Training of intracranial hemorrhage recognition model and methods, systems and devices for intracranial hemorrhage recognition

Abstract

This invention discloses a training method, system, and device for intracranial hemorrhage recognition model training and intracranial hemorrhage recognition. The method includes: S1, constructing an intracranial hemorrhage recognition model, which includes a first-stage network for region segmentation and a second-stage network for hemorrhage recognition; S2, training the first-stage network; and S3, training the second-stage network to complete the training of the intracranial hemorrhage recognition model. The intracranial hemorrhage recognition method includes: acquiring CT images scanned from a patient, inputting the CT images into the intracranial hemorrhage recognition model, and outputting the predicted hemorrhage area and hemorrhage type. This invention proposes a two-stage intracranial hemorrhage analysis model, which can simultaneously perform region segmentation and site identification of intracranial hemorrhage, helping to provide more specific and effective auxiliary diagnostic opinions in practical applications.

Description

Technical Field

[0001] This invention relates to the field of intracranial hemorrhage recognition, and in particular to an intracranial hemorrhage recognition model training method, system and device. Background Technology

[0002] Intracranial hemorrhage is a relatively common critical illness in clinical practice. It is a general term for traumatic and non-traumatic bleeding in different parts of the skull. The specific treatment plan depends on the amount and location of the bleeding. Therefore, before determining the treatment plan, doctors often need to use medical imaging technology to examine the patient's intracranial condition.

[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] In recent years, automated segmentation methods based on computer algorithms have been increasingly used to assist in clinical diagnosis and treatment decisions due to their advantages of eliminating subjective errors and saving time and effort. Compared to manual labeling methods, this method is more time-saving and labor-saving, significantly improving efficiency. Furthermore, in terms of accuracy, the hematoma boundaries of most patients are not clearly defined, and manual labeling methods heavily rely on the doctor's experience, leading to significant segmentation errors. Computer algorithms, through learning from standardized labeled data, achieve higher consistency in results. However, current computer segmentation algorithms often only predict and segment the hemorrhage area. While this can calculate the amount of bleeding, it cannot identify the location of the hemorrhage within the brain. Hemorrhages in different locations often produce different symptoms, and the diagnosis, treatment, and prognosis will also differ. For example, in basal ganglia hemorrhages with a volume less than 30 ml, the patient is often conscious and can be treated conservatively. However, in cerebellar hemorrhages with a volume greater than 10 ml, the patient may become comatose, requiring surgical removal of the hematoma. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and device for training a model for recognizing intracranial hemorrhage and for recognizing intracranial hemorrhage, thereby solving the problems of training a model for recognizing intracranial hemorrhage and recognizing hemorrhage.

[0006] This invention provides a training method for a model to identify intracranial hemorrhage and a method for identifying intracranial hemorrhage, including:

[0007] S1. Construct a brain hemorrhage recognition model, the recognition model including: a first-stage network for region segmentation and a second-stage network for hemorrhage recognition;

[0008] S2, Training the first-stage network;

[0009] S3. Train the second-stage network to complete the training of the intracranial hemorrhage recognition model.

[0010] This invention provides a method for identifying intracranial hemorrhage using a model: acquiring CT images of a patient, inputting the CT images into the intracranial hemorrhage identification model, and outputting the predicted hemorrhage area and hemorrhage type.

[0011] The present invention also provides a training model for intracranial hemorrhage recognition and an intracranial hemorrhage recognition system, comprising:

[0012] A construction module is used to build a brain hemorrhage recognition model, the recognition model including: a first-stage network for region segmentation and a second-stage network for hemorrhage recognition;

[0013] First training module: Training the first-stage network;

[0014] The second training module trains the second-stage network to complete the training of the intracranial hemorrhage recognition model.

[0015] The recognition module is used to acquire CT images of the patient, input the CT images into the intracranial hemorrhage recognition model, and output the predicted hemorrhage area and hemorrhage type.

[0016] This invention also provides a training method for a brain hemorrhage recognition model and a brain hemorrhage recognition device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above method.

[0017] Using the embodiments of the present invention, the present invention proposes a two-stage intracranial hemorrhage analysis model. This model can simultaneously perform regional segmentation and site identification of intracranial hemorrhage, which helps to provide more specific and effective auxiliary diagnostic opinions in practical applications.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the intracranial hemorrhage recognition model training method according to an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the training method for the intracranial hemorrhage recognition model according to an embodiment of the present invention;

[0022] Figure 3 This is a flowchart illustrating the use of the intracranial hemorrhage identification method according to an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the intracranial hemorrhage recognition model training system according to an embodiment of the present invention;

[0024] Figure 5 This is a schematic diagram of the intracranial hemorrhage recognition model training device according to an embodiment of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Method Implementation Examples

[0027] According to embodiments of the present invention, a training model for recognizing intracranial hemorrhage and a method for recognizing intracranial hemorrhage are provided. Figure 1 This is a flowchart of the intracranial hemorrhage recognition model training method according to an embodiment of the present invention, such as... Figure 1 As shown, it specifically includes:

[0028] This invention proposes a method for hemorrhage lesion segmentation and location identification based on deep neural networks, aiming to address the limitation of existing CT-based intracranial hemorrhage identification algorithms, which can only perform lesion segmentation and hemorrhage volume calculation. This method constructs a two-stage convolutional neural network. First, a network with an encoder-decoder structure extracts the region of the hemorrhage lesion from the CT image. Then, a multi-channel input network containing image CT value information, 3D sequence information, and lesion region information accurately determines the location of each hemorrhage lesion within the brain. Building upon existing hemorrhage lesion segmentation algorithms based on intracranial CT images, this method fully utilizes sequence location information, dividing the brain into 42 regional partitions and analyzing the location of each hemorrhage lesion within the brain, providing more valuable information for assisted diagnosis and treatment.

[0029] A method for training a brain hemorrhage recognition model, comprising,

[0030] S1. Construct a brain hemorrhage recognition model, the recognition model including: a first-stage network for region segmentation and a second-stage network for hemorrhage recognition;

[0031] S2, Training the first-stage network;

[0032] S3. Train the second-stage network to complete the training of the intracranial hemorrhage recognition model.

[0033] S2 specifically includes:

[0034] S21. Construct a training data set based on CT images of the correct bleeding area and type as marked by the radiologist, and send the CT images in the training data set into the first-stage network.

[0035] S22. The first-stage network extracts features from the CT image to obtain the first feature. Based on the first feature, the probability that each pixel in the CT image belongs to the hemorrhage area is obtained. If the probability is higher than the set probability, it is a predicted hemorrhage area.

[0036] S23. Calculate the loss value based on the probability of the bleeding area and the correct bleeding area, and backfeed the first gradient to update the first-stage network.

[0037] S24. Repeat S21 to S24 until the first gradient no longer decreases, then freeze the network parameters for the first stage.

[0038] S3 specifically includes:

[0039] S31. Input the CT images in the training data set into the first-stage network to obtain the predicted hemorrhage area. Input the gray value of the CT image of the predicted hemorrhage area, the 3D location information in the CT sequence of the predicted hemorrhage area, and the binarized image of the predicted hemorrhage area into the second-stage network.

[0040] S32. The second-stage network obtains bleeding category features based on the grayscale value of CT images, 3D location information and binarized images, predicts which part of the brain the bleeding area belongs to based on the bleeding category features, and obtains the predicted bleeding category based on the brain part.

[0041] S33. Calculate the loss value based on the predicted bleeding category and the correct bleeding category and backfeed the second gradient to update the parameters of the second-stage network.

[0042] S34. Repeat S31 to S33 until the second gradient no longer decreases, freeze the network parameters in the second stage, and complete the model training.

[0043] A method for identifying intracranial hemorrhage, based on an intracranial hemorrhage identification model, includes,

[0044] The system acquires CT images of the patient, inputs these images into the intracranial hemorrhage recognition model, and outputs the predicted hemorrhage area and hemorrhage type.

[0045] The specific implementation method is as follows:

[0046] This invention proposes a method for segmenting intracranial hemorrhage regions and identifying hemorrhage sites in CT images based on a two-stage deep neural network. Figure 2 This is a schematic diagram of the training method for the intracranial hemorrhage recognition model according to an embodiment of the present invention. It includes two aspects: firstly, the training of a two-stage neural network. The overall network structure includes a hemorrhage region segmentation module consisting of an encoder and a decoder, and a hemorrhage site recognition module consisting of a feature extractor and a classifier. The model is trained using a step-by-step training method, with specific steps as follows: Figure 1 As shown; the second is to use the trained model to segment the bleeding area and classify the location of the bleeding in the input CT image.

[0047] The proposed first-stage neural network can be divided into two sub-modules: an encoder module for feature extraction and a decoder module for bleed probability prediction. The encoder module uses 3x3 convolutional kernels, each followed by a Leaky ReLU activation function. Four downsampling layers with a stride of 2 are used to obtain a feature map with a size of 1 / 16 of the input image. In the decoder module, similarly 3x3 convolutional kernels and a Leaky ReLU activation function are used. The upsampling method is applied to the feature map four times until it returns to the original image size. Finally, a softmax function is used to obtain the probability of each pixel being a bleed region. Skip connections are also used to directly propagate features learned at different scales from the downsampling network to the layers used for bleed probability prediction.

[0048] The second-stage neural network proposed in this invention can also be divided into two sub-modules: a downsampling module for feature extraction and a classifier module for bleeding site identification. The downsampling module uses 3x3 convolutional kernels, each followed by a Leaky ReLU activation function. Four downsampling layers with a stride of 2 are used to obtain a feature map with a size of 1 / 16 of the input image. In the classifier module, two fully connected layers are used to map the high-dimensional feature map to a low-dimensional dimension, and then a sigmoid function is used for multi-label classification of the bleeding site. Specifically, the input to the second-stage neural network is multi-channel, including the HU value of the CT image, the location coordinates in the CT sequence, and the bleeding area information output by the first-stage network.

[0049] In network training, the first stage uses a hybrid loss function for parameter optimization, including cross-entropy loss for pixel-level supervision and DICE loss for full-image-level supervision. The second stage uses cross-entropy loss for parameter optimization.

[0050] During the training phase, a training dataset was constructed using CT images correctly labeled with hemorrhage areas and locations by radiologists. A supervised training method was used to train the proposed two-stage deep neural network. The training steps can be summarized as follows: Step 1: Construct the training dataset by feeding single-slice CT images into the model. Step 2: For the input image, the encoder extracts image features, and then the decoder network obtains the probability that each pixel belongs to the hemorrhage area. Step 3: Calculate the loss value using the predicted hemorrhage probability and label, and backpropagate the gradient to update the parameters of the first-stage network. Step 4: Repeat steps 1 to 3 until the gradient no longer decreases, then freeze the parameters of the first-stage network. Step 5: Input a single-slice CT image into the model. The first-stage network obtains the predicted hemorrhage area, and then the grayscale value of the CT image, its 3D position information in the sequence, and the binarized image of the predicted hemorrhage area are fed into the second-stage network as input. Step 6: The feature extraction network in the second-stage network extracts features, and then the classifier network predicts which of the 42 brain regions the hemorrhage area of ​​the input CT image belongs to. Step 7: Calculate the loss value using the predicted location and the actual labeled location, and backpropagate the gradient to update the parameters of the second-stage network. Step 8: Repeat steps 5 to 7 until the gradient stops decreasing, then freeze the parameters of the second-stage network to complete model training.

[0051] Figure 3 This is a flowchart illustrating the usage of the intracranial hemorrhage recognition method according to an embodiment of the present invention. In the model usage phase, the trained model is used to segment the hemorrhage region and identify the hemorrhage site according to the following steps: First, acquire the patient's CT scan sequence, input the data to be analyzed into the model in sequence, and output the probability map of the hemorrhage region in the target layer. Second, extract the hemorrhage region according to a set threshold, use a median filter to remove discrete small targets in the result, and use a connected component analysis algorithm to obtain connected hemorrhage regions. Third, predict the intracranial location of the hemorrhage region based on the CT image and the connected hemorrhage regions.

[0052] As can be seen from the above description, this invention addresses the fact that current algorithms for analyzing intracranial hemorrhage in CT images based on computer algorithms can only segment the hemorrhage region and calculate the hemorrhage volume based on the segmentation results, but cannot identify the equally important hemorrhage site in diagnosis and treatment. Therefore, this invention innovatively proposes a two-stage intracranial hemorrhage analysis model that can simultaneously segment the hemorrhage region and identify the site, helping to provide more specific and effective auxiliary diagnostic opinions in practical applications. The intracranial CT hemorrhage region segmentation and site identification method proposed in this invention has the characteristics of high speed, high accuracy, and strong robustness, and has strong practical application prospects.

[0053] System Implementation Examples

[0054] According to embodiments of the present invention, a training system for a brain hemorrhage recognition model is provided. Figure 4 This is a schematic diagram of the intracranial hemorrhage recognition model training system according to an embodiment of the present invention, as shown below. Figure 4 As shown, it specifically includes:

[0055] A construction module is used to build a brain hemorrhage recognition model, the recognition model including: a first-stage network for region segmentation and a second-stage network for hemorrhage recognition;

[0056] First training module: Training the first-stage network;

[0057] The second training module trains the second-stage network to complete the training of the intracranial hemorrhage recognition model.

[0058] The first training module is specifically used for:

[0059] S51. Construct a training data set based on CT images of the correct bleeding area and type as marked by the radiologist, and send the CT images in the training data set into the first-stage network.

[0060] S52. Control the first-stage network to extract features from the CT image to obtain the first feature. Based on the first feature, obtain the probability that each pixel in the CT image belongs to the hemorrhage area. If the probability is higher than the set probability, it is a predicted hemorrhage area.

[0061] S53. Calculate the loss value based on the probability of the bleeding area and the correct bleeding area, and backfeed the first gradient to update the first-stage network.

[0062] S54. Repeat S51 to S54 until the first gradient no longer decreases, and save the network parameters of the first stage.

[0063] The second training module is specifically used for:

[0064] S61. Input the CT images in the training data set into the first-stage network to obtain the predicted hemorrhage area. Input the gray value of the CT image of the predicted hemorrhage area, the 3D location information in the CT sequence of the predicted hemorrhage area, and the binarized image of the predicted hemorrhage area into the second-stage network.

[0065] S62. The second-stage control network obtains bleeding category features based on the grayscale value of the CT image, 3D location information and binarized image, predicts which part of the brain the bleeding area belongs to based on the bleeding category features, and obtains the predicted bleeding category based on the brain part.

[0066] S63. Calculate the loss value based on the predicted bleeding category and the correct bleeding category and backfeed the second gradient to update the parameters of the second-stage network.

[0067] S64. Repeat S61 to S63 until the second gradient no longer decreases, save the network parameters for the second stage, and complete the model training.

[0068] A system for recognizing intracranial hemorrhage includes a recognition module for acquiring CT images of a patient, inputting the CT images into an intracranial hemorrhage recognition model, and outputting a predicted hemorrhage area and hemorrhage type.

[0069] The embodiments of the present invention are system embodiments corresponding to the above method embodiments. The specific operation of each module can be understood by referring to the description of the method embodiments, and will not be repeated here.

[0070] Device Example 1

[0071] This invention provides a training device for a brain hemorrhage recognition model, such as... Figure 5 As shown, it includes: a memory 50, a processor 52, and a computer program stored on the memory 50 and executable on the processor 52. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions to the technical solutions of the embodiments of the present invention do not cause the essence of the corresponding technical solutions to deviate from the scope of the present solution.

Claims

1. A method for training a model to identify intracranial hemorrhage, characterized in that, include, S1. Construct a brain hemorrhage recognition model, the recognition model including: a first-stage network for region segmentation and a second-stage network for hemorrhage recognition; S2, Training the first-stage network; S3. Train the second-stage network to complete the training of the intracranial hemorrhage recognition model; S2 specifically includes: S21. Construct a training data set based on CT images of the correct bleeding area and type as marked by the radiologist, and send the CT images in the training data set into the first-stage network. S22. The first-stage network extracts features from the CT image to obtain the first feature. Based on the first feature, the probability that each pixel in the CT image belongs to the hemorrhage area is obtained. If the probability is higher than the set probability, it is a predicted hemorrhage area. S23. Calculate the loss value based on the probability of the bleeding area and the correct bleeding area, and backfeed the first gradient to update the first-stage network. S24. Repeat S21 to S24 until the first gradient no longer decreases, then freeze the network parameters for the first stage. S3 specifically includes: S31. Input the CT images in the training data set into the first-stage network to obtain the predicted hemorrhage area. Input the gray value of the CT image of the predicted hemorrhage area, the 3D location information in the CT sequence of the predicted hemorrhage area, and the binarized image of the predicted hemorrhage area into the second-stage network. S32. The second-stage network obtains bleeding category features based on the grayscale value of CT images, 3D location information and binarized images, predicts which part of the brain the bleeding area belongs to based on the bleeding category features, and obtains the predicted bleeding category based on the brain part. S33. Calculate the loss value based on the predicted bleeding category and the correct bleeding category and backfeed the second gradient to update the parameters of the second-stage network. S34. Repeat S31 to S33 until the second gradient no longer decreases, freeze the network parameters in the second stage, and complete the model training.

2. A training system for a model to identify intracranial hemorrhage, characterized in that, include, A construction module is used to build a brain hemorrhage recognition model, the recognition model including: a first-stage network for region segmentation and a second-stage network for hemorrhage recognition; First training module: Training the first-stage network; Second training module: Training the second-stage network to complete the training of the intracranial hemorrhage recognition model; The first training module is specifically used for: S51. Construct a training data set based on CT images of the correct bleeding area and type as marked by the radiologist, and send the CT images in the training data set into the first-stage network. S52. Control the first-stage network to extract features from the CT image to obtain the first feature. Based on the first feature, obtain the probability that each pixel in the CT image belongs to the hemorrhage area. If the probability is higher than the set probability, it is a predicted hemorrhage area. S53. Calculate the loss value based on the probability of the bleeding area and the correct bleeding area, and backfeed the first gradient to update the first-stage network. S54. Repeat S51 to S54 until the first gradient no longer decreases, then save the network parameters for the first stage. The second training module is specifically used for: S61. Input the CT images in the training data set into the first-stage network to obtain the predicted hemorrhage area. Input the gray value of the CT image of the predicted hemorrhage area, the 3D location information in the CT sequence of the predicted hemorrhage area, and the binarized image of the predicted hemorrhage area into the second-stage network. S62. The second-stage control network obtains bleeding category features based on the grayscale value of the CT image, 3D location information and binarized image, predicts which part of the brain the bleeding area belongs to based on the bleeding category features, and obtains the predicted bleeding category based on the brain part. S63. Calculate the loss value based on the predicted bleeding category and the correct bleeding category and backfeed the second gradient to update the parameters of the second-stage network. S64. Repeat S61 to S63 until the second gradient no longer decreases, save the network parameters for the second stage, and complete the model training.

3. A method for identifying intracranial hemorrhage, characterized in that, The intracranial hemorrhage recognition model obtained based on the intracranial hemorrhage recognition model training method as described in claim 1 includes, The system acquires CT images of the patient, inputs these images into the intracranial hemorrhage recognition model, and outputs the predicted hemorrhage area and hemorrhage type.

4. A system for identifying intracranial hemorrhage, characterized in that, include: The recognition module is used to acquire CT images scanned by the patient, input the CT images into the intracranial hemorrhage recognition model obtained by the intracranial hemorrhage recognition model training method as described in claim 1, and output the predicted hemorrhage area and hemorrhage type.

5. A training device for a model to identify intracranial hemorrhage, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the intracranial hemorrhage recognition model training method as described in claim 1.

6. A device for identifying intracranial hemorrhage, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the intracranial hemorrhage identification method as described in claim 3.

Images