CNN and RNN-based intracranial hemorrhage CT image classification method
By combining CNN and RNN in intracranial hemorrhage detection, using fine-grained convolution kernel and timing information, combined with principal component analysis, the problems of feature expression ability and computational complexity in the prior art are solved, and high accuracy and The efficient classification of intracranial hemorrhage subtypes significantly improves the accuracy and real-time detection.
Patent Information
- Application Number
- CN202410976605.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-20
- Publication Date
- 2024-11-08
AI Technical Summary
The existing intracranial hemorrhage detection methods have shortcomings in feature expression ability, spatial relationship utilization, fine-grained classification, computational complexity and data imbalance, resulting in inaccuracy and inefficiency, making it difficult to achieve accurate classification of subtypes of intracranial hemorrhage and early diagnosis. .
Using the CT image classification method based on CNN and RNN, the image details are captured through fine-grained convolution kernels, RNN processing timing information is introduced, and the calculation complexity is reduced in combination with principal component analysis, and the model is improved through data preprocessing and model combination. Awesomeness and generalization ability.
It significantly improved the accuracy and efficiency of the classification of intracranial hemorrhage subtypes, with the AUC reaching 96.75%, with an accuracy rate of 97.71%, leading to other methods by 0.76 percentage points, significantly improving the accuracy of classification of intracranial hemorrhage subtypes and real-time detection.
Smart Images

Figure CN118918374A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of intelligent medical image processing and relates to a CNN and RNN-based intracranial hemorrhage CT image classification method. Background Art
[0002] Intracranial hemorrhage is a potentially life-threatening neurological emergency characterized by ruptured blood vessels and bleeding. Its main subtypes are based on the location of the hemorrhage, including intracerebral hemorrhage (ICH), subarachnoid hemorrhage (SAH), intraventricular hemorrhage (IVH), epidural hemorrhage (EDH), and subdural hemorrhage (SDH). The accuracy of the classification of intracranial hemorrhage subtypes is very important for the early diagnosis of ICH. Accurate classification helps guide doctors to assess the progression of the disease, provide appropriate clinical intervention as early as possible, and formulate treatment plans for different categories, which is an important means to reduce the mortality rate of intracranial hemorrhage. The traditional method of detecting intracranial hemorrhage is for radiologists to visually inspect and quantitatively evaluate the images. This diagnostic process is not only time-consuming and labor-intensive, but also the accuracy of the judgment depends on the work experience of the radiologist. With the advancement of information technology and the continuous development of artificial intelligence, computer-aided diagnosis (CAD) uses medical image processing technology and machine learning technology to automatically analyze and interpret pathological data and clinical information, and has begun to attract the attention of a large number of doctors and scholars. At present, the mainstream way of computer-aided diagnosis of intracranial hemorrhage is to use deep learning technology to analyze, extract features and predict medical images.
[0003] Existing methods for detecting intracranial hemorrhage have several deficiencies. First, methods using a single CNN model for detection (such as ResNet and DenseNet) have limited feature expression capabilities on intracranial CT images and are difficult to fully capture rich feature information. Due to the diversity of morphology and location of intracranial hemorrhage, more powerful models are needed to extract discriminative features for accurate classification. Second, existing methods often ignore the spatial relationship between slices of intracranial CT images and fail to fully utilize the spatial information between slices. Spatial information between slices is very important for accurate classification and localization of intracranial hemorrhage. In order to improve the accuracy and precision of detection, new methods need to be developed that can effectively utilize the spatial information between slices. Third, there are also limitations in fine-grained classification, and it is difficult to accurately distinguish different subtypes of intracranial hemorrhage. This is crucial for early diagnosis and treatment. In order to improve the effect of fine-grained classification, it is necessary to design a more discriminative model and combine more feature information. Fourth, methods based on Transformer variants need to perform self-attention calculations on all positions when processing sequence data, resulting in high computational complexity. This limits the ability to quickly detect and diagnose intracranial hemorrhage in real-time scenarios. Therefore, it is necessary to find ways to reduce the computational complexity in order to improve the real-time and efficiency of detection. Fifth, the ratio of positive and negative samples in the intracranial CT image dataset is seriously unbalanced, which may lead to insufficient learning of the model for minority categories. In addition, a single hemorrhage CT slice may have multiple lesion types, and the correlation between these lesions is often overlooked. In order to improve the generalization and robustness of the model, it is necessary to adopt appropriate data balancing strategies and association modeling methods to better handle the diversity and complexity in intracranial CT images. Summary of the invention
[0004] The purpose of the present invention is to solve the problem that the current classification of intracranial hemorrhage subtypes relies on the professionalism of doctors, has low accuracy and is time-consuming, and proposes a method for classifying intracranial hemorrhage CT images based on CNN and RNN. In the CNN module, a fine-grained strategy is adopted to accurately capture the subtle features and local information in the intracranial hemorrhage CT images by using larger convolution kernels. This strategy helps the model to better understand the image details and improve the accuracy of classification. At the same time, the RNN module is introduced to process serialized image features, and the RNN is used to capture the timing information of the image sequence, further enhancing the model's ability to understand intracranial hemorrhage CT images. In order to reduce the data dimension and reduce the calculation time, the present method also applies principal component analysis technology. Principal component analysis can reduce the dimensionality of the data, thereby effectively extracting the most representative features, reducing the computational burden and speeding up the classification process.
[0005] In order to achieve the above object, the present invention provides the following technical solution: a CNN and RNN-based intracranial hemorrhage CT image classification method, comprising the following steps:
[0006] Step 1: Collect the intracranial hemorrhage dataset, remove invalid data, and integrate the real label information.
[0007] Step 2: Data preprocessing was performed on the acquired dataset, including converting pixel values into CT values, using brain windows [40, 80], subdural windows [80, 200], and soft tissue windows [40, 380] to enhance image contrast, overlaying channels, separating metadata information, changing image resolution to 512*512, using the Albumentations library for image standardization, converting to PNG format, and removing edge redundant information. The standardized mean was set to 0.456, the standard deviation was set to 0.224, and the maximum pixel value was set to 255. These values are empirical values obtained by calculating and analyzing millions of ImageNet images. The standardized formula is as follows:
[0008]
[0009]
[0010] Where image_standardization is the standardized image, x is the input image, μ is the image mean, adjusted_stdev is the adjusted image standard deviation, σ is the input image standard deviation, and N is the number of pixels in the input image.
[0011] Step 3: Build a network model. The model consists of a CNN module and a RNN module. The CNN module uses the ConvNext model. ConvNext consists of four feature resolution stages. The number of blocks selected in each stage is [3, 3, 9, 3]. The RNN module uses a bidirectional GRU model.
[0012] Step 4: Use the ImageNet dataset to train the ConvNext model to obtain pre-trained weights, construct a binary cross entropy loss function as the optimization target of the model, load the pre-trained weights onto the ConvNext model, use the pre-processed PNG format image as the input data of the ConvNext model, and adjust the image resolution to 224*224 to ensure that it matches the input size of the pre-trained model. By adjusting the weights and parameters of the model during the training process, try to reduce the value of the loss function so that the model's prediction results are as close as possible to the true label. When the loss function reaches the minimum value, obtain the feature values extracted by the ConvNext model and the predicted values for the six categories. The binary cross entropy loss function formula is defined as follows:
[0013] L BCE=-ylog(y′)-(1-y)log(1-y′);
[0014] Where L BCE Represents the loss value, y represents the true value of a certain category, y∈{0,1}, indicating that y can only take the value 0 or 1, y′ represents the probability of the corresponding category predicted by the ConvNext neural network, y′∈[0,1], indicating that y′ ranges from 0 to 1 (including 0 and 1).
[0015] Step 5: Use the principal component analysis method to reduce the dimension of the features extracted by ConvNext, and save the reduced features in Numpy format to prepare as the input of the RNN model.
[0016] Step 6: Read the features and metadata information after dimensionality reduction, sort the metadata information, and load the features after dimensionality reduction according to the sorted metadata information. Train them sequentially through the first and second parts of the bidirectional GRU to finally generate the prediction results.
[0017] Step 7: Based on the prediction results, evaluate the model performance through AUC, accuracy, sensitivity, and specificity.
[0018] Compared with the prior art, the present invention has the following advantages:
[0019] (1) In the preprocessing stage, the present invention enables the model to focus more on key image features by adjusting the window position and width, superimposing the number of channels, and removing black edges, thereby enhancing the robustness and anti-interference ability of the model and reducing the possibility of misclassification.
[0020] (2) The present invention solves the problem of intracranial hemorrhage subtype classification from a fine-grained perspective, using a network architecture that combines ConvNext and bidirectional GRU to fully utilize information within and between slices, thereby improving the expression ability of the neural network and the classification performance.
[0021] (3) The present invention effectively reduces the dimension and computational complexity of features and improves the reasoning speed by introducing the principal component analysis method.
[0022] (4) The experimental results of the present invention show that the AUC of the RSNA Intracranial Hemorrhage Detection public dataset reached 96.75%, 0.76 percentage points ahead of other methods, and the accuracy rate reached 97.71%, 0.2 percentage points higher than other methods. On the basis of ensuring the speed of the algorithm, the accuracy of intracranial hemorrhage subtype classification was significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1This is a flow chart of a method for classifying intracranial hemorrhage CT images based on CNN and RNN in the present invention.
[0024] Figure 2 This is a schematic diagram of the pretreatment operation of step 2 of the present invention.
[0025] Figure 3 This is the network model structure diagram of step 3 of the present invention.
[0026] Figure 4 This is a structural diagram of the ConvNext model of the present invention.
[0027] Figure 5 It is the Block structure diagram of the present invention.
[0028] Figure 6 This is a structural diagram of the bidirectional GRU model of the present invention.
[0029] Figure 7 It is a control experiment result diagram of the present invention.
[0030] Figure 8 This is a diagram of the ablation experiment results of the present invention. DETAILED DESCRIPTION
[0031] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.
[0032] like Figure 1 As shown, the present invention provides a method for classifying intracranial hemorrhage CT images based on CNN and RNN, and the method specifically comprises the following steps:
[0033] Step 1: Prepare the dataset.
[0034] Specifically, step 1 includes the following:
[0035] Step 1-1 Obtain the RSNA Intracranial Hemorrhage Detection public dataset, which contains 750,000 DICOM format files and a CSV file with true labels.
[0036] Step 1-2: Use MicroDicom software to browse DICOM format images and exclude invalid DICOM images.
[0037] Steps 1-3 reintegrate the valid DICOM image tags to generate a CSV file containing six tags: epidural hemorrhage, subdural hemorrhage, intraventricular hemorrhage, intracerebral hemorrhage, subarachnoid hemorrhage, and intracranial hemorrhage.
[0038] Step 2, data preprocessing, such as Figure 2 shown.
[0039] Specifically, step 2 includes the following:
[0040] Step 2-1 Convert the DICOM format image to the CT value. The formula for converting the DICOM format image to the CT value is as follows:
[0041] img=pixel*slope+intercept
[0042] Among them, img is the CT value, also known as HU value or Heinz value; pixel is the pixel value of the DICOM format image; slope is the scaling intercept, which can be obtained from the DICOM format image through tag (0028, 1053); intercept is the scaling slope, which can be obtained from the DICOM format image through tag (0028, 1052).
[0043] Step 2-2 Adjust the window width and window level.
[0044] Since DICOM format images generally occupy 16 bits, the HU value range after conversion is relatively large and the image contrast is poor, so it is necessary to adjust the window width and window position according to needs to enhance the image contrast. The window position refers to the center position of the window, and the window width refers to the displayed HU value range.
[0045] The window widths and window positions used were as follows: brain window [40,80], subdural window [80,200], and soft tissue window [40,380].
[0046] The first number represents the window level value, and the second number represents the window width value.
[0047] Steps 2-3 are superimposed into a three-channel image.
[0048] DICOM images are all grayscale images. In order to simulate real color images, the brain window, subdural window, and soft tissue window are superimposed to simulate RGB three-channel images.
[0049] Steps 2-4 separate the metadata information from the DICOM image.
[0050] In addition to image information, the DICOM format also contains rich metadata information. Making full use of metadata information can effectively improve the performance of the algorithm. The metadata information in the DICOM format images can be separated and saved as a CSV file.
[0051] The metadata information obtained includes Patient ID, Study instance UID, Series InstanceUID, Image Position, Pixel Spacing, Pixel Representation, Window Center, WindowWidth, Rescale Intercept, Rescale Slope, and Slice ID.
[0052] Among them, Patient ID uniquely identifies the patient's identity information; Study Instance UID uniquely identifies a study or examination, usually containing multiple images or sequences; Series Instance UID uniquely identifies a group of related images or sequences, usually obtained in the same study; Image Position represents the coordinate value with the patient's position as the reference; Pixel Spacing represents the pixel spacing in the horizontal and vertical directions; Pixel Representation represents the encoding method and data type of the image pixel value, which can be an unsigned integer or a signed integer; Window Center and WindowWidth represent the window position and window width, which are used to adjust the contrast and brightness of the image display. The window position represents the grayscale level at the center of the window, and the window width represents the width range of the window; Rescale Intercept and Rescale Slope represent the scaling slope and scaling intercept, which are linear conversion parameters used to convert pixel values into actual physical values; Slice ID uniquely identifies each slice in the image sequence, usually associated with the position and order of the image.
[0053] Steps 2-5 change the image resolution size.
[0054] In order to unify the image specifications, the resize function of the PIL library is used to adjust the image resolution to a uniform size of 512*512.
[0055] Steps 2-6: Image normalization.
[0056] Use the Albumentations library to normalize the superimposed images; set the standardized mean to 0.456, the standard deviation to 0.224, and the maximum pixel value to 255; these values are empirical values obtained by calculating and analyzing millions of ImageNet images. The mean and standard deviation are chosen to make the distribution of standardized image data closer to the standard normal distribution, while the maximum pixel value is chosen based on the bit depth of the image, which is usually 8 bits, that is, the pixel value range is 0 to 255. The normalization operation can scale the pixel value range of the image to an appropriate range, which helps to improve the training stability and performance of the model. The normalization formula is as follows:
[0057]
[0058]
[0059] Where image_standardization is the standardized image, x is the input image, μ is the image mean, adjusted_stdev is the adjusted image standard deviation, σ is the input image standard deviation, and N is the number of pixels in the input image.
[0060] Steps 2-7 are converted into PNG format images.
[0061] The DICOM format contains not only image information but also metadata information, so it takes up a lot of storage space. PNG is a lossless image format. After separating the DICOM metadata information, the standardized image information is saved in PNG format, which can effectively reduce storage space.
[0062] Step 2-8 removes the black edges of the image.
[0063] There is a lot of invalid information on the edge of the image. Through the morphological closing operation, that is, first perform an erosion operation on the image, which will expand the black area, and then perform an expansion operation on the eroded image, which will expand the white area and restore the eroded area. Through the combination of erosion and expansion, the closing operation can effectively remove the black edge and retain the main image structure. The morphological closing operation formula is as follows:
[0064]
[0065] Where A is the input image, B is the structural element, represents the corrosion operation, Represents an expansion operation.
[0066] Steps 2-9 divide the data set.
[0067] The processed images are randomly divided into training set, validation set and test set according to the ratio of 7:2:1. The training set is used to train the data, the validation set is used to verify whether the model is overfitting, and the test set is used to evaluate the performance.
[0068] Step 3: Build the network structure, such as Figure 3 shown.
[0069] Specifically, step 3 includes the following:
[0070] Step 3-1: Build the CNN network structure.
[0071] The CNN model used is a pre-trained ConvNext model. The ConvNext model structure is shown in the figure below. Figure 4 shown.
[0072] The ConvNext model first consists of a 4*4 convolutional layer and a LayerNorm layer. Then it consists of 4 feature resolution stages, each of which consists of multiple blocks, and finally a LayerNorm layer and a linear layer.
[0073] The Block consists of a 7*7 depthwise separable convolution, a LayerNorm layer, a 1*1 convolution layer, and a GELU layer. The number of Blocks selected at the four feature resolution stages is [3,3,9,3]. The Block structure is shown in the figure below. Figure 5 shown.
[0074] Step 3-2 builds the RNN network structure.
[0075] The RNN model used is a bidirectional GRU model. The structure of the bidirectional GRU model is as follows: Figure 6 shown.
[0076] The bidirectional GRU model consists of two sequence models. The first part consists of a feature convolution layer, a GRU layer, and a 1*1 convolution layer, and the second part consists of a residual convolution layer, a GRU layer, and a 1*1 convolution layer.
[0077] The feature convolution layer consists of a Dropout layer, a 1*1 convolution layer, a BatchNorm layer, and a ReLU layer, and the residual convolution layer consists of a 1*1 convolution layer, a BatchNorm layer, and a ReLU layer.
[0078] Step 4: Train the CNN network.
[0079] Specifically, step 4 includes the following:
[0080] Step 4-1 constructs the loss function.
[0081] Construct a binary cross entropy loss function, whose formula is defined as follows:
[0082] L BCE =-ylog(y′)-(1-y)log(1-y′)
[0083] Where L BCE Represents the loss value, y represents the true value of a certain category, y∈{0,1}, indicating that y can only take the value 0 or 1; y′ represents the probability of the corresponding category predicted by the ConvNext neural network, y′∈[0,1], indicating that y′ ranges from 0 to 1 (including 0 and 1).
[0084] Step 4-2 Adjust the image resolution.
[0085] In order to use the pre-trained weights, we need to uniformly adjust the resolution of the input image to a size that matches the pre-trained model, and use the resize function of the PIL library to adjust the image resolution to 224*224.
[0086] Step 4-3 uses pre-trained weights.
[0087] Using the transfer learning method, we first train the ConvNext model using the ImageNet 1K dataset to obtain pre-trained weights. ImageNet 1K is a dataset containing more than 1 million images and 1,000 categories. By using pre-trained weights, we can directly obtain the model's ability to recognize certain features, avoiding feature learning from scratch, thereby accelerating the convergence of training. In addition, through transfer learning, we can also learn rich image feature representations, thereby improving the performance and generalization ability of image classification models.
[0088] Step 4-4 trains the ConvNext model.
[0089] Load the pre-trained weights of the ConvNext model, construct a data loader, and load the pre-processed images and label information. Use the images of the training set to learn the model, and then use the images of the validation set to verify the model. Use the loss function to back-propagate the verification results, and then continue learning and verification until the validation set loss converges. After convergence, retain part of the features and pass part of the classifier to predict the results.
[0090] The training uses the stochastic gradient descent method and the Adam optimizer, setting the hyperparameters epoch=200, batchsize=64, the learning rate uses the cosine annealing algorithm, the initial learn rate=0.0005, T max=5, learn ratemin=1e-5, and weight decay=0.00002.
[0091] Among them, epoch represents the number of learning rounds, batch size represents the batch size read at one time, learn rate represents the initial learning rate, T max represents the maximum number of cycles, learn rate min represents the minimum learning rate, and weight decay represents the weight decay factor of L2 regularization.
[0092] Step 5: Use principal component analysis to reduce dimensionality.
[0093] The feature information retained in the previous step is reduced in dimension using principal component analysis (PCA). Through numerical tests, the feature dimension is reduced from 768 to 120. Dimensionality reduction through PCA can reduce storage space and computing costs, remove redundant features, and improve model efficiency and accuracy.
[0094] Then save the features after dimensionality reduction into Numpy files.
[0095] Step 6: Train the RNN network.
[0096] Specifically, step 6 includes the following:
[0097] Step 6-1 Load feature information.
[0098] Read the numpy file saved in the previous step, and load the metadata information separated from the DICOM format image and the prediction results of the ConvNext model.
[0099] Step 6-2 sorts the data information.
[0100] Sort the metadata information from small to large according to the value of Study Instance UID. For the same Study Instance UID value, sort it according to the value of Image Position to restore the spatial position information of the image.
[0101] Step 6-3 trains the first part of the bidirectional GRU model.
[0102] Load the numpy file into the first part of the bidirectional GRU model according to the sorted metadata information for training, and then concatenate the feature vector obtained by training with the prediction result obtained by the ConvNext model.
[0103] The training used five-fold cross validation, epoch=40, batch size=128.
[0104] Step 6-4 trains the second part of the bidirectional GRU model.
[0105] The concatenated feature vector is input into the second part of the bidirectional GRU model for training, and then the trained feature vector is passed through the classifier to obtain the final prediction result.
[0106] Step 7: Evaluate the model performance based on the test results.
[0107] Model evaluation indicators, the formula is as follows:
[0108]
[0109]
[0110]
[0111]
[0112] AUC stands for the area under the ROC curve, and its value range is between 0 and 1. The larger the AUC value, the better the model performance. M stands for the number of positive samples, and N stands for the number of negative samples.
[0113] Accuracy stands for accuracy, which means the proportion of correct predictions to the total number.
[0114] Sensitivity stands for sensitivity, also known as the true positive rate or recall rate. It indicates the proportion of positive samples correctly identified by the model to the total number of actual positive samples. High sensitivity means reducing the risk of missed diagnosis.
[0115] Specificity stands for specificity, which indicates the proportion of negative samples correctly identified by the model to the total number of actual negative samples. High specificity means reducing the risk of missed diagnosis.
[0116] TP means that the actual value is positive but is judged as positive, TN means that the actual value is positive but is judged as negative, FP means that the actual value is negative but is judged as negative, and FN means that the actual value is negative but is judged as positive.
[0117] This method is implemented under the Pytorch framework. In the first stage, batchsize=64, workers=28, epoch=200, and 677460 brain CT images are used as the data set. The learning rate for the first 80 epochs is 0.0005, and the learning rate is unified to 1e-5 thereafter. In the second stage, batchsize=128, workers=8, epoch=40, and five-fold cross validation is used. No overfitting occurs. This experiment uses a 14-core Intel Xeon(R) Gold 6330CPU and an NVIDIARTX3090(24G) graphics card for accelerated training.
[0118] like Figure 7As shown, it is shown that the method proposed in the present invention has achieved significant improvement effect in the field of intelligent medical impact processing.
[0119] like Figure 8 As shown, it is demonstrated that the module proposed by the present invention is effective.
[0120] The baseline indicates the use of the ConvNext model, the pre-trained weights indicate the use of pre-trained weights for fine-tuning, the three-channel superposition indicates the use of a 3-channel window for superposition, the morphological operation indicates the use of a morphological closing operation to remove invalid information, the recurrent network indicates the use of a bidirectional GRU model, and the principal component analysis indicates the use of principal component analysis for dimensionality reduction.
Claims
1. A classification method for intracranial hemorrhage CT images based on CNN and RNN, Its characteristics include the following steps: Step 1: Collect the intracranial hemorrhage dataset, remove invalid data, and integrate the real label information; Step 2: Data preprocessing was performed on the acquired dataset, including converting pixel values into CT values, using brain windows [40, 80], subdural windows [80, 200], and soft tissue windows [40, 380] to enhance image contrast, overlaying channels, separating metadata information, changing image resolution to 512*512, using the Albumentations library for image standardization, converting to PNG format, and removing edge redundant information. The standardized mean was set to 0.456, the standard deviation was set to 0.224, and the maximum pixel value was set to 255. These values are empirical values obtained by calculating and analyzing millions of ImageNet images. The standardized formula is as follows: Where image_standardization is the standardized image, x is the input image, μ is the image mean, adjusted_stdev is the adjusted image standard deviation, σ is the input image standard deviation, and N is the number of pixels in the input image; Step 3: Build a network model, which consists of a CNN module and an RNN module. The CNN module uses the ConvNext model, which consists of four feature resolution stages. The number of blocks selected in each stage is [3, 3, 9, 3]. The RNN module uses a bidirectional GRU model. Step 4: Use the ImageNet dataset to train the ConvNext model to obtain pre-trained weights, construct a binary cross entropy loss function as the optimization target of the model, load the pre-trained weights onto the ConvNext model, use the pre-processed PNG format image as the input data of the ConvNext model, and adjust the image resolution to 224*224 to ensure that it matches the input size of the pre-trained model. By adjusting the weights and parameters of the model during the training process, try to reduce the value of the loss function so that the model's prediction results are as close as possible to the true label. When the loss function reaches the minimum value, obtain the feature values extracted by the ConvNext model and the predicted values for the six categories. The binary cross entropy loss function formula is defined as follows: L BCE (-ylog(y′)-(1-y)log(1-y′)); Where L BCE represents the loss value, y represents the true value of a certain category, y∈{0,1}, indicating that y can only take the value 0 or 1, y′ represents the probability of the corresponding category predicted by the ConvNext neural network, y′∈[0,1], indicating that y′ ranges from 0 to 1 (including 0 and 1); Step 5: Use the principal component analysis method to reduce the dimension of the features extracted by ConvNext, and save the reduced features in Numpy format to prepare as the input of the RNN model; Step 6: Read the features and metadata information after dimensionality reduction, sort the metadata information, and load the features after dimensionality reduction according to the sorted metadata information, train them sequentially through the first part and the second part of the bidirectional GRU, and finally generate the prediction results; Step 7: Based on the prediction results, evaluate the model performance through AUC, accuracy, sensitivity, and specificity.
2. The intracranial hemorrhage CT image classification method based on CNN and RNN according to claim 1 is characterized in that The metadata information separated in step 2 includes Patient ID, Study instance UID, Series Instance UID, Image Position, Pixel Spacing, Pixel Representation, Window Center, Window Width, Rescale Intercept, Rescale Slope, and Slice ID; Among them, Patient ID uniquely identifies the patient's identity information; Study Instance UID uniquely identifies a study or examination; Series Instance UID uniquely identifies a group of related images or sequences; Image Position represents the coordinate value with the patient's position as the reference; Pixel Spacing represents the pixel spacing in the horizontal and vertical directions; PixelRepresentation represents the encoding method and data type of the image pixel value; Window Center and Window Width represent the window position and window width, the window position represents the grayscale level of the window center, and the window width represents the width range of the window; RescaleIntercept and Rescale Slope represent the scaling slope and scaling intercept; Slice ID uniquely identifies each slice in the image sequence.
3. The intracranial hemorrhage CT image classification method based on CNN and RNN according to claim 1 is characterized in that In step 5, principal component analysis is used to reduce the dimension of the extracted features. The dimension before the dimension reduction is 768, and the dimension after the dimension reduction is 120.
