A CT image cerebral hemorrhage region segmentation method based on deep learning

Abstract

The present application relates to the technical field of machine vision, in particular to a CT image cerebral hemorrhage region segmentation method based on deep learning. The present application obtains a cerebral hemorrhage CT image, labels a cerebral hemorrhage region in the cerebral hemorrhage CT image, establishes a cerebral hemorrhage CT image data set according to the cerebral hemorrhage CT image and the corresponding cerebral hemorrhage region, divides the cerebral hemorrhage CT image data set into a training set and a test set; data augmentation is performed on the training set, and the enhanced picture obtained after data augmentation is added to the training set; the training set is transmitted into a fast subtraction network model and trained to obtain an optimal weight; the test set is transmitted into the fast subtraction network model for testing, and the fast subtraction network model that passes the test is output for cerebral hemorrhage region segmentation. The accuracy of the present application on multiple evaluation indexes is high, and the data augmentation method helps to alleviate the problem of insufficient number of medical image data set images.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of machine vision, in particular to a CT image cerebral hemorrhage region segmentation method based on deep learning. BACKGROUND

[0002] After analyzing the CT image cerebral hemorrhage region segmentation methods at home and abroad, it is found that there are two difficulties in labeling the cerebral hemorrhage region in the CT image: (1) the discrimination of the cerebral hemorrhage region and the brain calcification region; (2) the edge of the cerebral hemorrhage region is relatively fuzzy, and a low-density ring appears on the edge. In order to achieve higher segmentation accuracy, the deep learning model structure used for segmenting the deep cerebral hemorrhage region is becoming more and more complex, and the complex model structure brings an increase in the number of parameters, which increases the difficulty of model convergence and prolongs the training time.

[0003] Resnet network and dense connection are used in the ablation experiment as a comparison. Resnet is widely used in medical image segmentation models as a feature extraction backbone network. Dense connection is a connection mode of a deep neural network, which means that each convolutional layer is connected to each other. SUMMARY

[0004] The application aims to solve the above technical problems and provides a CT image cerebral hemorrhage region segmentation method based on deep learning.

[0005] In order to achieve the above purpose, the technical scheme provided by the application is:

[0006] The CT image cerebral hemorrhage region segmentation method based on deep learning comprises the following steps:

[0007] Obtaining a cerebral hemorrhage CT image, marking a cerebral hemorrhage region in the cerebral hemorrhage CT image, establishing a cerebral hemorrhage CT image data set according to the cerebral hemorrhage CT image and the corresponding cerebral hemorrhage region, and dividing the cerebral hemorrhage CT image data set into a training set and a test set;

[0008] Data augmentation is performed on the training set, and the enhanced picture obtained after data augmentation is added to the training set;

[0009] The training set is transmitted into a fast subtraction network model and trained to obtain optimal weights;

[0010] The test set is transmitted into the fast subtraction network model based on the optimal weights for testing, and the fast subtraction network model that passes the test is output and used for cerebral hemorrhage region segmentation;

[0011] The fast subtraction network model comprises an encoding module, a multi-scale transition module and a decoding module, the encoding module, the multi-scale transition module and the decoding module each comprise an I-layer structure, an i-th layer of the encoding module outputs an i-th layer of the multi-scale transition module, and an i-th layer of the multi-scale transition module outputs an i-th layer of the decoding module, wherein i [1, I].

[0012] Further, the data augmentation on the training set comprises:

[0013] Calculate the image similarity of each brain hemorrhage CT image in the training set and the plurality of brain hemorrhage-free region CT images;

[0014] Sort all image similarities of the current brain hemorrhage CT image, and select the brain hemorrhage-free region CT image with the highest similarity to the current brain hemorrhage CT image as the mother;

[0015] Extract the brain hemorrhage region of the current brain hemorrhage CT image, replace the pixels of the corresponding region of the mother with the pixels of the brain hemorrhage region of the brain hemorrhage CT image to obtain an enhanced picture of the current brain hemorrhage CT image;

[0016] Obtain the enhanced picture of each brain hemorrhage CT image in the training set.

[0017] Further, the image similarity of the brain hemorrhage-free region CT image and the brain hemorrhage CT image in the training set is calculated using a mutual information algorithm, and the calculation formula of the mutual information algorithm is as follows:

[0018]

[0019] Wherein, I(X; Y) represents the mutual information algorithm, (X; Y) represents two random variables: the brain hemorrhage-free region CT image and the brain hemorrhage CT image, p(x; y) represents the joint distribution of the random variable (X; Y), and p(x), p(y) are the marginal distributions of the two random variables.

[0020] Further, the extraction of the brain hemorrhage region of the brain hemorrhage CT image comprises:

[0021] The size of the cropped region is calculated by a probability distribution algorithm, and the calculation formula is as follows:

[0022] Y i =D(Y j )~λ

[0023] X j =X i +Y i

[0024] Wherein, X i represents the mother, Y jrepresents a cerebral hemorrhage area in a cerebral hemorrhage CT image dataset, λ represents a probability distribution, Y i represents a generated cropped area, X j represents an enhanced picture.

[0025] According to the size of the cropped area, the cerebral hemorrhage area of the cerebral hemorrhage CT image is cropped to obtain a cropped cerebral hemorrhage area.

[0026] Further, the encoding module comprises a group of convolutional layers, and a training process of the encoding module is represented by a formula as follows:

[0027] A'(x) = (1-λ)A(x) + λx

[0028]

[0029] wherein A'(x) is an activation function of the second layer in the group of convolutional layers, λ is a non-linear hyperparameter of the activation function A'(x), A(x) represents an activation function of the first layer in the group of convolutional layers, x is an image input into the fast subtraction network model, e is a current training round number, and E is a set total training round number.

[0030] Before training to a predetermined round number, A'(x) = A(x), and after training to the predetermined round number, A'(x) = x.

[0031] Further, the multi-scale transition module comprises I-1 subtraction units, and the multi-scale transition module is represented by a formula as follows:

[0032]

[0033]

[0034] wherein DS i represents a feature map output by the i-th layer of the multi-scale transition module, represents a feature map output by the i-th layer of the encoding module, represents a feature map output by the subtraction module in the i-th layer of the multi-scale transition module, represents a feature map input into the i+1-th layer of the multi-scale transition module, represents a feature map output by the subtraction module in the i+1-th layer of the multi-scale transition module, represents a feature map After element-wise subtraction, the feature map is input into a convolutional layer again.

[0035] Further, the first layer to the I-1 layer of the decoding module comprises a convolution layer and an up-sampling layer, the I layer of the decoding module comprises a convolution layer, the up-sampling layer samples the feature map output by the multi-scale transition module to the same size as the previous layer and outputs to the convolution layer of the same layer, the convolution layer fuses the feature map output by the multi-scale transition module and the feature map output by the previous layer of the decoding module after up-sampling and decodes to input the next layer of the decoding module, and the first layer of the decoding module outputs a segmentation result map.

[0036] Further, the optimal weight of the fast subtraction network model is obtained according to the segmentation result map and the CT image of cerebral hemorrhage input into the fast subtraction network model.

[0037] Compared with the prior art, the present application has the following advantages:

[0038] (1) The new segmentation method for segmenting the cerebral hemorrhage region of the CT image provided by the present application constructs a fast subtraction network model, the fast subtraction network model comprises three modules, an encoding module, a multi-scale transition module and a decoding module. The encoding module is used for encoding the features of each level of the CT image and inputting the features into the multi-scale transition module, the multi-scale transition module outputs the features after processing to the decoding module, and the decoding module outputs a final segmentation result map after fusing the output of the multi-scale transition module. The average intersection over union (Miou) of the segmentation result of the fast subtraction network model on the cerebral hemorrhage CT image dataset is 0.553, the average Dice coefficient (MDice) is 0.6266, the accuracy (Acc) is 0.9971, the weighted Dice coefficient is 0.8316, the structural similarity coefficient (SS α ) is 0.7743, and the enhanced alignment coefficient is 0.8588. Compared with the most advanced medical image segmentation models MsNet, PraNet and U2Net, the fast subtraction network model of the present application is superior in accuracy in multiple evaluation indexes, and provides a new method and a theoretical basis for clinical application.

[0039] (2) The present application designs a new data enhancement method suitable for CT images of brain lesions. First, a cerebral hemorrhage CT image is selected, then the similarity between the cerebral hemorrhage CT image and a CT image without a cerebral hemorrhage region is calculated by using a mutual information algorithm, the most similar CT image without a cerebral hemorrhage region is selected as a parent, the cerebral hemorrhage region of the cerebral hemorrhage CT image is extracted, and the pixels of the corresponding region of the CT image without a cerebral hemorrhage region are replaced by the pixels of the cerebral hemorrhage region of the cerebral hemorrhage CT image, finally a new cerebral hemorrhage CT image and a label are obtained. This method helps to alleviate the problem of insufficient number of images in the medical image dataset. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1is a quick subtraction network model structure diagram of the application;

[0041] Figure 2 is a CT image of cerebral hemorrhage and a corresponding cerebral hemorrhage area diagram of the application;

[0042] Figure 3 is a coding module structure diagram of the application;

[0043] Figure 4 is a data enhancement flowchart of the application;

[0044] Figure 5 is a training loss value diagram of the application;

[0045] Figure 6 is a segmentation result comparison diagram, A is a CT image of cerebral hemorrhage, B is a cerebral hemorrhage area manually labeled by a doctor, C is a segmentation result diagram of the quick subtraction network model of the application, D is a segmentation result diagram of the prior art MsNet, E is a segmentation result diagram of the prior art PraNet, and F is a segmentation result diagram of the prior art U2Net;

[0046] Figure 7 is a segmentation effect evaluation diagram of the quick subtraction network model of the application;

[0047] Figure 8 is a training speed evaluation diagram of the quick subtraction network model of the application;

[0048] Figure 9 is an effect comparison diagram of the transition connection of the application and the dense connection segmentation of the prior art;

[0049] Figure 10 is a segmentation effect comparison diagram of different backbone models of the application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0051] As shown in Figures 1-3 , a CT image cerebral hemorrhage area segmentation method based on deep learning of the application includes the following steps:

[0052] Step S1: an image department doctor manually marks a cerebral hemorrhage area in a CT image of cerebral hemorrhage, establishes a CT image data set of cerebral hemorrhage, and divides the CT image data set of cerebral hemorrhage into a training set and a test set.

[0053] Step S2: using a SimiliarMix data enhancement method to perform data enhancement on the training set.

[0054] Step S2.1: asFigure 4 As shown, a marked image (i.e. a brain hemorrhage CT image) is first selected, the marked brain hemorrhage area is extracted, and the image similarity of each brain hemorrhage CT image in the training set and the brain hemorrhage-free area CT image is calculated using the mutual information algorithm. The calculation formula of mutual information is shown in formula (1):

[0055]

[0056] Where I(X;Y) represents the mutual information algorithm, (X;Y) represents two random variables: brain hemorrhage-free area CT image and brain hemorrhage CT image, p(x; y) represents the joint distribution of random variables (X;Y), and p(x), p(y) are the marginal distributions of the two random variables.

[0057] After the calculation is completed, all the similarity scores of the brain hemorrhage-free area CT images and the brain hemorrhage CT images are sorted, and the brain hemorrhage-free area CT image with the highest similarity to the brain hemorrhage CT image is selected as the mother.

[0058] Step S2.2: Process the brain hemorrhage CT image, extract the brain hemorrhage area pixels of the brain hemorrhage CT image according to the brain hemorrhage area marked by the doctor, calculate the size of the cropped area by the probability distribution algorithm, and the calculation formula is as follows:

[0059] Y i =D(Y j )~λ#(2)

[0060] X j =X i +Y i #(3)

[0061] Where X i represents the mother, Y j represents the brain hemorrhage area in the brain hemorrhage CT image dataset, λ represents the probability distribution, Y i represents the generated cropped area, and X j represents the enhanced picture.

[0062] After obtaining the size of the cropped area, the brain hemorrhage area of the brain hemorrhage CT image is cropped to obtain the cropped brain hemorrhage area.

[0063] Step S2.3: Replace the corresponding pixels in the brain hemorrhage-free area CT image (mother) with the pixels of the cropped brain hemorrhage area, and use the cropped brain hemorrhage area as the marked area of the brain hemorrhage-free area CT image. At this time, the brain hemorrhage-free area CT image becomes a marked image.

[0064] Step S3: Import the training set into the fast subtraction network model FsubNet and train it.

[0065] The architecture of the fast subtraction network model FsubNet of the embodiment is shown in Figure 1 which mainly includes three parts. The left part is an encoding module (Van encoder), the middle part is a multi-scale transition module, and the right part is a decoding module.

[0066] The structure of the encoding module is shown in Figure 3 In the embodiment, the pictures in the training set are sequentially passed through five groups of convolutional layers (corresponding to E 1 ~E 5 in the figure), and each group of convolutional layers is composed of a 4*4 and a 1*1 convolutional layer and a normalization layer. The processed feature maps are respectively input into the multi-scale transition module through the output of each group of convolutional layers.

[0067] The encoding module adopts a unique deep training strategy. At the beginning of training, a convolutional layer is divided into two similar convolutional layers, and the two convolutional layers are trained at the same time instead of one convolutional layer at the beginning of training, and the two convolutional layers are gradually merged into one convolutional layer as the training round number increases. In this way, the training time is reduced while maintaining the same feature extraction effect. The specific training process is described by formulas (4) and (5):

[0068] A'(x) = (1 - λ)A(x) + λx (4)

[0069]

[0070] wherein A'(x) is the activation function of the second layer in a group of convolutional layers, λ is a hyperparameter of the nonlinearity of the activation function A'(x), A(x) represents the activation function of the first layer in a group of convolutional layers, x is an image input into the fast subtraction network model, e is the current training round number, and E is the total training round number set. At the beginning of training (for example, within 100 training rounds), two convolutional layers A(x) and A'(x) are trained at the same time, A'(x) = A(x), and after a period of training (for example, after 100 training rounds), the model converges, at which time A'(x) = x, and at this time the two convolutional layers can be regarded as one convolutional layer.

[0071] The multi-scale transition module adopts a subtraction unit, and the basic subtraction unit is described by formula (6):

[0072] F = conv(|A-B|) (6)

[0073] wherein F represents the output of the subtraction unit, A and B represent the input feature maps, - represents element-wise subtraction, and conv represents that the output F is obtained by inputting the element-wise subtraction of the feature maps A and B into a convolutional layer again.

[0074] The specific process of the multi-scale transition module is described by formulas (7) and (8).

[0075]

[0076]

[0077] wherein DS i denotes the feature map output by the i-th layer of the multi-scale transition module, denotes the feature map output by the i-th layer of the encoding module, denotes the feature map output by the subtraction module in the i-th layer of the multi-scale transition module, denotes the feature map input into the i+1-th layer of the multi-scale transition module, denotes the feature map output by the subtraction module in the i+1-th layer of the multi-scale transition module, denotes the feature map After element-wise subtraction, the feature map is input into a convolutional layer again.

[0078] The decoding module includes four up-sampling layers and five 3*3 convolutional layers. Each of the four up-sampling layers is responsible for up-sampling the feature map output by the multi-scale transition module to the same size as the previous layer and outputting to the convolutional layer of the same layer. The five 3*3 convolutional layers are responsible for fusing the feature map output by the multi-scale transition module and the feature map output by the previous layer after up-sampling and inputting to the next layer after decoding. Finally, the first layer (i.e., the uppermost layer) outputs the segmentation result map after segmentation.

[0079] Step S4: After obtaining the optimal weight, the test picture is input into the model for brain hemorrhage area segmentation.

[0080] In order to further verify the accuracy of the CT image brain hemorrhage area segmentation method based on deep learning of the embodiment, an experimental example is provided below, CT image brain hemorrhage area segmentation experiment.

[0081] Data set collection: This project cooperates with the radiology department of the hospital, and retrospectively collects CT images of 128 patients with symptoms of cerebral hemorrhage from February 2021 to December 2022. Inclusion criteria: ① Patients who are clinically diagnosed as having symptoms of cerebral hemorrhage; ② CT image data is complete and of good quality. Exclusion criteria: CT image quality is poor and cannot be segmented. A total of 128 patients were included, of which 68 were male and 56 were female, with an age range of 24-94 years and an average age of 66.95 years. According to the collected data, a cerebral hemorrhage CT image data set was made, which included 236 images in the training set and 20 images in the test set, and all CT image brain hemorrhage areas were labeled by radiologists.

[0082] Firstly, the training set is augmented by using the SimiliarMix data augmentation method. After data augmentation, the training set is expanded to 472 images.

[0083] The processed image dataset is trained by the fast subtraction network FsubNet. The model in this paper is implemented based on the Pytorch deep learning framework. It is trained for 100 rounds on a server with i7 as Cpu and 3090 as gGpu. The Batchsize is 16, and the size of the input picture is 1578*893. SimiliarMix is used for data augmentation before training. The stochastic gradient descent function (SGD) is used as the optimizer. The coding module (Vanencoder) uses an independent learning rate, and the learning rate of the transition subtraction module and the decoder module is the same.

[0084] As shown in Figure 5 , after 100 rounds of training, the training loss value tends to 0 and remains stable and does not decrease, which indicates that the model has converged.

[0085] The Dice coefficient, average intersection over union (Miou), weighted Dice coefficient structural similarity coefficient (S α ) and enhanced alignment coefficient are used to test the model.

[0086] As shown in Figure 6 , the model effect comparison chart of the fast subtraction network model FsubNet and U2Net, PraNet and MsNet. It can be found that the weighted Dice coefficient, structural similarity coefficient, enhanced alignment coefficient, average Dice coefficient and average intersection over union of the fast subtraction network model FsubNet are higher than those of other comparison models, and the accuracy of the FsubNet model is the same as that of other models.

[0087] As shown in Figure 7 , this experiment uses U2Net, PraNet and MsNet as comparison models. Figure 7 It can be seen that the weighted Dice coefficient of the FsubNet model segmentation result is 0.8316, the structural similarity coefficient is 0.7443, the enhanced alignment coefficient is 0.8588, the average Dice coefficient is 0.6266, the average intersection over union is 0.553, and the accuracy is 0.9971. The weighted Dice coefficient, structural similarity coefficient, enhanced alignment coefficient, average Dice coefficient and average intersection over union of the FsubNet model are higher than those of other comparison models, and the accuracy of the FsubNet model is the same as that of other models.

[0088] As shown in Figure 8As shown in the figure, the training speed of U2Net is 75 images per minute, the training speed of PraNet is 90.225 images per minute, the training speed of MsNet is 107.14 images per minute, and the training speed of FsubNet is 109.09 images per minute, it can be seen that the training speed of FsubNet is higher than that of other comparative models.

[0089] In order to prove the effectiveness of the encoding module and the multi-scale transition module, a perfect ablation experiment needs to be carried out. In order to verify the effectiveness of the Vanencoder module, this paper uses the ResNet module to replace the Vanencoder module in the network, and after retraining the network, the average Dice coefficient (mDice), the average intersection over union (Miou), the weighted Dice coefficient , the structural similarity coefficient (SSIM) and the enhanced alignment coefficient (EAC) of the network segmentation image and the manual segmentation image are calculated α in the validation set. The segmentation effect of the network using the Vanencoder module or the ResNet module is compared. Secondly, in order to verify the effectiveness of the multi-scale transition module, the general convolution module is used to replace the multi-scale transition module, and after retraining the network, the average Dice coefficient (mDice), the average intersection over union (Miou), the weighted Dice coefficient , the structural similarity coefficient (SSIM) and the enhanced alignment coefficient (EAC) of the network segmentation image and the manual segmentation image are calculated α in the validation set. The segmentation effect of the network using the general convolution module or the multi-scale transition module is compared.

[0090] As Figure 9 shown in the figure, it can be seen that the weighted Dice coefficient of the dense connection is 0.6984, the structural similarity coefficient is 0.6690, the enhanced alignment coefficient is 0.7818, the average Dice coefficient is 0.5120, the average intersection over union is 0.4270, and the accuracy is 0.9979. From Figure 10 , it can be seen that the weighted Dice coefficient of the ResNet backbone network is 0.7469, the structural similarity coefficient is 0.7080, the enhanced alignment coefficient is 0.8100, the average Dice coefficient is 0.5725, the average intersection over union is 0.5025, and the accuracy is 0.9986.

[0091] The above-described embodiments only express one or several embodiments of the present application, and the description is more specific and detailed, but it cannot be understood as limiting the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1.A CT image cerebral hemorrhage region segmentation method based on deep learning, characterized in that, The deep learning-based CT image cerebral hemorrhage region segmentation method comprises the following steps: Obtain a cerebral hemorrhage CT image, mark a cerebral hemorrhage region in the cerebral hemorrhage CT image, establish a cerebral hemorrhage CT image dataset according to the cerebral hemorrhage CT image and the corresponding cerebral hemorrhage region, and divide the cerebral hemorrhage CT image dataset into a training set and a test set; Data augmentation is performed on the training set, and the enhanced picture obtained after data augmentation is added to the training set; The training set is input into a fast subtraction network model and trained to obtain optimal weights; The test set is input into the fast subtraction network model based on the optimal weights for testing, and the fast subtraction network model that passes the test is output for cerebral hemorrhage region segmentation. The fast subtraction network model comprises an encoding module, a multi-scale transition module and a decoding module, and the encoding module, the multi-scale transition module and the decoding module each comprise an I-layer structure, the i-th layer of the encoding module outputs an i-th layer of the multi-scale transition module, and the i-th layer of the multi-scale transition module outputs an i-th layer of the decoding module, wherein i ∈ [1, I]; The data augmentation on the training set comprises the following steps: Calculate the image similarity of each cerebral hemorrhage CT image in the training set and a plurality of cerebral hemorrhage-free region CT images using a mutual information algorithm; Sort all the image similarities of the current cerebral hemorrhage CT image, and select the cerebral hemorrhage-free region CT image with the highest similarity to the current cerebral hemorrhage CT image as the parent; Extract the cerebral hemorrhage region of the current cerebral hemorrhage CT image, replace the pixels of the corresponding region of the parent with the pixels of the cerebral hemorrhage region of the cerebral hemorrhage CT image to obtain an enhanced picture of the current cerebral hemorrhage CT image; Obtain the enhanced picture of each cerebral hemorrhage CT image in the training set; The multi-scale transition module comprises I-1 subtraction units, and the multi-scale transition module is expressed by the following formula: Calculate the image similarity of the cerebral hemorrhage-free region CT image and the cerebral hemorrhage CT image in the training set using a mutual information algorithm, and the calculation formula of the mutual information algorithm is as follows: The extraction of the cerebral hemorrhage region of the cerebral hemorrhage CT image comprises the following steps: ； ； wherein, denotes the feature map output by the i-th layer of the multi-scale transition module, denotes the feature map output by the i-th layer of the encoding module, denotes the feature map output by the subtraction module in the i-th layer of the multi-scale transition module, denotes the feature map input to the i+1-th layer of the multi-scale transition module, denotes the feature map output by the subtraction module in the i+1-th layer of the multi-scale transition module, denotes the feature map , is input to the convolution layer again after element-wise subtraction. 2.The CT image cerebral hemorrhage region segmentation method based on deep learning according to claim 1, characterized in that, Calculate the size of the cropped region by a probability distribution algorithm, and the calculation formula is as follows: ； wherein, denotes the mutual information algorithm, denotes two random variables: the CT image of the non-intracerebral hemorrhage region and the CT image of the intracerebral hemorrhage, denotes the joint distribution of the random variables , , are the marginal distributions of the two random variables. 3.The CT image cerebral hemorrhage region segmentation method based on deep learning according to claim 1, characterized in that, Crop the cerebral hemorrhage region of the cerebral hemorrhage CT image according to the size of the cropped region to obtain the cropped cerebral hemorrhage region. The encoding module comprises I groups of convolution layers, and the training process of the encoding module is expressed by the following formula: ； ； wherein, denotes the maternal parent, denotes a cerebral hemorrhage region in the cerebral hemorrhage CT image dataset, denotes a probability distribution, denotes a generated cropped region, denotes an enhanced picture; The first layer to the I-1 layer of the decoding module comprises an up-sampling layer and a convolution layer, the I layer of the decoding module comprises a convolution layer, the up-sampling layer up-samples the feature map output by the multi-scale transition module to the same size as the previous layer and outputs it to the convolution layer of the same layer, the convolution layer fuses the feature map output by the multi-scale transition module and the feature map output by the previous layer of the decoding module after up-sampling and inputs the fused feature map to the next layer of the decoding module, and the first layer of the decoding module outputs a segmentation result picture. 4.The CT image cerebral hemorrhage region segmentation method based on deep learning according to claim 1, characterized in that, Obtain the optimal weights of the fast subtraction network model according to the segmentation result picture and the cerebral hemorrhage CT image input into the fast subtraction network model. ； ； wherein, is an activation function of a 2nd layer in the set of convolutional layers, is an activation function is a hyperparameter of the nonlinearity, represents an activation function of a 1st layer in the set of convolutional layers, is an image input into the fast subtraction network model, is a current training round, is a total number of training rounds set; before training to a predetermined number of rounds, after training to a predetermined number of rounds, . 5.The CT image cerebral hemorrhage region segmentation method based on deep learning according to claim 1, characterized in that, ​ 6.The CT image cerebral hemorrhage region segmentation method based on deep learning according to claim 5, characterized in that, ​

Images