align the flipped image with the original brain CT image based on the matching pixels; and splice each flipped image with the original brain CT image corresponding to each flipped image to obtain a spliced ​​image. The data enhancement process includes horizontal flipping, random cropping, and principal component analysis.

[0084] The construction module 30 is used to construct an improved UNet model with dilated convolution; wherein the improved UNet model includes an encoder and a decoder, downsampling is performed by the encoder, upsampling is performed by the decoder, the encoder is separated from the decoder by a single convolutional layer, and each downsampling step is connected to the corresponding upsampling step.

[0085] The training module 40 is used to introduce the GAN network structure into the improved UNet model, input the preprocessed data into the improved UNet model to train the improved UNet model, so as to obtain the optimal improved UNet model.

[0086] The prediction module 50 is used to perform data preprocessing on the brain CT image to be identified, input it into the optimal improved UNet model, and output the identification result.

[0087] Figure 6 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as Figure 6 As shown, the electronic device 60 includes: a processor 601 (processor), a memory 602 (memory) and a bus 603;

[0088] The processor 601 and the memory 602 communicate with each other via the bus 603;

[0089] The processor 601 is used to call the program instructions in the memory 602 to execute the method provided by the embodiment of the present invention.

[0090] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the method provided by the embodiment of the present invention.

[0091] A person skilled in the art can understand that all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various storage media that can store program codes.

[0092] The device embodiments described above are 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 scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0093] 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 embodiment or some parts of the embodiment.

[0094] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A method for automatic identification of cerebral hemorrhage based on deep learning, characterized in that: include: Acquiring an original brain CT image, and performing data preprocessing on the original brain CT image to obtain preprocessed data, wherein the data preprocessing includes data symmetry processing and data enhancement processing; Constructing an improved UNet model with dilated convolution; wherein the improved UNet model includes an encoder and a decoder, downsampling is performed by the encoder, upsampling is performed by the decoder, the encoder is separated from the decoder by a single convolutional layer, and each downsampling step is connected to the corresponding upsampling step; The GAN network structure is introduced into the improved UNet model, and the preprocessed data is input into the improved UNet model to train the improved UNet model to obtain an optimal improved UNet model, including: Calculated by formula (1) ; Formula (1): In the formula, It represents the function F, the expansion factor Discrete convolution operation between and function k, p is the independent variable, the result of the convolution is taken at p, function For input data, is the dilation factor, the filter or operation rule used by the function to process or transform the function F, is the independent variable of the result in the convolution operation, indicating the position of the output, is the input variable of the convolution, representing the input of the function F, is the input variable of the convolution, representing the input of function k, is the value of function F at position s, is the value of function k at position t, * is the discrete convolution operator; When training the improved UNet model, the input preprocessed data is subjected to feature extraction and transformation by formula (1), and training is continuously iterated to obtain the optimal model parameters; After data preprocessing, the brain CT image to be identified is input into the optimal improved UNet model to output the recognition result.

2. The method for automatic identification of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: The data symmetric processing includes: Generate a flipped image for each original brain CT image; A matching pixel pair between the flipped image and the original brain CT image is obtained, and a rigid transformation is used to align the flipped image with the original brain CT image based on the matching pixels.

3. The method for automatic identification of cerebral hemorrhage based on deep learning according to claim 2, characterized in that: After the rigid transformation is used to align the flipped image with the original brain CT image based on the matching pixels, the method further comprises: Each of the flipped images and the original brain CT image corresponding to each of the flipped images are spliced ​​to obtain a spliced ​​image.

4. The method for automatic identification of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: The data enhancement processing includes horizontal flipping processing, random cropping processing and principal component analysis processing.

5. The method for automatic identification of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: The step of inputting the preprocessed data into an improved UNet model to train the improved UNet model to obtain an optimal improved UNet model further includes: By improving the UNet model, the target category is predicted at the pixel level to obtain the segmentation result; Based on the segmentation results, the performance of the improved UNet model is verified based on the segmentation results.

6. The method for automatic identification of cerebral hemorrhage based on deep learning according to claim 5, characterized in that: The step of inputting the preprocessed data into an improved UNet model to train the improved UNet model to obtain an optimal improved UNet model further includes: The real data and the segmentation result are received through the discriminant network, the prediction probability is output, and the performance of the improved UNet model is evaluated based on the prediction probability to obtain the optimal improved UNet model.

7. A deep learning-based automatic identification system for cerebral hemorrhage, characterized in that: include: An acquisition module, used for acquiring original brain CT images; A data processing module, used for performing data preprocessing on the original brain CT image to obtain preprocessed data, wherein the data preprocessing includes data symmetry processing and data enhancement processing; A construction module for constructing an improved UNet model with dilated convolutions; wherein the improved UNet model includes an encoder and a decoder, downsampling is performed by the encoder, upsampling is performed by the decoder, the encoder is separated from the decoder by a single convolutional layer, and each downsampling step is connected to a corresponding upsampling step; A training module, used for introducing the GAN network structure into the improved UNet model, inputting the preprocessed data into the improved UNet model to train the improved UNet model, so as to obtain an optimal improved UNet model; The training module is also used to: Calculated by formula (1) ; Formula (1): In the formula, It represents the function F, the expansion factor Discrete convolution operation between and function k, p is the independent variable, the result of the convolution is taken at p, function For input data, is the dilation factor, the filter or operation rule used by the function to process or transform the function F, is the independent variable of the result in the convolution operation, indicating the position of the output, is the input variable of the convolution, representing the input of the function F, is the input variable of the convolution, representing the input of function k, is the value of function F at position s, is the value of function k at position t, * is the discrete convolution operator; When training the improved UNet model, the input preprocessed data is subjected to feature extraction and transformation by formula (1), and training is continuously iterated to obtain the optimal model parameters; The prediction module is used to preprocess the brain CT image to be identified, input it into the optimal improved UNet model, and output the recognition result.

8. 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 computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A non-transitory computer readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images