Hemorrhagic stroke microwave image classification method based on multi-input neural network

Abstract

The invention belongs to the technical field of medical imaging, and relates to a cerebral hemorrhagic stroke microwave image classification method based on a multi-input neural network, which comprises the following steps: obtaining a microwave brain image and corresponding time domain waveform data, inputting the microwave brain image and the time domain waveform data into an image classification model, and obtaining a cerebral hemorrhagic stroke classification result; the image classification model is obtained by training a multi-input neural network through a training set, and a weighted cross entropy loss function is introduced in the training process; the image classification model adopts a dual-channel feature extraction network architecture to perform image feature extraction and time domain waveform feature extraction on the microwave brain image and the time domain waveform data respectively, and adopts channel dimension splicing to obtain an hemorrhagic cerebral apoplexy classification result. According to the method, multi-modal data feature extraction is realized through a multi-input neural network architecture, the network training stability is ensured by using residual connection, and the classification capability of hemorrhagic cerebral apoplexy is effectively improved by combining long and short term memory network modeling waveform time sequence features.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and in particular to a microwave image classification method for hemorrhagic stroke based on a multi-input neural network. Background Technology

[0002] Stroke is a common and serious traumatic disease with high mortality and disability rates. According to stroke treatment guidelines, it is necessary to complete the diagnosis of the injured person as soon as possible, determine the extent of the injury, and provide surgical treatment before brain herniation occurs.

[0003] The diagnosis of stroke mainly relies on medical imaging equipment. The imaging time, the length of the time interval between diagnosis and treatment and injury are directly related to the patient's life and postoperative recovery ability.

[0004] With the development of medical imaging technology, microwave imaging, as a non-invasive detection method, has gradually attracted researchers' attention due to its advantages such as speed, no radiation, and ease of on-site application. However, existing microwave imaging technology still faces three major challenges: First, due to the physical characteristics of microwaves, its spatial resolution is low, making it difficult to clearly identify tiny hemorrhages; second, the complex anatomical structure of the brain makes microwave signals susceptible to interference, resulting in a low signal-to-noise ratio and increasing the difficulty of feature extraction; and finally, traditional single-modal analysis methods cannot fully utilize the complementary information between microwave images and time-domain waveform data. Summary of the Invention

[0005] To address the three major challenges in current microwave brain imaging—insufficient spatial resolution, low temporal signal-to-noise ratio, and complex multimodal correlations—this invention proposes a microwave image classification method for hemorrhagic stroke based on a multi-input neural network. A dual-channel deep fusion network is designed, where the image channel extracts spatial gradient features of the dielectric constant distribution through a residual convolutional network to overcome resolution limitations; the waveform channel utilizes five-level convolutional compression and LSTM temporal modeling to separate hemorrhagic characteristic frequency bands from noise; and the feature fusion layer employs channel stitching and hierarchical decision mechanisms to establish a nonlinear mapping relationship between the image and the waveform. This design significantly improves the classification ability and noise robustness of microwave diagnostic systems for hemorrhagic stroke lesions, enabling more efficient assistance to physicians in classifying image data.

[0006] To achieve the above objectives, the present invention provides the following solution: A microwave image classification method for hemorrhagic stroke based on multi-input neural networks includes: A microwave brain image and its corresponding time-domain waveform data are acquired. The microwave brain image and the time-domain waveform data are then input into an image classification model to obtain a classification result for hemorrhagic stroke. The image classification model is obtained by training a multi-input neural network using a training set, and a weighted cross-entropy loss function is introduced during the training process. The training set includes: the original microwave brain image and its corresponding original time-domain waveform data. The image classification model employs a dual-channel feature extraction network architecture to extract image features and time-domain waveform features from the microwave brain image and the time-domain waveform data, respectively, and uses channel-dimensional concatenation to obtain the classification result of the hemorrhagic stroke.

[0007] Optionally, training the multi-input neural network using the training set includes: The training samples in the training set are preprocessed; the preprocessing includes: standardization and dynamic data augmentation, and the original time-domain waveform data in the training samples are transformed separately, that is, all spatial channels are expanded according to the time dimension, so as to train the multi-input neural network using the preprocessed training samples. The dynamic data enhancement processing includes random horizontal flipping, vertical flipping, brightness and contrast adjustment, and random rotation.

[0008] Optionally, the image classification model includes: The image feature extraction module is used to extract image features from the microwave brain image and obtain an image feature vector; The waveform feature extraction module is used to extract time-domain waveform features from the time-domain waveform data and obtain a time-domain waveform feature vector. The multimodal feature fusion module is used to concatenate the image feature vector and the temporal waveform feature vector in the channel dimension to form a joint feature vector; The classification decision module is used to reduce the dimensionality of the joint feature vector, generate a two-dimensional probability distribution, and normalize it using the Softmax function to obtain the classification result of the hemorrhagic stroke.

[0009] Optionally, the image feature extraction module includes: The image feature extraction submodule uses a ResNet18 model as the backbone network, removes the fully connected layers in the backbone network, and retains the target layer convolutional modules to construct a deep residual structure. The deep residual structure includes multiple residual units, each residual unit consisting of multiple consecutive convolutional layers. Batch normalization layers and ReLU activation functions are inserted between the convolutional layers, and the input features are fused with the convolutional output features through skip connections at the end of the residual unit. A deactivation layer is added after the output layer of the backbone network, and global average pooling is used to compress the output feature map of the output layer to the first target dimension to generate the image feature vector.

[0010] Optionally, the backbone network includes: ; in, For convolution kernel weights, This is the output feature map of the residual module. For the residual function, This is the input feature map for the residual module. To modify the activation function of the linear unit, This is a batch normalization operation.

[0011] Optionally, generating the image feature vector includes: ; in, This provides global information for the entire image feature map. It is the spatial location of the feature map The channel vector at position 1, i.e., the first... line, number All channel values ​​of the column, This represents the height of the feature map, i.e., the number of rows in the feature map. This represents the width of the feature map, i.e., the number of columns in the feature map.

[0012] Optionally, the waveform feature extraction module includes: The waveform feature extraction submodule is used to input the time-domain waveform data into a five-level one-dimensional convolutional neural network. The time-domain waveform data is received through different channels in each level of the one-dimensional convolutional neural network. Local features are extracted from the received time-domain waveform data using convolutional kernels of target size. After batch normalization and ReLU activation, the signal length is compressed to the second target dimension based on max pooling to generate high-level features. Deep temporal modeling is performed on the unidirectional LSTM layer with stacked high-level feature inputs. The output feature map of the LSTM layer is then input into a multi-layer fully connected layer to reduce the dimensionality to the third target dimension, thereby generating the image feature vector.

[0013] Optionally, performing local feature extraction on the received time-domain waveform data using a convolutional kernel of the target size includes: ; in, For the first Level convolution operation, For the first Level convolution operation, This represents a one-dimensional convolution operation. For convolution kernel parameters, For max pooling operation, For the first Bias term of multi-level convolution.

[0014] Optionally, deep temporal modeling of the unidirectional LSTM layer with stacked high-level feature inputs includes: ; in, This is the output of the fifth-level convolution. Let be the hidden state vector at time step t. Let be the cell state vector at time step t. These are the parameters for the input gate, forget gate, and output gate.

[0015] Optionally, the classification decision module includes: The classification decision submodule is used to reduce the dimensionality of the joint feature vector, and after ReLU activation and Dropout regularization, further compress the joint feature vector to a fourth objective dimension. Based on the compression of the joint feature vector under the fourth objective dimension, a two-dimensional probability distribution is generated and normalized by the Softmax function to obtain the classification result of hemorrhagic stroke. ; in, For the discard rate, This is the output of the first layer, with a dimension of 256. The second layer output has a dimension of 128. The layer is randomly deactivated with a dropout rate of 50%. This is the weight matrix of the first fully connected layer. The input feature vector after fusion. For the bias term of the first layer fully connected layer, This is the weight matrix for the second fully connected layer. It is the second-level bias term. This is the output layer weight matrix. It is the output layer bias term.

[0016] The beneficial effects of this invention are as follows: This invention achieves efficient classification processing of microwave images by integrating the spatial features of microwave images and the dynamic features of time-domain waveforms, combined with deep residual networks and temporal modeling techniques. This effectively reduces the workload of doctors in the image data classification process, improves the efficiency and accuracy of image data classification, provides technical support for data analysis of portable microwave detection instruments for craniocerebral injuries, and helps to process relevant image data more efficiently.

[0017] This invention achieves accurate and rapid classification of microwave images of hemorrhagic stroke through multimodal feature fusion and dynamic optimization strategies. Based on a dual-channel deep neural network architecture, it deeply integrates the spatial distribution features of microwave images with the dynamic evolution of temporal signals, enabling classification decisions to be made directly without relying on traditional iterative scattering inversion algorithms, significantly reducing computational complexity. Experimental results demonstrate excellent classification performance. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the classification network model in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, this embodiment discloses a microwave image classification method for hemorrhagic stroke based on a multi-input neural network, including: acquiring microwave brain images and corresponding time-domain waveform data; inputting the microwave brain images and time-domain waveform data into an image classification model to obtain hemorrhagic stroke classification results; the image classification model is obtained by training a multi-input neural network using a training set, and a weighted cross-entropy loss function is introduced during the training process; the training set includes: the original microwave brain images and the corresponding original time-domain waveform data; the image classification model adopts a dual-channel feature extraction network architecture to extract image features and time-domain waveform features from the microwave brain images and time-domain waveform data respectively, and uses channel-dimensional concatenation to obtain hemorrhagic stroke classification results.

[0023] Furthermore, training a multi-input neural network using the training set includes: preprocessing the training samples in the training set; the preprocessing includes: standardization and dynamic data augmentation, and separately transforming the original time-domain waveform data in the training samples, that is, expanding all spatial channels according to the time dimension, so as to train the multi-input neural network using the preprocessed training samples; wherein, the dynamic data augmentation includes: random horizontal flipping, vertical flipping, brightness and contrast adjustment, and random rotation.

[0024] Specifically, microwave brain images and corresponding temporal waveform data are acquired. Multimodal data preprocessing is performed, and the dataset is divided into training and testing sets. Specifically, the original microwave images are standardized to eliminate equipment noise interference. Dynamic data augmentation strategies are employed, including random horizontal and vertical flipping, brightness and contrast adjustment, and random rotation, to improve the model's generalization ability. The temporal waveform data undergoes a format transformation, unfolding all spatial channels along the time dimension, enabling the one-dimensional convolutional network to extract features more effectively on the time axis. The preprocessed microwave brain image (244×244 size) and temporal waveform data (64×144 size) are input into a multi-input neural network (MIN). The MPN is a dual-channel feature extraction network architecture. The image feature extraction channel uses a ResNet convolutional neural network as its basic architecture, while the temporal waveform feature extraction channel consists of one-dimensional convolutions and cascaded unidirectional long short-term memory (LSTM) modules following the convolutional layers.

[0025] like Figure 2 As shown, the image classification model includes: an image feature extraction module for extracting image features from microwave brain images to obtain image feature vectors; a waveform feature extraction module for extracting time-domain waveform features from time-domain waveform data to obtain time-domain waveform feature vectors; a multimodal feature fusion module for concatenating image feature vectors and time-domain waveform feature vectors along the channel dimension to form a joint feature vector; and a classification decision module for reducing the dimensionality of the joint feature vector, generating a two-dimensional probability distribution, and normalizing it using the Softmax function to obtain the classification result for hemorrhagic stroke.

[0026] Furthermore, the image feature extraction module includes: an image feature extraction submodule, which uses a ResNet18 model as the backbone network, removes the fully connected layers in the backbone network, and retains the target layer convolutional modules to construct a deep residual structure; the deep residual structure includes: multiple residual units, each residual unit consisting of multiple consecutive convolutional layers, with batch normalization layers and ReLU activation functions inserted between the convolutional layers, and the input features are fused with the convolutional output features at the end of the residual unit through skip connections; a deactivation layer is added after the output layer of the backbone network, and global average pooling is used to compress the output feature map of the output layer to the first target dimension to generate an image feature vector.

[0027] Specifically, in the multi-input neural network, the image feature extraction channel uses a pre-trained ResNet18 model as the backbone network. By removing its fully connected layers and retaining the first 17 convolutional modules, a deep residual structure is constructed. This structure contains multiple residual modules, each consisting of two consecutive 3×3 convolutional layers. Batch Normalization (BN) layers and ReLU activation functions are inserted between the convolutional layers. The input features are transferred to the end of the module via skip connections. With convolution output Adding them together effectively alleviates the vanishing gradient problem. Its mathematical expression is as follows: ; in For convolution kernel weights, This is the output feature map of the residual module. For the residual function, This is the input feature map for the residual module. To modify the activation function of the linear unit, This is a batch normalization operation.

[0028] Step 202: Add a random deactivation layer after the ResNet18 output layer, and use global average pooling to extract the feature maps. Compressed into a 512-dimensional vector : ; in, It is a feature map output by ResNet18. This provides global information for the entire image feature map. It is the spatial location of the feature map The channel vector at position 1, i.e., the first... line, number All channel values ​​of the column, This represents the height of the feature map, i.e., the number of rows in the feature map. The width of the feature map is the number of columns in the feature map.

[0029] This design preserves the spatial distribution characteristics of hemorrhage areas in microwave images, such as anomalous gradient changes in tissue dielectric constant. Through transfer learning from a pre-trained model, this design significantly improves feature generalization ability.

[0030] Furthermore, the waveform feature extraction module includes: a waveform feature extraction submodule, which is used to input time-domain waveform data into a five-level one-dimensional convolutional neural network, receive time-domain waveform data through different channel numbers in each level of the one-dimensional convolutional neural network, perform local feature extraction on the received time-domain waveform data using convolutional kernels of target size, and after batch normalization and ReLU activation, compress the signal length to the second target dimension based on max pooling to generate high-level features; perform deep temporal modeling on the stacked unidirectional LSTM layers of high-level features, and input the output feature map of the LSTM layer into a multi-layer fully connected layer to reduce the dimension to the third target dimension to generate image feature vectors.

[0031] Specifically, the time-domain waveform feature extraction channel employs a hybrid architecture combining a five-level one-dimensional convolutional neural network and a unidirectional LSTM. The first level processes the input 64 channels and acquires values ​​at 144 time points. Further, the continuous signal is discretized on the time axis, resulting in 64×144 time-domain waveform data. Local feature extraction is performed using a 5-size convolutional kernel. Level convolution operation Represented as: ; in, For the first Level convolution operation, For the first Level convolution operation, This represents a one-dimensional convolution operation. For max pooling operation, For the first The bias term of a convolutional layer. The convolution kernel parameters are padded with 2 to maintain the length, and the number of channels in the five-level channel is: .

[0032] After batch normalization and ReLU activation, the signal length is compressed to 72 using max pooling with a stride of 2. The second and third stages increase the number of channels to 32 and 64 respectively, and the second and third stages increase the number of channels to 128 and 256 respectively. The convolution-pooling operation is repeated, and finally, a 256-dimensional high-level feature with 4 time steps is output.

[0033] More specifically, in a five-level one-dimensional convolutional neural network, the connection relationship of each level is achieved through repeated operations of convolution (Conv) → batch normalization (BN) → ReLU activation → max pooling (MaxPool), which gradually compresses the signal length and increases the number of channels, and finally outputs high-level features of 4 time steps × 256 dimensions.

[0034] The aforementioned high-level features are then fed into stacked unidirectional LSTM layers for deep temporal modeling: ; in, It is the output of the fifth-level convolution. It is the hidden state vector at time step t. It is the cell state vector at time step t. Includes parameters for the input gate, forget gate, and output gate; the final state is then determined. Characterizes the dynamic mode of the waveform.

[0035] The LSTM output is then reduced to a 64-dimensional vector through two fully connected layers: ; in, .

[0036] Furthermore, multimodal feature fusion and classification decision-making concatenates microwave image feature vectors and time-domain waveform data feature vectors along the channel dimension, then inputs them into a fully connected layer for classification, and finally outputs whether hemorrhagic stroke exists.

[0037] The multimodal feature fusion stage integrates image feature vectors With time-domain waveform feature vector By concatenating the features along the channel dimension, a 576-dimensional joint feature vector is formed: ; This fusion strategy fully leverages the accuracy of image spatial features in locating hemorrhage areas, as well as the sensitivity of waveform temporal features to subtle pathological changes.

[0038] Furthermore, the classification decision module includes a classification decision submodule, which uses dimensionality reduction of the joint feature vector, followed by ReLU activation and Dropout regularization, to further compress the joint feature vector to a fourth objective dimension. Based on the compression of the joint feature vector under the fourth objective dimension, a two-dimensional probability distribution is generated and normalized using the Softmax function to obtain the classification result of hemorrhagic stroke. Specifically, the fused features are used for classification decisions through a three-layer fully connected network: the first layer reduces the dimension to 256 and applies ReLU activation and Dropout regularization; the second layer further compresses it to 128 dimensions; and the final output layer generates a 2-dimensional probability distribution and normalizes it using the Softmax function. The process is as follows: ; in, , For the discard rate, This is the output of the first layer, with a dimension of 256. The second layer output has a dimension of 128. The layer is randomly deactivated with a dropout rate of 50%. This is the weight matrix of the first fully connected layer. The input feature vector after fusion (dimension 576). For the bias term of the first layer fully connected layer, This is the weight matrix for the second fully connected layer. It is the second-level bias term. This is the output layer weight matrix. This is the output layer bias term. The random deactivation mechanism between fully connected layers (dropout rate of 0.5) forces the network to learn redundant feature representations, enhancing the model's robustness to noise interference.

[0039] The loss function uses weighted cross-entropy: ; in, It is the weighted cross-entropy loss value. It is the total number of samples (batch size), that is, the number of samples used in the loss calculation. It is the sample index (i=1,2,…,N) i =1,2,…, N ). No. The true label of each sample (takes a value of 0 or 1 in binary classification). The model predicts the first The probability that a sample belongs to category 1 ( . No. The class weights of each sample (based on the true label) (Dynamically adjusted): If The weights are positive class weights. ;like The weight is the negative class weight. .

[0040] Because the experimental data from microwave stroke detection instruments is limited and insufficient for model training, this invention is implemented on a microwave simulation dataset. The dataset, constructed around intracranial hemorrhage detection, contains three category folders: normal without hematoma, one hematoma, and two hematomas. Each category folder contains two types of data: 100×100 pixel BMP format training microwave brain images and .npy format raw time-domain waveform data. Specifically, the two-hematoma dataset contains 808 data sets, while the one-hematoma and normal without-hematoma datasets each contain 2000 data sets. The data acquisition process is as follows: First, an elliptical brain model with a major axis of 17 cm and a minor axis of 13 cm is constructed, and a circular hemorrhage region with a diameter of 2-4 cm is embedded within it to simulate different hemorrhage scenarios. The imaging frequency is set to the optimal brain imaging band of 0.5-3 GHz, and the time-domain signal-to-noise ratio is maintained at 5 dB for all experimental groups. A uniform circular array consisting of 12 antennas surrounds the brain model, and simulation and imaging experiments are conducted for brain models with and without hemorrhage, with 20 simulations performed for each configuration.

[0041] Regarding the model process, the number of network iterations (epochs) used for training was set to 50, the batch size to 16, Adam was used as the optimizer, the initial learning rate was set to 0.0001, and the dropout rate was set to 0.5.

[0042] To verify the effectiveness of the present invention, Table 1 compares the present invention with existing technologies using accuracy, recall, precision, F1 score, area under the receiver operating characteristic curve (ROC-AUC), and area under the precision-recall curve (PRC-AUC). As shown in Table 1, the present invention achieves better results on multiple metrics, thus demonstrating its effectiveness.

[0043]

[0044] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A microwave image classification method for hemorrhagic stroke based on a multi-input neural network, characterized in that, include: Acquire microwave brain images and corresponding time-domain waveform data, input the microwave brain images and the time-domain waveform data into an image classification model, and obtain the classification results of hemorrhagic stroke; The image classification model is obtained by training a multi-input neural network using a training set, and a weighted cross-entropy loss function is introduced during the training process; the training set includes: original microwave brain images and corresponding original time-domain waveform data; The image classification model employs a dual-channel feature extraction network architecture to extract image features and time-domain waveform features from the microwave brain image and the time-domain waveform data, respectively, and uses channel-dimensional concatenation to obtain the classification result of the hemorrhagic stroke.

2. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 1, characterized in that, Training the multi-input neural network using the training set includes: The training samples in the training set are preprocessed; the preprocessing includes: standardization and dynamic data augmentation, and the original time-domain waveform data in the training samples are transformed separately, that is, all spatial channels are expanded according to the time dimension, so as to train the multi-input neural network using the preprocessed training samples. The dynamic data enhancement processing includes random horizontal flipping, vertical flipping, brightness and contrast adjustment, and random rotation.

3. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 1, characterized in that, The image classification model includes: The image feature extraction module is used to extract image features from the microwave brain image and obtain an image feature vector; The waveform feature extraction module is used to extract time-domain waveform features from the time-domain waveform data and obtain a time-domain waveform feature vector. The multimodal feature fusion module is used to concatenate the image feature vector and the temporal waveform feature vector in the channel dimension to form a joint feature vector; The classification decision module is used to reduce the dimensionality of the joint feature vector, generate a two-dimensional probability distribution, and normalize it using the Softmax function to obtain the classification result of the hemorrhagic stroke.

4. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 3, characterized in that, The image feature extraction module includes: The image feature extraction submodule uses a ResNet18 model as the backbone network, removes the fully connected layers in the backbone network, and retains the target layer convolutional modules to construct a deep residual structure. The deep residual structure includes multiple residual units, each residual unit consisting of multiple consecutive convolutional layers. Batch normalization layers and ReLU activation functions are inserted between the convolutional layers, and the input features are fused with the convolutional output features through skip connections at the end of the residual unit. A deactivation layer is added after the output layer of the backbone network, and global average pooling is used to compress the output feature map of the output layer to the first target dimension to generate the image feature vector.

5. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 4, characterized in that, The backbone network includes: ； in, For convolution kernel weights, This is the output feature map of the residual module. For the residual function, This is the input feature map for the residual module. To modify the activation function of the linear unit, This is a batch normalization operation.

6. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 4, characterized in that, Generating the image feature vector includes: ； in, This provides global information for the entire image feature map. It is the spatial location of the feature map The channel vector at position 1, i.e., the first... line, number All channel values ​​of the column, This represents the height of the feature map, i.e., the number of rows in the feature map. This represents the width of the feature map, i.e., the number of columns in the feature map.

7. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 3, characterized in that, The waveform feature extraction module includes: The waveform feature extraction submodule is used to input the time-domain waveform data into a five-level one-dimensional convolutional neural network. The time-domain waveform data is received through different channels in each level of the one-dimensional convolutional neural network. Local features are extracted from the received time-domain waveform data using convolutional kernels of target size. After batch normalization and ReLU activation, the signal length is compressed to the second target dimension based on max pooling to generate high-level features. Deep temporal modeling is performed on the unidirectional LSTM layer with stacked high-level feature inputs. The output feature map of the LSTM layer is then input into a multi-layer fully connected layer to reduce the dimensionality to the third target dimension, thereby generating the image feature vector.

8. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 7, characterized in that, Local feature extraction of the received time-domain waveform data using a convolutional kernel of the target size includes: ； in, For the first Level convolution operation, For the first Level convolution operation, This represents a one-dimensional convolution operation. For convolution kernel parameters, For max pooling operation, For the first Bias term of multi-level convolution.

9. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 7, characterized in that, Deep temporal modeling of the unidirectional LSTM layer with stacked high-level feature inputs includes: ； in, This is the output of the fifth-level convolution. Let be the hidden state vector at time step t. Let be the cell state vector at time step t. These are the parameters for the input gate, forget gate, and output gate.

10. The microwave image classification method for hemorrhagic stroke based on a multi-input neural network according to claim 3, characterized in that, The classification decision module includes: The classification decision submodule is used to reduce the dimensionality of the joint feature vector, and after ReLU activation and Dropout regularization, further compress the joint feature vector to a fourth objective dimension. Based on the compression of the joint feature vector under the fourth objective dimension, a two-dimensional probability distribution is generated and normalized by the Softmax function to obtain the classification result of hemorrhagic stroke. ； in, For the discard rate, This is the output of the first layer, with a dimension of 256. The second layer output has a dimension of 128. The layer is randomly deactivated with a dropout rate of 50%. This is the weight matrix of the first fully connected layer. The input feature vector after fusion. For the bias term of the first layer fully connected layer, This is the weight matrix for the second fully connected layer. It is the second-level bias term. This is the output layer weight matrix. It is the output layer bias term.

Images