An automatic segmentation method for intracranial hemorrhage area based on multi-layer CT images
By calculating the correlation of features between upper and lower planes on CT images and combining cross-entropy and Dice loss function, the accuracy and consistency problems of intracranial hemorrhage region segmentation on CT images are solved, and efficient hemorrhage region detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU ZHUOXI INST OF BRAIN & INTELLIGENCE
- Filing Date
- 2022-08-31
- Publication Date
- 2026-07-03
AI Technical Summary
Existing automatic segmentation methods for CT images are difficult to fully utilize information from upper and lower layers to segment intracranial hemorrhage areas, and the segmentation results of adjacent layers differ greatly, resulting in insufficient accuracy and consistency.
An automatic segmentation method for intracranial hemorrhage regions based on multi-slice CT images is constructed. The correlation between features in the upper and lower layers is calculated through a feature extraction network and a feature fusion network. The method is trained using a hybrid loss function of cross-entropy and Dice loss to generate a probability map of the hemorrhage region.
With a slight increase in computational load, the accuracy of bleeding region segmentation and consistency between adjacent layers are improved, thereby enhancing the detection precision and reliability of bleeding region statistics.
Smart Images
Figure CN116205930B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and in particular to an automatic segmentation method for intracranial hemorrhage regions based on multi-slice CT images. Background Technology
[0002] Intracranial hemorrhage is a common and critical clinical condition, encompassing both traumatic and non-traumatic bleeding within the skull. Depending on the location, it can be classified as epidural hematoma, subdural hematoma, subarachnoid hemorrhage, parenchymal hemorrhage, and intraventricular hemorrhage. The specific treatment plan depends on the size and location of the hemorrhage; therefore, doctors often need to use medical imaging techniques to examine the patient's intracranial condition before determining the treatment plan.
[0003] Currently, magnetic resonance imaging (MRI) and computed tomography (CT) are the two most commonly used medical imaging technologies. CT imaging technology, in particular, has gained wider application in the detection and diagnosis of intracranial hemorrhage areas due to its high speed, low cost, and ability to obtain good contrast images for precise research.
[0004] Currently, manual labeling and segmentation remains the gold standard for extracting hemorrhage areas and calculating hemorrhage volume from CT images. However, this method has significant drawbacks. First, it heavily relies on the physician's experience, making it time-consuming, labor-intensive, and inefficient for the labeler. Second, regarding accuracy, since the hematoma boundaries are not always clearly defined in most patients, there are significant individual and inter-individual segmentation errors. These shortcomings can easily lead to reduced detection accuracy. Therefore, automatic segmentation methods based on computer algorithms are increasingly being used to assist in clinical diagnosis and treatment decisions due to their advantages of eliminating subjective errors and saving time and effort. Existing automatic CT image segmentation methods can be divided into two types. One type is the layer-by-layer analysis method based on 2D algorithms. This type of method inputs multiple layers of images from the CT sequence into the algorithm model one by one, analyzes each image individually, and finally integrates the results. These methods are characterized by low memory usage and simple pre- and post-processing. However, because they do not consider the correlation between different layers in the CT sequence, their accuracy is relatively low, and the analysis results between adjacent layers differ significantly, leading to large errors when analyzing the number and volume of hemorrhage areas. Another type of method is the holistic sequence analysis method based on 3D algorithms. This method inputs the complete CT sequence into the model and directly outputs the segmentation results of the hemorrhage areas for the entire sequence. This method fully integrates information from different layers in the sequence during analysis, resulting in stronger consistency between results from adjacent layers and higher accuracy. However, due to the need for extensive 3D computation, this method places high demands on computer memory and other performance aspects. Furthermore, because the slice thickness and number of slices in CT sequences obtained from different devices or scanning modes vary, it is often difficult for a single algorithm to achieve compatibility.
[0005] Patent application CN201610595691.3 discloses a method and system for segmenting hemorrhage regions in brain CT images based on semi-supervised learning. Specifically, the method includes a semi-supervised model training stage and a hemorrhage region segmentation stage based on the semi-supervised model. The semi-supervised model training stage is used to train the semi-supervised model. The hemorrhage region segmentation stage involves converting the format of the 2D CT image sequence to be segmented into intracranial hemorrhage regions, reconstructing the 2D CT images into 3D space, then using a supervoxel algorithm to divide the 3D image into supervoxels of similar size, extracting features from each supervoxel as a sample, and finally, based on the features, using the trained semi-supervised model to divide the supervoxels into foreground and background parts. This invention is based on a 3D algorithm, requires a large amount of computation, and does not consider the correlation between different CT slices.
[0006] Chinese patent application CN202010785690.1 discloses an algorithm for detecting intracranial hemorrhage based on CNN and NLSTM neural networks applied to CT images. Specifically, the method involves using a CNN neural network to extract image features from CT images, combining the extracted image embedding with the patient's sequence information as input to an NLSTM neural network, calculating the loss using a cross-entropy loss function, and then backpropagating the network. The final network structure is then used for testing. This invention combines CNN and RNN neural networks, and the trained NLSTM network can focus on the differences between the current image and the previous / next image in the same sequence. Summary of the Invention
[0007] This invention aims to solve the problems of difficulty in fully utilizing information from upper and lower layers for segmentation and large differences in segmentation results between adjacent layers when automatically segmenting intracranial hemorrhage areas on CT images. Therefore, it proposes an automatic segmentation method for intracranial hemorrhage areas based on multi-slice CT images.
[0008] This invention provides an automatic segmentation method for intracranial hemorrhage regions based on multi-slice CT images, comprising:
[0009] Step S01: Construct a training dataset based on historical CT images with annotated hemorrhage areas; the training dataset includes multiple training data groups, each training data group consisting of a target CT image, an upper-layer CT image preceding the target CT image sequence, and a lower-layer CT image following the target CT image sequence;
[0010] Step S02 involves inputting the training data set into the bleeding region segmentation model for training, specifically as follows:
[0011] Step S21: Use a feature extraction network to extract image features of the target CT image, upper CT image and lower CT image in the training data group respectively;
[0012] Step S22: Using a feature fusion network, the image features of the target CT image are fused based on the correlation between the image features of the target CT image and the image features of the upper and lower CT images, respectively, to obtain the target fused features.
[0013] Step S23: Input the target fusion features into the hemorrhage probability prediction network to obtain the probability map of the target CT image belonging to the hemorrhage area;
[0014] Step S24: Calculate the loss value using a hybrid loss function consisting of cross-entropy and dice, and calculate the gradient value using the gradient backpropagation algorithm.
[0015] Input multiple training data sets and repeat steps S21 to S24 until the loss value reaches the minimum, and the bleeding region segmentation model is trained.
[0016] Step S03: Obtain multiple data groups to be analyzed in the CT sequence to be tested. Each data group to be analyzed consists of the CT image to be tested, the pre-CT image before the CT image sequence to be tested, and the post-CT image after the CT image sequence to be tested.
[0017] Step S04: According to the order of the CT sequence to be tested, the data groups to be analyzed in the CT sequence to be tested are sequentially input into the trained hemorrhage region segmentation model to obtain the hemorrhage region probability map of all layers of the CT image to be tested in the CT sequence to be tested.
[0018] Since even thin-slice CT scans have a thickness exceeding 1 mm, image information from layers above offers little help in analyzing the current layer. Therefore, this invention constructs a fully convolutional deep neural network and designs a mechanism capable of automatically calculating the correlation between features of upper and lower layers. This allows the model to predict hemorrhage areas using information from upper and lower layers without performing 3D computations. Furthermore, during training, a hybrid loss method incorporating cross-entropy and DICE loss is employed to provide both pixel-level and full-image-level supervision. This method fully utilizes information from upper and lower layers in the CT sequence with only a small increase in computation, while simultaneously improving the accuracy of hemorrhage area segmentation and the consistency between adjacent layers.
[0019] Preferably, step S04 further includes extracting the bleeding area in the CT image under test when the probability of the bleeding area in the CT image under test is greater than the bleeding probability threshold.
[0020] Preferably, step S04 further includes filtering the hemorrhage area in the extracted CT image to be tested using a stop filter, and filling the holes using morphological operations.
[0021] Preferably, the method further includes: step S05, using a connected component algorithm to process the hemorrhage areas of the CT images of all layers in the CT sequence to obtain the number and location of hemorrhage areas in the CT sequence.
[0022] Preferably, the feature extraction network includes 3*3 convolutional kernels and 4 downsampling layers with a stride of 2, each convolutional kernel having a Leaky ReLU activation function.
[0023] Preferably, step S22 includes:
[0024] Step S22.1: Perform feature transformation on the image features obtained in step S21;
[0025] Step S22.2, using the correlation formula Y=Softmax((W*X′)X T ) Calculate the correlation between the image features of the target CT image and the image features of the upper CT image, and the correlation between the image features of the target CT image and the image features of the lower CT image, respectively. Y is the correlation.
[0026] Where W is the weight layer, X′ represents the image features of the upper CT image or the lower CT image, and X represents the image features of the target CT image;
[0027] Step S22.3: Calculate the target fusion features using the fusion formula O = X + Y1*X1 + Y2*X2;
[0028] Where X1 and X2 represent the image features of the upper CT image and the lower CT image, respectively; Y1 is the correlation between the image features of the target CT image and the image features of the upper CT image; Y2 is the correlation between the image features of the target CT image and the image features of the lower CT image; and O is the target fusion feature.
[0029] Preferably, step S22.1 specifically involves performing feature transformation on the image features obtained in step S21 using a 1*1 convolution kernel.
[0030] Preferably, step S23 includes:
[0031] Step S23.1: Using a bleed probability prediction network including 3*3 convolutional kernels and skip connection layers and an upsampling method, the target fusion features obtained in step S22 are processed to obtain target fusion features that conform to the original image size; each convolutional kernel has a Leaky ReLU activation function;
[0032] Step S23.2, using Calculate the probability that each pixel belongs to the hemorrhage region and form a probability map of the target CT image belonging to the hemorrhage region; where zi represents the i-th output node, and C represents the total number of output nodes, which is set to 2.
[0033] Preferably, step S24 includes:
[0034] Step S24.1: Calculate the loss value using the hybrid loss function Tatol Loss = CELoss + DiceLoss, which consists of cross-entropy and dice.
[0035] Where, CELoss=-∑ i y i log(p i ), where y represents the true value and p represents the predicted value obtained through step S23; X represents the set of pixels predicted as bleeding, and Y represents the set of pixels actually labeled as bleeding;
[0036] Step S24.2: Calculate the gradient value based on the loss value and gradient backpropagation algorithm obtained in step S24.1.
[0037] The loss value is calculated using a hybrid loss function consisting of cross-entropy and dice, and the gradient value is calculated using a gradient backpropagation algorithm.
[0038] Preferably, in step S03, the data set to be analyzed is obtained from the CT sequence to be tested using a sliding window method.
[0039] The present invention has the following beneficial effects:
[0040] This invention presents an automatic segmentation method for intracranial hemorrhage regions based on multi-slice CT images. It addresses the challenges of fully utilizing information from upper and lower slices and the significant differences in segmentation results between adjacent slices when automatically segmenting intracranial hemorrhage regions on CT images. During training, the method obtains the fusion features of the target layer by calculating the correlation between upper and lower slices. Then, a hybrid loss method incorporating cross-entropy loss and DICE loss is employed to provide both pixel-level and full-image-level supervision. Ultimately, this invention fully utilizes information from upper and lower slices in the CT sequence to achieve accurate and consistent segmentation of the hemorrhage region across adjacent slices. Attached Figure Description
[0041] Figure 1 This is a flowchart of an automatic segmentation method for intracranial hemorrhage regions based on multi-slice CT images according to the present invention.
[0042] Figure 2This is a training flowchart for an automatic segmentation method for intracranial hemorrhage regions based on multi-slice CT images according to the present invention.
[0043] Figure 3 This is a schematic diagram illustrating the process of using the bleeding region segmentation model. Detailed Implementation
[0044] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.
[0045] like Figure 1 , 2 An automatic segmentation method for intracranial hemorrhage regions based on multi-slice CT images, comprising:
[0046] Step S01: Construct a training dataset based on historical CT images with annotated hemorrhage areas; the training dataset includes multiple training data groups, each training data group consisting of a target CT image, an upper-layer CT image preceding the target CT image sequence, and a lower-layer CT image following the target CT image sequence;
[0047] Step S02 involves inputting the training data set into the bleeding region segmentation model for training, specifically as follows:
[0048] Step S21: Use a feature extraction network to extract image features of the target CT image, upper CT image and lower CT image in the training data group respectively;
[0049] Step S22: Using a feature fusion network, the image features of the target CT image are fused based on the correlation between the image features of the target CT image and the image features of the upper and lower CT images, respectively, to obtain the target fused features.
[0050] Step S23: Input the target fusion features into the hemorrhage probability prediction network to obtain the probability map of the target CT image belonging to the hemorrhage area;
[0051] Step S24: Calculate the loss value using a hybrid loss function consisting of cross-entropy and dice, and calculate the gradient value using the gradient backpropagation algorithm.
[0052] Input multiple training data sets and repeat steps S21 to S24 until the loss value reaches the minimum, and the bleeding region segmentation model is trained.
[0053] Step S03: Obtain multiple data groups to be analyzed in the CT sequence to be tested. Each data group to be analyzed consists of the CT image to be tested, the pre-CT image before the CT image sequence to be tested, and the post-CT image after the CT image sequence to be tested.
[0054] Step S04: According to the order of the CT sequence to be tested, the data groups to be analyzed in the CT sequence to be tested are sequentially input into the trained hemorrhage region segmentation model to obtain the hemorrhage region probability map of all layers of the CT image to be tested in the CT sequence to be tested.
[0055] In step S01, CT images of the correctly labeled bleeding areas by radiologists are acquired and used as historical CT images to construct the training dataset. The historical CT images are stored in sequence. The target CT image is acquired, along with the upper and lower CT images preceding the target CT image sequence. These three consecutive images are used as a training data set, and multiple training data sets acquired from the sequence are then used to construct the training dataset.
[0056] Step S02 trains the bleeding region segmentation model based on the training data set. First, the three images in the training data set undergo the same image processing and are then fed into the bleeding region segmentation model; features are extracted from the three images using a feature extraction network. The feature extraction network includes 3*3 convolutional kernels and four downsampling layers with a stride of 2, each convolutional kernel having a Leaky ReLU activation function. After passing through this feature extraction network, a feature map representing 1 / 16 of the input image is obtained. Next, feature fusion processing is performed.
[0057] Step S22 includes:
[0058] Step S22.1: Perform feature transformation on the image features obtained in step S21; F(X) = X*k, where k represents a convolution kernel with a scale of 1*1.
[0059] Step S22.2, using the correlation formula Y=Softmax((W*X′)X T ) Calculate the correlation between the image features of the target CT image and the image features of the upper CT image, and the correlation between the image features of the target CT image and the image features of the lower CT image, respectively. Y is the correlation; where W is the weight layer, X′ represents the image features of the upper CT image or the image features of the lower CT image, and X represents the image features of the target CT image.
[0060] Step S22.3: Calculate the target fusion features using the fusion formula O = X + Y1*X1 + Y2*X2;
[0061] Where X1 and X2 represent the image features of the upper CT image and the lower CT image, respectively; Y1 is the correlation between the image features of the target CT image and the image features of the upper CT image; Y2 is the correlation between the image features of the target CT image and the image features of the lower CT image; and O is the target fusion feature.
[0062] The weight layer, also known as the parameter layer, is built in the initial stage of the model and its parameters are continuously updated during the training process until the optimal values are obtained.
[0063] By calculating the correlation between adjacent layers and using the calculated correlation as weights, a target fusion feature that incorporates information from adjacent layers is then calculated. Subsequently, based on the target fusion feature, a probability map is obtained showing that the target CT image belongs to a hemorrhage region.
[0064] Step S23 includes:
[0065] Step S23.1 involves using a bleed probability prediction network comprising 3x3 convolutional kernels and skip connection layers, along with an upsampling method, to process the target fusion features obtained in step S22 to obtain target fusion features that conform to the original image size. Each convolutional kernel has a Leaky ReLU activation function. For example, the fusion information feature map is upsampled four times until it returns to the original image size. The skip connection layers directly propagate features learned at different scales from the downsampling layers to the layers used for bleed probability prediction.
[0066] Step S23.2, using Calculate the probability that each pixel belongs to the hemorrhage region and form a probability map of the target CT image belonging to the hemorrhage region; where zi represents the i-th output node, and C represents the total number of output nodes, which is set to 2.
[0067] In network training, we use a hybrid loss function for parameter optimization, including cross-entropy loss for pixel-level supervision and DICE loss for full-image-level supervision. Step S24 includes:
[0068] Step S24.1: Calculate the loss value using the hybrid loss function Tatol Loss = CELoss + DiceLoss, which consists of cross-entropy and dice.
[0069] Where, CELoss=-∑ i y i log(p i ), where y represents the pre-labeled true value, labeled by professional labelers, and p represents the predicted value obtained through step S23; X represents the set of pixels predicted as bleeding, and Y represents the set of pixels actually labeled as bleeding;
[0070] Step S24.2: Calculate the gradient value based on the loss value and gradient backpropagation algorithm obtained in step S24.1.
[0071] Based on the calculated loss value, the gradient value of the last layer of neurons is calculated by differentiation. The gradient values of the neurons in the preceding layers are calculated step-by-step using the chain rule. The gradient formula is the derivative formula obtained according to the chain rule. For example, for F(H(x)), the gradient value of the neurons in layer F is dF / dH·dH / dx. Then, the gradient descent algorithm is used to update the values of all neurons in the network, so that the loss value continuously decreases, that is, the model prediction value continuously approaches the labeled value of the training data, and finally obtains the optimal network parameters. Neurons include all parts of the feature extraction network, feature fusion network, and probabilistic prediction network that need to be trained. These three networks are sequential, and gradient calculation can be directly passed. The gradient descent algorithm is used for updating, and the updated parameters include all trainable parameters in the network, also known as neurons or weights.
[0072] Repeat the training process until the gradient stops decreasing, then freeze the parameters and output the model.
[0073] In steps S03 and S04, the data set to be analyzed is input into the model, and the probability map of the hemorrhage area of the CT image to be tested is output. Specifically, after acquiring the CT sequence, a sliding window of 3 frames is used to obtain the data set to be analyzed. When the obtained 3 frames of data lack upper and lower layer images due to the beginning and end of the sequence, they are replaced with a completely black image. The processed data set to be analyzed is input into the model. First, feature extraction is performed on the input image, and then the extracted features are fused. The obtained fused features are input into the hemorrhage probability prediction network to obtain the probability that each pixel belongs to the hemorrhage area, and a probability map of the hemorrhage area of the CT image to be tested is formed. The process of feature extraction, transformation, fusion, and calculation of the probability of the hemorrhage area in the model is the same as steps S21 to S23 during model training. Since the optimal network parameters have been obtained after training, in the testing phase, after inputting the data set to be analyzed, an accurate probability map of the hemorrhage area of the CT image to be tested can be obtained based on the optimal network parameters.
[0074] Step S04 of the method of the present invention further includes extracting the hemorrhage area in the CT image under test when the probability of the hemorrhage area is greater than the hemorrhage probability threshold. The hemorrhage probability threshold is a critical value for the hemorrhage probability of healthy tissue, for example, 0.5. When it is higher than 0.5, the hemorrhage area in the CT image under test is considered abnormal hemorrhage. The hemorrhage probability threshold can fluctuate slightly up or down according to actual judgment needs. In addition, during the extraction process, the following processing is also required: using a stop filter to filter the extracted hemorrhage area in the CT image under test, removing discrete small targets in the result, and using morphological operations to fill holes. Where B represents a symmetric structural element, A c This is the complement of the original image, and X represents the filled result. The above process is performed on each layer, ultimately yielding the hemorrhage areas in all layers of the CT sequence.
[0075] The method of this invention further includes step S05, which uses a connected component algorithm to process the hemorrhage regions of the CT images in all layers of the CT sequence to obtain the number and location of hemorrhage regions in the CT sequence. The connected component algorithm refers to an algorithm that determines whether pixels are connected based on their four adjacent positions (up, down, left, and right) to construct connected regions.
[0076] Figure 3 The diagram illustrates the practical application phase. After acquiring the patient's CT scan sequence, a 3-frame sliding window is used to obtain the data set to be analyzed. Then, the data set is input into the model sequentially, and the probability of hemorrhage areas in the target layer is output. Next, hemorrhage areas are extracted based on a set threshold, and the image is processed. This process of inputting into the model and extracting hemorrhage areas is repeated to obtain the hemorrhage areas in all layers of the CT sequence. Based on a connected component algorithm, the number of hemorrhage areas in the CT sequence and their corresponding locations are output.
[0077] This invention proposes a method for segmenting hemorrhage regions by fusing image information from upper and lower layers of the target layer. With only a small increase in parameters and computation time, it improves the accuracy of hemorrhage region segmentation and demonstrates good consistency in segmentation results between adjacent layers, which is helpful for counting the number of hemorrhage lesions and calculating the volume of each hemorrhage region. The proposed method for segmenting hemorrhage regions in cranial CT scans is characterized by high speed, high accuracy, and strong robustness, and has strong potential for practical application.
[0078] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any modifications or variations of the implementation of the present invention may be made without departing from the stated principles.
Claims
1. An automatic segmentation method of intracranial hemorrhage area based on multi-slice CT images, characterized in that, include: Step S01: Construct a training dataset based on historical CT images with labeled hemorrhage areas; The training dataset includes multiple training data groups, each training data group consisting of a target CT image, an upper-layer CT image preceding the target CT image sequence, and a lower-layer CT image following the target CT image sequence. Step S02 involves inputting the training data set into the bleeding region segmentation model for training, specifically as follows: Step S21: Use a feature extraction network to extract image features of the target CT image, upper CT image and lower CT image in the training data group respectively; Step S22: Using a feature fusion network, the image features of the target CT image are fused based on the correlation between the image features of the target CT image and the image features of the upper and lower CT images, respectively, to obtain the target fused features. Step S23: Input the target fusion features into the hemorrhage probability prediction network to obtain a probability map of the target CT image belonging to the hemorrhage region; including: Step S23.1: Using a hemorrhage probability prediction network including 3*3 convolutional kernels and skip connection layers and an upsampling method, perform image size processing on the target fusion features obtained in Step S22 to obtain target fusion features that conform to the original image size; each convolutional kernel has a Leaky ReLU activation function; Step S23.2: Utilize Calculate the probability that each pixel belongs to the hemorrhage region, and generate a probability map of the target CT image belonging to the hemorrhage region; where z i Let represent the i-th output node, and C represent the total number of output nodes, which is set to 2; Step S24 involves calculating the loss value using a hybrid loss function consisting of cross-entropy and dice, and calculating the gradient value using a gradient backpropagation algorithm; this includes: Step S24.1, calculating the loss value using the hybrid loss function Tatol Loss = CELoss + DiceLoss, consisting of cross-entropy and dice; wherein, y represents the pre-labeled true value, and p represents the predicted value obtained through step S23; M represents the set of pixels predicted as bleeding, and N represents the set of pixels actually labeled as bleeding; Step S24.2, based on the loss value calculated in step S24.1 and the gradient backpropagation algorithm, calculate the gradient value; Input multiple training data sets and repeat steps S21 to S24 until the loss value reaches the minimum, and the bleeding region segmentation model is trained. Step S03: Obtain multiple data groups to be analyzed in the CT sequence to be tested. Each data group to be analyzed consists of the CT image to be tested, the pre-CT image before the CT image sequence to be tested, and the post-CT image after the CT image sequence to be tested. Step S04: According to the order of the CT sequence to be tested, the data groups to be analyzed in the CT sequence to be tested are sequentially input into the trained hemorrhage region segmentation model to obtain the hemorrhage region probability map of all layers of the CT image to be tested in the CT sequence to be tested.
2. The method for automatic segmentation of intracranial hemorrhage region based on multi-slice CT images according to claim 1, characterized in that, Step S04 further includes extracting the bleeding area in the CT image under test when the probability of bleeding area in the CT image under test is greater than the bleeding probability threshold.
3. The method for automatic segmentation of intracranial hemorrhage region based on multi-slice CT images according to claim 2, characterized in that, Step S04 further includes filtering the hemorrhage area in the extracted CT image to be tested using a stop filter, and filling the holes using morphological operations.
4. The method for automatic segmentation of intracranial hemorrhage region based on multi-slice CT images according to claim 2, characterized in that, It also includes: step S05, using a connected component algorithm to process the hemorrhage areas of the CT images of all layers in the CT sequence to obtain the number and location of hemorrhage areas in the CT sequence.
5. The method for automatic segmentation of intracranial hemorrhage region based on multi-slice CT images according to claim 1, characterized in that, The feature extraction network includes 3*3 convolutional kernels and 4 downsampling layers with a stride of 2. Each convolutional kernel has a LeakyReLU activation function.
6. The method for automatic segmentation of intracranial hemorrhage region based on multi-slice CT images according to claim 1, characterized in that, Step S22 includes: Step S22.1: Perform feature transformation on the image features obtained in step S21; Step S22.2, using the correlation formula Calculate the correlation between the image features of the target CT image and the image features of the upper CT image, and the correlation between the image features of the target CT image and the image features of the lower CT image, respectively. Y represents the correlation. Where W is the weight layer, This represents the image features of the upper or lower CT image, where X represents the image features of the target CT image. Step S22.3, using the fusion formula Calculate target fusion features; Where X1 and X2 represent the image features of the upper CT image and the lower CT image, respectively, Y1 is the correlation between the image features of the target CT image and the image features of the upper CT image, Y2 is the correlation between the image features of the target CT image and the image features of the lower CT image, and O is the target fusion feature.
7. The method for automatic segmentation of intracranial hemorrhage region based on multi-slice CT images according to claim 6, characterized in that, Specifically, step S22.1 involves performing feature transformation on the image features obtained in step S21 using a 1*1 convolution kernel.
8. The method for automatic segmentation of intracranial hemorrhage region based on multi-slice CT images according to claim 1, characterized in that, In step S03, the data set to be analyzed is obtained from the CT sequence to be tested using a sliding window method.