Cerebral hemorrhage microwave detection method based on deep learning
By using the VITS-UNet3plus model to process microwave data in human cerebral hemorrhage microwave detection, the problems of detection accuracy and inefficiency in the prior art are solved, and more efficient and accurate human cerebral hemorrhage detection is achieved.
Patent Information
- Application Number
- CN202510067828.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-16
- Publication Date
- 2025-05-27
AI Technical Summary
The prior art has problems of accuracy and inefficiency in microwave detection of human cerebral hemorrhage, especially imaging artifacts and slow imaging speed.
The HFSS simulation system is used to establish a three-dimensional human cerebral hemorrhage model, and the microwave data is processed and analyzed through the VITS-UNet3plus model in deep learning, and feature information is extracted to achieve the detection of human cerebral hemorrhage.
The accuracy and efficiency of human cerebral hemorrhage detection is improved, and the imaging artifacts and slow imaging speed in traditional microwave imaging methods are avoided.
Smart Images
Figure CN120048512A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical fields of deep learning and microwave imaging, and particularly relates to a microwave detection method for cerebral hemorrhage based on deep learning. Background Art
[0002] Deep neural network is an important technology in the field of machine learning. Its characteristic is to construct a multi-layer neural network model to process and analyze complex data by simulating the structure and function of the human brain neuron network. At present, the application of deep neural network in medical imaging has achieved remarkable results. Its powerful data processing and learning ability enables the neural network to perform tasks such as feature extraction, image classification, detection, and segmentation on medical image data, thereby improving the accuracy and efficiency of diagnosis and treatment.
[0003] The combination of deep neural network and medical microwave imaging is bringing new breakthroughs to the field of medical imaging. Medical microwave imaging technology is a technology that uses microwave signals to scan the human body and reconstruct internal images of the human body by receiving the reflected signals. Compared with traditional imaging technologies such as X-rays and CT, microwave imaging has remarkable characteristics and advantages: non-invasive: Microwave imaging technology can obtain internal images without performing invasive operations on the human body, reducing the pain and risk of patients; radiation-free: Microwave imaging technology does not use radiation, so it will not cause radiation damage to the human body, improving safety; strong real-time performance: Microwave imaging technology can obtain real-time internal image information of the human body, helping doctors make timely diagnosis and treatment decisions; high resolution: Microwave imaging technology has high spatial resolution and temperature resolution, and can clearly display the fine structure and temperature changes inside the human body. Therefore, using the powerful learning ability of deep neural network to perform more refined processing and analysis on microwave imaging data, thereby improving the accuracy and reliability of imaging, has become an innovative idea and exploration direction in medical microwave imaging research. Summary of the Invention
[0004] The purpose of the present invention is to solve the problems in the prior art, and a microwave detection method for cerebral hemorrhage based on deep learning is proposed.
[0005] A microwave detection method for cerebral hemorrhage based on deep learning includes the following steps:
[0006] S1. Use the HFSS simulation system to perform microwave simulation on the established three-dimensional human cerebral hemorrhage model to obtain microwave data, take screenshots of the x-y plane of the three-dimensional model, and label the bleeding positions in the figure to obtain label data, and establish a microwave imaging dataset;
[0007] S2. Delete the row names of the first column of the microwave data, then convert it into a tensor format, perform normalization processing, adjust the size of the label image, and perform central cropping on it to ensure that the label images are of the same size and have the same number of channels;
[0008] S3. Divide the samples into two parts: a training set and a test set;
[0009] S4. Use a deep neural network optimization algorithm to extract feature information from the microwave data, make the generated image fit the label image. Based on the UNet3plus model, build the VITS-UNet3plus model, and use the training set and the test set to optimize and debug the model parameters of the built VITS-UNet3plus model;
[0010] S5. Save the optimal model structure and model parameters;
[0011] S6. Input the microwave data with different bleeding positions and bleeding sizes in the test set into the model structure in step S5 for forward propagation to obtain a predicted grayscale image of the human cerebral hemorrhage situation. The white part of the grayscale image is the bleeding block;
[0012] S7. According to the distance of the white bleeding block part in the grayscale image from the center of the image and the corresponding scale, judge the coordinate information and size range of the bleeding block.
[0013] Preferably, in step S1, establishing the microwave imaging dataset includes:
[0014] Construct a three-dimensional human cerebral hemorrhage model through the HFSS simulation system, and adjust the antenna array to perform microwave simulations on the model from two directions of zero degree and thirty degrees to obtain two sets of simulation data, and subtract the two sets of simulation data to obtain microwave data;
[0015] The label data obtained by taking screenshots of the x-y plane of the three-dimensional model and using the labelme data annotation package to annotate the bleeding positions of the images.
[0016] Preferably, in step S2, the tensor format of the microwave data after data processing is [56, 101], the size of the label data is uniformly 64*64, and the number of channels is 1. The normalization formula in data processing is as follows:
[0017] data=(tensor–tensor_min) / (tensor_max - tensor_min)
[0018] where tensor is the data of each dimension in the tensor data, tensor_min is the minimum value in the whole data, tensor_max is the maximum value in the whole data, and data is the input data.
[0019] Preferably, in the step S3, there are 594 sets of data in the training set and 30 sets of data in the test set.
[0020] Preferably, in the step S4, the deep neural network optimization algorithm adopted is the RMSprop optimization algorithm, and the formula of the RMSprop optimization algorithm is as follows:
[0021] s dw = βs dw +(1 - β)dW 2
[0022] s db = βs db +(1 - β)db 2
[0023]
[0024] The RMSProp algorithm uses the weighted average of the squared differentials of the gradients of the weight W and the bias b. Among them, s dw and s db are the gradient momenta accumulated in the previous t - 1 rounds of iteration of the loss function respectively, and β is an exponent for gradient accumulation.
[0025] Preferably, in the step S4, the VITS - UNet3plus model is built based on the UNet3plus model by adding the SE attention mechanism and the Vision transformer module. The last convolutional layer and pooling layer in the encoder of the UNet3plus model are replaced by the Vision transformer module and the upsampling module respectively, and the SE module is added after the pooling layer and after the upsampling module in the encoder, with a total of five layers of SE modules added.
[0026] Preferably, in the step S4, the BCELoss loss function is used, and the learning rate and the EPOCHS value are set to 0.00001 and 500 respectively. Given a set of predicted values y and the actual labels The formula of BCELoss is as follows:
[0027]
[0028] Among them, N is the number of samples, y i is the actual label of the i - th sample, is the predicted value of the i - th sample.
[0029] Compared with the existing technologies, the advantages of the present invention are as follows:
[0030] 1. The present invention uses the HFSS simulation system to establish a three-dimensional model of human cerebral hemorrhage, and uses the VITS-UNet3plus model in deep learning to sample the microwave data obtained by simulation, extract effective feature information from a large amount of microwave data, so as to effectively judge the relative position of human cerebral hemorrhage and the size of the hemorrhage mass, realize microwave detection of human cerebral hemorrhage, improve the accuracy and efficiency of human cerebral hemorrhage detection, and avoid the problems of imaging artifacts and slow imaging speed in traditional microwave imaging methods.
[0031] 2. The present invention uses the VITS-UNet3plus neural network to abandon the adversarial generation method, avoid mode collapse caused by performance differences or overly intense confrontation between the generator and the discriminator, has a faster training speed, fewer parameters, and the computing resources are concentrated in the generation part, resulting in better generation accuracy. Description of the Drawings
[0032] Figure 1 is the flowchart of the microwave detection method for cerebral hemorrhage in the present invention.
[0033] Figure 2 is the three-dimensional simulation model and antenna array for establishing the above-mentioned dataset in the present invention.
[0034] Figure 3 is the x-y plane screenshot of the three-dimensional model and the grayscale map of data annotation in the present invention.
[0035] Figure 4 is the VITS-UNet3plus neural network model proposed by the present invention.
[0036] Figure 5 is the Vision Transformer (VIT) module of the model proposed by the present invention.
[0037] Figure 6 is the prediction result of the model on the test set and the label data of the test set in the present invention. Detailed Embodiments
[0038] To make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments.
[0039] Referring to Figure 1 as shown, a microwave detection method for cerebral hemorrhage based on deep learning is characterized in that it includes the following steps:
[0040] S1. Use the HFSS simulation system to perform microwave simulation on the established three-dimensional human cerebral hemorrhage model to obtain microwave data, take a screenshot of the x-y plane of the three-dimensional model and perform data annotation on the bleeding position in the figure to obtain label data, and establish a microwave imaging dataset;
[0041] S2. Delete the row names of the first column of the microwave data, then convert it into a tensor format, perform normalization processing, adjust the label image size, and perform central cropping on it to ensure that the label images are of the same size and have the same number of channels;
[0042] S3. Divide the samples into two parts: a training set and a test set to train and test the performance of the model;
[0043] S4. Use a deep neural network optimization algorithm to extract feature information from the microwave data, make the generated images fit the label images. Based on the UNet3plus model, build the VITS-UNet3plus model, and use the training set and the test set to optimize and debug the model parameters of the built VITS-UNet3plus model;
[0044] S5. Save the optimal model structure and model parameters;
[0045] S6. Input the microwave data with different bleeding positions and bleeding sizes in the test set into the model structure in step S5 for forward propagation to obtain a predicted grayscale image of the human cerebral hemorrhage situation. The white part of the grayscale image is the bleeding block;
[0046] S7. According to the distance of the white bleeding block part in the grayscale image relative to the center of the image and the corresponding scale, judge the coordinate information and size range of the bleeding block.
[0047] In the above step S1, establishing the microwave imaging dataset includes:
[0048] Construct a three-dimensional human cerebral hemorrhage model through the HFSS simulation system, and adjust the antenna array to perform microwave simulations on the model from two directions of zero degree and thirty degrees to obtain two sets of simulation data, and subtract the two sets of simulation data to obtain microwave data;
[0049] Label data obtained by taking screenshots of the x-y plane of the three-dimensional model and using the labelme data annotation package to annotate the bleeding positions of the images.
[0050] In the above step S2, the tensor format of the microwave data after data processing is [56, 101], the size of the label data is uniformly 64*64, and the number of channels is 1. The normalization formula in data processing is as follows:
[0051] data=(tensor–tensor_min) / (tensor_max - tensor_min)
[0052] Among them, tensor is the data of each dimension in the tensor data, tensor_min is the minimum value in the entire data, tensor_max is the maximum value in the entire data, and data is the input data.
[0053] In the step S3, there are 594 groups of data in the training set and 30 data in the test set.
[0054] In the step S4, the deep neural network optimization algorithm adopted is the RMSprop optimization algorithm. The formula of the RMSprop optimization algorithm is as follows:
[0055] s dw =βs dw +(1 - β)dW 2
[0056] s db =βs db +(1 - β)db 2
[0057]
[0058] The RMSProp algorithm uses the weighted average of the squared differentials for the gradients of the weight W and the bias b. Among them, s dw and s db are the gradient momenta accumulated in the previous t - 1 rounds of iteration of the loss function respectively. β is an exponent for gradient accumulation. The difference is that the RMSProp algorithm calculates the weighted average of the squared differentials for the gradients. This approach is beneficial for eliminating the directions with large swing amplitudes, used to correct the swing amplitudes, making the swing amplitudes of each dimension smaller. On the other hand, it also makes the network function converge faster. For example, when one of the values of dW or db is relatively large, then when we update the weight or the bias, we divide it by the square root of the previously accumulated gradient, so that the update amplitude can be made smaller. To prevent the denominator from being zero, a very small value ε is used for smoothing, generally taking the value of 1e - 8.
[0059] In step S4, the VITS-UNet3plus model is constructed by adding the SE attention mechanism and the Vision transformer module to the UNet3plus model. The last convolutional layer and pooling layer in the encoder of the UNet3plus model are replaced with the Vision transformer module and the upsampling module respectively, and the SE module is added after the pooling layer and after the upsampling module in the encoder. A total of five SE modules are added. By adding the SE attention mechanism and the Vision transformer module to the UNet3plus model, the problems of incorrect target position and artifacts in the generated image are prevented. Compared with the traditional generative neural network DCGAN, the VITS-UNet3plus neural network has the following main advantages: abandoning the adversarial generation method, avoiding mode collapse caused by performance differences between the generator and the discriminator or overly intense confrontation; faster training speed and fewer parameters; the computing resources are concentrated in the generation part, and the generation accuracy is better.
[0060] In step S4, the BCELoss loss function is used, and the learning rate and EPOCHS values are set to 0.00001 and 500 respectively, and a set of predicted values y and actual labels are given The formula of BCELoss is as follows:
[0061]
[0062] where N is the number of samples, y i is the actual label of the i-th sample, is the predicted value of the i-th sample.
[0063] Embodiment
[0064] Step 1: Establishment of the dataset: Use the HFSS simulation system to construct a three-dimensional simulation model of human cerebral hemorrhage, as Figure 2 shown.
[0065] The 3D model is mainly divided into several modules: plasma, skin, skull, and tissue. Eight antenna arrays are set on the periphery of the model. The size of the innermost tissue module in the model is a sphere with a radius of 66 mm. To simulate the actual situation of human cerebral hemorrhage and enhance data diversity, the central coordinates of the hemorrhage block with random x, y, and z coordinates within plus or minus 30 are generated, and the hemorrhage range with a random radius between 0 - 9 mm is generated, ensuring that the hemorrhage range is within the tissue module. Microwave simulation experiments are carried out on the hemorrhage conditions with different coordinates and radii through the HFSS simulation system. The antenna arrays are adjusted to perform microwave simulations on the model from two directions: zero degree and thirty degrees to obtain two sets of simulation data, and the two sets of simulation data are subtracted to obtain microwave data. The obtained microwave experimental data is exported as the input data of the model, and the x - y plane screenshot of the corresponding 3D model is used as the label data of the model. As Figure 3 shown, as can be seen from Figure 3 a in, the label data is in color, and it includes both the human brain tissue part and the hemorrhage block part. In reality, the size and orientation of the human brain tissue part are unchanged, only the orientation and size of the hemorrhage block are changing. To exclude the noise in the label data and the parts that are definitely unchanged in the 3D model, the labelme data annotation tool is used to annotate the label pictures to obtain grayscale pictures, and the color pictures are discarded. The grayscale pictures are used as the label data, which enables the model to only fit the data pictures of a single channel, reducing the features that the model needs to fit. Figure 3 The grayscale image corresponding to a in is shown as b.
[0066] Step 2: Before training the neural network model, the microwave data will be normalized. The row names of the first column of the microwave data are deleted, then it is converted into a tensor format and normalized. The size of the label pictures is adjusted, and they are centrally cropped to ensure that the label pictures are of the same size and the number of channels is the same. Most of the microwave data is below the order of magnitude of 1e - 2 and has large differences. After normalization, the data has similar scales, which helps optimization algorithms such as gradient descent converge to the optimal solution faster. When the data scale differences are large, the optimizer may take a longer time to find the optimal solution because different features have different degrees of influence on the loss function. Normalization helps reduce the sensitivity of the model to the initial weights, making the model more stable. By adjusting the data scale, normalization can make different features have similar contributions in the model, thus improving the accuracy of the model. Normalization can reduce the amount of calculation because the normalized data has a smaller numerical range, which can reduce the error of floating - point operations and the calculation time.
[0067] Step 3: The dataset used for training the neural network contains a total of 624 groups of data, which are divided into a training dataset and a test dataset. 594 groups of data are used in the training dataset, and the test dataset contains 30 groups of data.
[0068] Step 4: Use the RMSprop optimizer to train the VITS-UNet3plus network model. The RMSprop optimizer, whose full name is Root Mean Square Propagation, is a mean square root propagation algorithm and an optimizer for neural network training. It is an improved version of the Adagrad optimizer, aiming to solve the problem of the learning rate dropping too fast in Adagrad.
[0069] Based on the UNet3plus neural network model, the VITS-UNet3plus neural network model is established. The Vision Transformer module and the SE attention mechanism are added to the UNet3plus model to form the VITS-UNet3plus model. The specific structure diagram of the above model is as Figure 4 shown. Since the format of the microwave simulation data obtained through the HFSS simulation system is a csv file with 57 columns and 101 rows, to ensure that the tensor format output by the model is [batchsize, 1, 64, 64], a fully connected layer is added before the first layer of ConvBlock. During the process of processing the data into the tensor format, the first column that is the same in all csv files is deleted, and the data with 56 columns and 101 rows is converted into tensor data with the format of [batchsize, 5656]. After the data is input, it first passes through the fully connected layer to obtain tensor data with the format of [batchsize, 4096], and then through tensor transformation operations to obtain tensor data with the format of [batchsize, 1, 64, 64], making the sampling process of subsequent modules more convenient. The structure diagram of the Vision Transformer module is as Figure 5 shown in a of the figure. The Blocks and MLPBlocks included in this module are respectively as Figure 5 shown in b of the figure and Figure 5 shown in c of the figure.
[0070] Use the training set and the test set to optimize and debug the model parameters of the built VITS-UNet3plus model. Use the BCELoss loss function, and set the learning rate and the EPOCHS value to 0.00001 and 500 respectively.
[0071] Step 5: Save the optimal model structure and model parameters.
[0072] Step 6: Input the microwave data with different bleeding positions and bleeding sizes in the test set into the model structure in Step 5 for forward propagation, and obtain the comparison between the predicted grayscale image of the human cerebral hemorrhage situation and the label image, as Figure 6 shown. The white part of the grayscale image is the bleeding block.
[0073] Step 7: Determine the bleeding area. Based on the distance of the white bleeding block part in the grayscale image from the center of the picture and the corresponding scale, determine the coordinate information and size range of the bleeding block.
[0074] As can be seen Figure 6 from the test results shown, the imaging accuracy of the VITS-UNet3plus model is relatively high, the images formed are similar to the label data, and it has corresponding reference value for judging the condition of human cerebral hemorrhage.
[0075] In summary, the present invention proposes to use the VITS-UNet3plus neural network model for microwave imaging, which improves the accuracy and efficiency of human cerebral hemorrhage detection compared with traditional microwave imaging, and avoids the problems of imaging artifacts and slow imaging speed of traditional microwave imaging methods.
[0076] As is known by common technical knowledge, the present invention can be implemented by other embodiments that do not depart from its spiritual essence or essential features. Therefore, the above-disclosed embodiments are illustrative in all aspects and not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are encompassed by the present invention.
Claims
1. A microwave detection method for cerebral hemorrhage based on deep learning, characterized by: The following steps are involved: S1. Use the HFSS simulation system to perform microwave simulation on the established three-dimensional human brain hemorrhage model to obtain microwave data, take a screenshot of the xy plane of the three-dimensional model and annotate the bleeding position in the image to obtain label data, and establish a microwave imaging data set; S2. Delete the row name of the first column of the microwave data, convert it into a tensor format, normalize it, adjust the size of the label image, and perform center cropping on it to ensure that the label images have the same size and the same number of channels; S3, divide the samples into two parts: training set and test set; S4. Use a deep neural network optimization algorithm to extract feature information from microwave data, fit the generated image to the label image, build a VITS-UNet3plus model based on the UNet3plus model, and use the training set and test set to optimize and debug the model parameters of the built VITS-UNet3plus model; S5, saving the optimal model structure and model parameters; S6, inputting the microwave data of different bleeding locations and bleeding sizes in the test set into the model structure of step S5 for forward propagation, and obtaining a predicted grayscale image of human brain bleeding, wherein the white part of the grayscale image is the bleeding clot; S7. According to the distance of the white bleeding block part in the grayscale image relative to the center of the image and the corresponding scale, the coordinate information and size range of the bleeding block are determined.
2. The method for microwave detection of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: In step S1, establishing a microwave imaging data set includes: A three-dimensional human brain hemorrhage model was constructed using the HFSS simulation system, and the antenna array was adjusted to perform microwave simulation on the model from two directions, zero degrees and thirty degrees, to obtain two sets of simulation data, which were then subtracted to obtain microwave data. The label data is obtained by taking a screenshot of the xy plane of the 3D model and using the labelme data annotation package to annotate the bleeding position of the image.
3. The method for microwave detection of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: In step S2, the format of the microwave data tensor after data processing is [56, 101], the label data size is unified to 64*64, the number of channels is 1, and the normalization formula in data processing is as follows: data=(tensor–tensor_min) / (tensor_max-tensor_min) Among them, tensor is the data of each dimension in the tensor data, tensor_min is the minimum value of the entire data, tensor max is the maximum value of the entire data, and data is the input data.
4. The method for microwave detection of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: In step S3, the training set has 594 sets of data and the test set has 30 sets of data.
5. The method for microwave detection of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: In step S4, the deep neural network optimization algorithm used is the RMSprop optimization algorithm, and the formula of the RMSprop optimization algorithm is as follows: s dw =βs dw +(1-β)dW 2 s db =βs db +(1-β)db 2 The RMSProp algorithm uses a differential square weighted average of the gradients of weight W and bias b, where s dw and db are the accumulated gradient momentum of the loss function during the first t-1 rounds of iterations, and β is an index of gradient accumulation.
6. The method for microwave detection of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: In step S4, the VITS-UNet3plus model is built based on the UNet3plus model by adding the SE attention mechanism and the Vision transformer module. The last convolutional layer and the pooling layer in the UNet3plus model encoder are replaced with the Vision transformer module and the upsampling module, respectively, and the SE module is added after the pooling layer and the upsampling module in the encoder, and a total of five layers of SE modules are added.
7. The method for microwave detection of cerebral hemorrhage based on deep learning according to claim 1, characterized in that: In step S4, the BCELoss loss function is used, and the learning rate and EPOCHS value are set to 0.00001 and 500 respectively. Given a set of predicted values y and actual labels The formula for BCELoss is as follows: Where N is the number of samples, y i is the actual label of the i-th sample, is the predicted value of the ith sample.