Microwave thermal therapy ultrasonic noninvasive temperature measurement system based on RNN-ARMB model

Abstract

The invention relates to the technical field of non-invasive temperature measurement, in particular to a microwave hyperthermia ultrasonic non-invasive temperature measurement system based on an RNN-ARMB model, and the system comprises a microwave hyperthermia instrument. The waveguide antenna is connected with the output end of the microwave hyperthermia instrument, the microwave hyperthermia instrument feeds energy into the waveguide antenna, and the waveguide antenna generates heat; the B-type ultrasonic equipment is used for acquiring an ultrasonic image video; the computer is connected with the B-type ultrasonic equipment and used for receiving the ultrasonic image video collected by the B-type ultrasonic equipment and converting the ultrasonic image video into ultrasonic image data; the temperature is input into the trained RNN-ARMB model, and a predicted temperature is obtained; the system can synchronously predict the temperature at the moment while acquiring an ultrasonic image.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of non-invasive temperature measurement, and particularly relates to a microwave hyperthermia ultrasonic non-invasive temperature measurement system based on an RNN-ARMB model. BACKGROUND

[0002] Tumor hyperthermia technology, as a safe and efficient new tumor treatment method, has developed rapidly in recent years; tumor hyperthermia can be divided into traditional hyperthermia and thermal ablation; tumor hyperthermia takes advantage of the fact that tumor tissue is more sensitive to temperature than normal tissue, that is, human normal cells will not be damaged at 42.5-43 DEG C, but most tumor cells will be induced to apoptosis at this temperature.

[0003] At present, the commonly used in clinical is still a lossy temperature measurement technology, that is, the temperature measurement probe is placed inside the biological tissue for measurement; this method can obtain relatively accurate temperature information, but it will cause pain and trauma to the patient and may cause metastasis of cancer cells; therefore, the present application proposes a non-invasive temperature measurement technology. SUMMARY

[0004] In order to solve the problems in the prior art, the present application provides a microwave hyperthermia ultrasonic non-invasive temperature measurement system based on an RNN-ARMB model, which aims to non-invasively measure the temperature during tumor hyperthermia.

[0005] The present application aims to provide a microwave hyperthermia ultrasonic non-invasive temperature measurement system based on an RNN-ARMB model, comprising:

[0006] a microwave hyperthermia instrument;

[0007] a waveguide antenna, the waveguide antenna being connected to the output end of the microwave hyperthermia instrument, the microwave hyperthermia instrument feeding energy to the waveguide antenna, and the waveguide antenna generating heat;

[0008] a B-mode ultrasonic device, the B-mode ultrasonic device being used to collect ultrasonic video images;

[0009] a computer, the computer being connected to the B-mode ultrasonic device and being used to receive the ultrasonic video images collected by the B-mode ultrasonic device and convert them into ultrasonic image data; and predict the temperature.

[0010] The computer is provided with,

[0011] a conversion module configured to convert the ultrasonic video images into ultrasonic image data to be detected;

[0012] a preprocessing processing module configured to perform frame average filtering and median filtering on the ultrasonic image data to be detected to obtain filtered ultrasonic image data;

[0013] an acquisition module configured to acquire the ultrasonic image data to be detected;

[0014] The preprocessing processing module is configured to perform frame average filtering and median filtering on the to-be-detected ultrasound image data, and obtain filtered ultrasound image data.

[0015] The region of interest positioning module is configured to position the filtered ultrasound image data, and obtain image data after positioning of the region of interest.

[0016] The extraction module is configured to obtain a gray mean value, a gray standard deviation, an average gradient and a hybrid entropy of the image data after positioning of the region of interest.

[0017] The signal-to-noise ratio extraction module is configured to obtain a signal-to-noise ratio of the image data after positioning of the region of interest.

[0018] The k parameter calculation module is configured to obtain an improved k parameter of the image data after positioning of the region of interest by means of an improved homodyne K distribution model.

[0019] The prediction module is configured to input the improved k parameter, the gray mean value, the gray standard deviation, the average gradient and the hybrid entropy of the image data after positioning of the region of interest into a trained RNN-ARMB model, and obtain a predicted temperature.

[0020] The application achieves the following beneficial effects:

[0021] (1) The microwave hyperthermia ultrasound non-invasive temperature measurement system based on the RNN-ARMB model can simultaneously collect ultrasound images and predict the temperature at the moment.

[0022] (2) The application estimates the value of the Gamma function by means of the Stirling approximation formula, which is a relatively simple function, so that the calculation process can be significantly simplified, and the error can be reduced; in the calculation of the k parameter of the homodyne K model, the error of the Stirling formula gradually decreases with the increase of n, so that the error can be ignored when z is large enough.

[0023] (3) The application proposes a special RNN-ARMB model, which performs excellently in terms of mean square error and prediction accuracy, and each module makes an important contribution to the overall performance of the model, and their synergistic effect enables the RNN-ARMB model to achieve superior performance in a given task. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is a microwave hyperthermia ultrasound non-invasive temperature measurement system diagram of the RNN-ARMB model of the application.

[0025] Figure 2is the accuracy comparison chart of the Gamma function value of the application and the approximate formula replacement value.

[0026] Figure 3 is the linear fitting relationship chart of the improved k parameter (left side) and the improved k parameter (right side) and temperature of the application.

[0027] Figure 4 is the RNN-ARMB model structure diagram of the application.

[0028] Figure 5 is the logical block diagram of the Attention module of the application.

[0029] Figure 6 is the logical block diagram of the residual connection module of the application.

[0030] Figure 7 is the ultrasound image after filtering processing of the application.

[0031] Figure 8 is the ultrasound image positioning diagram of the application.

[0032] Figure 9 is the fitting relationship chart between each parameter and temperature of the application.

[0033] Figure 10 is the RNN-ARMB model training loss chart of the application.

[0034] Figure 11 is the RNN-ARMB model temperature actual value prediction value comparison chart of the application.

[0035] Figure 12 is the RNN-ARMB model temperature prediction error bar chart of the application.

[0036] Figure 13 is the microwave hyperthermia ultrasound non-invasive temperature measurement system of the RNN-ARMB model of the application. DETAILED DESCRIPTION

[0037] In order to facilitate those skilled in the art to understand the application, the specific embodiments of the application will be described below in conjunction with the drawings.

[0038] As shown in Figure 1 , the application proposes a microwave hyperthermia ultrasound non-invasive temperature measurement system based on RNN-ARMB model, which comprises

[0039] a microwave hyperthermia instrument, the microwave hyperthermia instrument adopts the MH-IY model microwave hyperthermia instrument manufactured by Beijing Muxiuyu Electronics Co., Ltd., the output power of the hyperthermia instrument is adjustable, the range is 0-60W; under the working frequency of 2.45GHz, the peak gain reaches 2.2dB;

[0040] The waveguide antenna is connected with the output end of the microwave hyperthermia instrument, the microwave hyperthermia instrument feeds energy to the waveguide antenna, and the waveguide antenna generates heat to perform hyperthermia on the tumor; the waveguide antenna adopts a compact metamaterial-filled waveguide aperture hyperthermia antenna, the size of which is 10mm*17.4mm, and the peak gain reaches 2.2dB at a working frequency of 2.45GHz;

[0041] The B-mode ultrasound device is used to collect an ultrasound video, and the internal parameter setting of the B-mode ultrasound device is 30 frames per second, that is, there are 30 pictures per second; the B-mode ultrasound device adopts a Mindray M9Vet portable ultrasound diagnostic instrument produced by the Mindray Company; and the ultrasound probe is a linear array L20-5s ultrasound probe which is adapted to the ultrasound measurement system and is suitable for small organs, has 192 array elements, and has a sampling depth of 5cm.

[0042] The computer is connected with the B-mode ultrasound device and is used to receive the ultrasound video collected by the B-mode ultrasound device and convert the ultrasound video into ultrasound image data; the ultrasound image data is input into the trained RNN-ARMB model to obtain a predicted temperature; the operating system of the computer is Windows 11, the processor is Intel(R)Core(TM)i7-13700H, the CPU is @2.40GHz, the memory is 64GB, the graphics card is NVIDIA GeForce RTX 4060, the graphics card memory is 40GB, and the parallel computing environment is CUDA 11.8.

[0043] The computer is connected with the B-mode ultrasound device and is used to receive the ultrasound video collected by the B-mode ultrasound device and convert the ultrasound video into ultrasound image data; the operating system of the computer is Windows 11, the processor is Intel(R)Core(TM)i7-13700H, the CPU is @2.40GHz, the memory is 64GB, the graphics card is NVIDIA GeForce RTX 4060, the graphics card memory is 40GB, and the parallel computing environment is CUDA 11.8.

[0044] The conversion module is configured to convert the ultrasound video into ultrasound image data to be detected.

[0045] The preprocessing processing module is configured to perform frame average filtering and median filtering on the ultrasound image data to be detected to obtain filtered ultrasound image data.

[0046] The region of interest positioning module is configured to position the filtered ultrasound image data to obtain image data after region of interest positioning.

[0047] The specific method for positioning the filtered ultrasound image data is as follows: taking the position of the temperature probe as the center, a square region with a side length of 256 pixels is defined as the image after region of interest positioning.

[0048] The extraction module is configured to obtain the gray mean value, gray standard deviation, average gradient and hybrid entropy of the image data after region of interest positioning.

[0049] The gray mean value, the gray standard deviation, the average gradient and the mixed entropy are calculated on the ultrasound image by the formula through the matlab; the specific calculation formula is as follows:

[0050] The gray mean value:

[0051] Wherein, M and N are the row number and the column number of the image respectively, and f(i,j) is the gray value of the image at the coordinate (i,j);

[0052] The gray standard deviation:

[0053] Wherein, M and N are the row number and the column number of the image respectively, and f(i,j) is the gray value of the image at the coordinate (i,j);

[0054] The average gradient:

[0055] Wherein, M and N are the row number and the column number of the image respectively, And Are the gradients of the image in the x and y directions respectively;

[0056] The mixed entropy:

[0057] Wherein, p ijk Is the probability of the simultaneous occurrence of the gray values i,j,k.

[0058] The signal-to-noise ratio extraction module is configured to obtain the signal-to-noise ratio of the image data after the positioning of the region of interest;

[0059] The signal-to-noise ratio is obtained by the following method:

[0060] Step 501, the envelope amplitude A of ultrasonic backscattering is calculated by using Hilbert transform;

[0061] The definition of Hilbert transform is the convolution of signal x(t) and the impulse response h(t) of the system:

[0062]

[0063] Wherein, Indicates a Hilbert transform in the time domain;

[0064] The analytical signal is constructed:

[0065] The envelope amplitude is calculated:

[0066] Step 502, the signal-to-noise ratio is calculated by using the envelope amplitude through the following formula:

[0067]

[0068] where P signal is the signal power, P noise is the noise power, and is calculated as follows:

[0069]

[0070] where μ E is the mean of the signal envelope amplitude, and is calculated as follows:

[0071] The k parameter calculation module is configured to obtain an improved k parameter by using an improved homodyne K distribution model on the signal-to-noise ratio of the image data after positioning the region of interest.

[0072] The improved homodyne K distribution model is a model improved on the basis of the original homodyne K distribution model.

[0073] The k parameter calculation formula before improvement is as follows:

[0074]

[0075] where Rv is the signal-to-noise ratio, μ is the number of effective scattering sub-particles, v is a positive real number, Γ(z) is the Gamma function, pFq(a1,…,ap;b1,…,bq;c) represents the generalized hypergeometric series, csc(·) is the cotangent function, and η is defined as:

[0076] The improved homodyne K distribution model is represented by the following formula:

[0077] Since the change amount of the signal-to-noise ratio Rv is small, the signal-to-noise ratio Rv is first amplified by a scaling factor,

[0078] The scaling factor calculation formula is as follows:

[0079] The new signal-to-noise ratio R` calculation formula is as follows: R`=(R-R min )*β (4)

[0080] R v in formula (1) is equal to R` in formula (4).

[0081] The Gamma function substitution formula is as follows:

[0082] That is, the improved homodyne K distribution model is calculated by substituting formula (3) and formula (4) into formula (1), and substituting formula (5) into formula (2) and then into formula (1).

[0083] The zero-difference K distribution model is a statistical model widely used in ultrasonic signal processing, which describes the statistical characteristics of the ultrasonic echo envelope. It is also the statistical distribution model of the ultrasonic backscattering signal envelope with the most physical meaning, and the main parameters include k (the ratio of coherent scattering to diffuse scattering) and μ (the effective number of scattering subgroups). The RSK method in the parameter estimation algorithm of the zero-difference K distribution model uses the signal-to-noise ratio, skewness and kurtosis of the envelope amplitude A to estimate the parameters of the zero-difference K distribution model. The present application uses the signal-to-noise ratio R to calculate the k parameter and the μ parameter, and the relationship between the signal-to-noise ratio R and k, μ is shown in formula (1).

[0084] After obtaining the estimated value of the signal-to-noise ratio R, the values of the two parameters k and μ can be obtained by using the formula. The Bessel function in the formula contains the Gamma function. Since the expression of the Gamma function in the frequency domain is in the form of integration, and it shows rapid exponential growth in the complex plane and is multivalued, for large factorial numbers, direct calculation of the Gamma function value may encounter numerical overflow or underflow problems, thereby causing certain errors in the calculation process. The present application estimates the value of the Gamma function by using the Stirling approximation formula. The Stirling approximation formula is a relatively simple function, which can significantly simplify the calculation process and reduce errors. In the calculation of the k parameter of the zero-difference K model, the error of the Stirling formula will gradually decrease with the increase of n, so that when z is large enough, the error can be ignored.

[0085] As shown in Figure 2 , when the value of z is large, the value of the Gamma function appears underflow and cannot be calculated. The spouge formula replacement will appear negative values at large z, so it cannot be accurately replaced by the gamma function. At this time, the Stirling approximation formula shows its advantages and can calculate the corresponding value of the gamma function.

[0086] Therefore, in the research of the zero-difference K distribution parameters, after the replacement of the Stirling formula and the scaling of the signal-to-noise ratio R', Figure 3 The figure shows the linear fitting relationship between the improved k parameter (left side) and the temperature before improvement and the improved k parameter (right side) and the temperature after improvement.

[0087] The prediction module is configured to input the improved k parameter, the gray mean value, the gray standard deviation, the average gradient and the hybrid entropy of the image data after positioning the region of interest into the trained RNN-ARMB model to obtain the predicted temperature.

[0088] As shown in Figure 4 , the RNN-ARMB model includes:

[0089] RNN module: The RNN module uses a standard RNN layer with an input feature dimension of 5, a hidden layer unit number of 10, a single-layer structure, and a batch_first setting of True to match the shape of the input data.

[0090] Attention mechanism module: The attention mechanism module is set between the hidden layer of the RNN module and the output layer of the RNN module, used to weight the output of the RNN and extract key information;

[0091] Given the excellent performance of the Attention mechanism in capturing key information in sequence data, we integrated an attention mechanism module after the hidden layer of the RNN network; the simple block diagram of the module is shown in Figure 5 The attention mechanism module can adaptively focus on the features most critical to temperature prediction by calculating the dynamic weights between each parameter and the temperature, thereby improving the model's sensitivity to important information. The implementation of the attention mechanism module not only enhances the model's ability to understand complex relationships between parameters, but also significantly improves the accuracy of temperature prediction.

[0092] Memory unit module: Set after the output layer of the RNN module, a learnable memory matrix of size 100x10 is introduced to interact with the RNN output through the attention mechanism;

[0093] Considering the importance of time dependence in time series data, we integrated a memory module after the final output layer of the RNN network; this module can capture and store long-term dependencies in the data, allowing the model to fully utilize historical information when making predictions. By introducing a memory mechanism, we expect the model to better understand and predict the trend of parameter changes over time, thereby improving the stability of temperature prediction.

[0094] Residual connection module: Set after the output layer of the RNN module, used to add the input of the RNN transformed by a linear layer to the output of the last time step of the RNN to enhance the model's ability to retain input information;

[0095] To alleviate the common gradient vanishing problem in deep neural networks, we introduced a residual connection module after the final output layer of the RNN model; the block diagram of the residual module is shown in Figure 6 By establishing a direct connection between adjacent layers, residual connection can ensure that gradients flow more smoothly during backpropagation, thereby facilitating the effective training of deep networks. This improvement not only increases the convergence speed of the model, but also significantly enhances its generalization ability.

[0096] Batch normalization module: Set after the output layer of the RNN module to speed up the training process and improve the stability of the model;

[0097] To improve the stability and efficiency of model training, we added a batch normalization module after the final output layer of the RNN network. Batch normalization can reduce internal covariate shift by normalizing input data, thereby accelerating the convergence process of the model and reducing the risk of overfitting. This improvement not only simplifies the model tuning process, but also significantly improves the final performance of the model.

[0098] Full connection layer module: located after the residual connection module, batch normalization module, and memory unit module, combining RNN output, residual connection results, and memory information, the full connection layer outputs the final temperature prediction value.

[0099] The trained RNN-ARMB model is constructed by the following method:

[0100] The trained RNN-ARMB model is constructed by the following method:

[0101] S1, obtain the ultrasound video, and convert the video into training ultrasound images by frame extraction;

[0102] S2, preprocess the training ultrasound image data to obtain the filtered ultrasound image;

[0103] The preprocessing method is as follows: in the time dimension, the frame average filtering method is used for the training ultrasound image data, and the average value of each pixel point of all frame images obtained per second is calculated; in the spatial dimension, the median filtering technique is used for the training ultrasound image data; the filtered ultrasound image data is obtained; Figure 7 A is the original image before filtering, Figure 7 B is the filtered ultrasound image.

[0104] We have preprocessed the ultrasound image, that is, filtered it, to reduce the influence of noise interference in the B-mode ultrasound image on the calculation result; first, in the time dimension, we use the frame average filtering method to calculate the average value of each pixel point of all frame images obtained per second; this step not only ensures the frequency synchronization of the ultrasound image data and the thermocouple temperature measurement data, but also effectively reduces the computational complexity in the subsequent image processing process, improving the processing efficiency; subsequently, in the spatial dimension, we apply the median filtering technique; as a nonlinear signal processing method, median filtering can effectively suppress high-frequency noise in the ultrasound image while preserving the edge details of the image, avoiding image distortion.

[0105] In the actual experiment data collection, this step takes the operation of frame by frame extraction of video, get 30 pictures per second, frame average filtering is to get only one picture per second, if the median filtering, will be filtered one by one for 30 per second, computer workload increases, first using frame average filtering method, then using median filtering technology to obtain the filtered image, this processing sequence will only operate on one picture per second, reduce the amount of calculation.

[0106] S3, positioning the filtered ultrasound image, obtaining the image of the positioning of the region of interest;

[0107] The specific operation is: taking the position of the temperature probe as the center, a square region with a side length of 256 pixels is defined as the image of the positioning of the region of interest.

[0108] In order to further improve the accuracy of temperature measurement and reduce the range of image processing, we implemented the positioning of the region of interest (ROI), Figure 7 C is the image of the positioning of the region of interest; through the subsequent analysis and processing of this specific area, we can focus more accurately on the image information directly related to temperature measurement.

[0109] Figure 8 The left side is the matlab interface of the positioning operation actual operation, in the picture moving mouse, in the center of the cross, select the center point, select the center point as the center, select the square of 128 up and 128 down and 128 left and 128 right, such as Figure 8 The right side, that is, a square with a side length of 256 pixels.

[0110] S4, obtaining the gray mean value, energy, correlation, gray standard deviation, contrast, average gradient, gray entropy and hybrid entropy of the image of the positioning of the region of interest; through matlab, the formula is calculated on the ultrasound image;

[0111] The specific calculation formula is as follows:

[0112] Gray mean value:

[0113] Where M and N are the number of rows and columns of the image, and f(i,j) is the gray value of the image at coordinate (i,j);

[0114] Energy:

[0115] Where M and N are the number of rows and columns of the image, and f(i,j) is the gray value of the image at coordinate (i,j);

[0116] Correlation:

[0117] where f(i,j) and g(i,j) are the gray values of the image at coordinates (i,j), and are the mean values of f and g, respectively;

[0118] Gray standard deviation:

[0119] Contrast:

[0120] where p(i,j) is the probability of the simultaneous occurrence of gray values i and j in the image;

[0121] Average gradient:

[0122] where and are the gradients of the image in the x and y directions, respectively;

[0123] Gray entropy:

[0124] where L is the number of gray levels, and pi is the probability of the occurrence of gray value i;

[0125] Mixed entropy:

[0126] where pijk is the probability of the simultaneous occurrence of gray values i, j, and k.

[0127] S5, screening out parameters with high temperature correlation coefficients;

[0128] To explore the relationship between image parameters and temperature, eight characteristic parameters were selected from the gray level co-occurrence matrix and the gray gradient co-occurrence matrix, including the gray mean (A), energy (B), correlation (C), gray standard deviation (D), contrast (E), average gradient (F), gray entropy (G), and mixed entropy (H). These parameters were applied to the preprocessed B-mode ultrasound images to extract relevant information. After completing the parameter extraction, each parameter was linearly fitted with the temperature data, and the corresponding scatter plot and fitting curve were obtained, as shown in FIG. 1. By comparing the fitting effects of each parameter, it was found that the gray mean, gray standard deviation, average gradient, and mixed entropy had a high correlation with the temperature. Figure 9 Figure 9 Figure 9 Figure 9 Figure 9 Figure 9 Figure 9 Figure 9 H). Figure 9

[0129] S6, obtaining the signal-to-noise ratio of the image after the region of interest is positioned,​​​​​​​​

[0130] S7, the signal-to-noise ratio of the image data after positioning the region of interest is obtained by the improved homodyne K distribution model, and the improved k parameter is obtained;

[0131] S8, the five parameters of the improved k parameter, the gray mean value, the gray standard deviation, the average gradient and the mixed entropy are input into the RNN-ARMB model for training, and the trained RNN-ARMB model is obtained.

[0132] The experimental verification and analysis are as follows:

[0133] Model training hyperparameter setting:

[0134] Set the learning rate: 0.001; optimizer: RMSprop; loss function: mean square error (MSE); training rounds: 5000. Model training is carried out on GPU to speed up the calculation process; training data is divided into training set and test set by data loading part, and the training set contains 10 groups of data totaling 424 samples; in each training cycle (epoch), the model first calculates the predicted value by forward propagation, then calculates the loss and updates the model parameters by back propagation. In order to monitor the training process, the training loss is printed once every 10 cycles.

[0135] Model prediction

[0136] After training, the model predicts on the test set to evaluate its generalization ability; the test set contains the remaining samples except the training set; the prediction result is obtained by gradient-free calculation to ensure that the model parameters are not updated during prediction; in order to intuitively show the prediction performance of the model, we draw the comparison chart of predicted temperature and actual temperature, and calculate the difference between predicted value and actual value of each sample.

[0137] Prediction result analysis

[0138] From the prediction results, the model can better capture the trend of temperature change, and the predicted value and the actual value maintain high consistency; the network model shows excellent performance in temperature prediction task. We follow the data set division principle of previous research, a total of 12 groups of data, the first 10 groups of data as the training set, the last 2 groups as the prediction set, to ensure that the model can learn fully and objectively verify its generalization ability, the experimental results are as follows Figures 10-12 shown.

[0139] Figure 10 The RNN-ARMB model training loss chart is shown in Figure 11 The RNN-ARMB model temperature actual value prediction value comparison chart (the left side is the first group of prediction, and the right side is the second group of prediction) is shown in Figure 12The RNN-ARMB model temperature prediction error bar chart (left is the first group of predictions, right is the second group of predictions).

[0140] The training loss of the RNN-ARMB network model is significantly reduced from 1.8411 to 1.4589 compared to the original model, and the prediction mean square error is also greatly reduced from 0.4783 to 0.1856. These improvements not only indicate that the model has higher optimization efficiency in the training process, but also significantly improve the fitting ability of the data; It is particularly worth mentioning that the temperature prediction accuracy of the model reaches an unprecedented 0.3℃, fully demonstrating the leading position of the RNN-ARMB network model in the field of temperature prediction; Figure 10 The loss curve of the RNN-ARMB model in the training process is shown, and compared with the previous model, its loss decreases more rapidly and stably, which highlights the efficiency of the model; Figure 11 By comparing the predicted value with the actual value, it is observed that the two curves in the figure almost completely coincide, which intuitively shows the accuracy of the RNN-ARMB model prediction; finally, Figure 12 The prediction error bar chart of the RNN-ARMB model further proves the excellent performance of the model, and the prediction error is extremely small and uniformly distributed.

[0141] In order to verify the effectiveness of each added module in the RNN-ARMB model, we conducted a series of ablation experiments; In these experiments, we gradually removed different modules in the model and observed the changes in model performance; The specific experimental results are shown in Table 1, where the training minimum loss (Loss) and mean square error (MSE) are used as the main indicators to evaluate the performance of the model.

[0142] Table 1 Network model ablation experiment

[0143]

[0144] As can be seen from Table 1, the complete RNN-ARMB model achieves the lowest Loss value (1.4589) and MSE value (0.1856), indicating that the model has achieved the best performance under the current configuration; then, we removed the Attention mechanism, residual connection (Res-Connection), memory unit (Memory), batch normalization (Batch-Norm), and optimization function modules one by one, and re-evaluated the performance of the model; the experimental results show that the removal of any single module will cause a significant decline in the performance of the model; for example, after removing the Attention mechanism (RNN-ARMB-Attention), the Loss value increased to 1.7681, and the MSE value also rose to 0.3522; similarly, the removal of other modules such as residual connection, memory unit, and batch normalization also led to the deterioration of the model performance; in addition, we also tried the combined experiment of removing multiple modules at the same time, and the results also showed further decline in the model performance; these ablation experiment results strongly prove the effectiveness of each added module in the RNN-ARMB model; each module makes an important contribution to the overall performance of the model, and their synergistic effect enables the RNN-ARMB model to achieve superior performance on the given task.

[0145] The RNN-ARMB model proposed in this study has achieved significant results in the temperature prediction task of multi-modal data fusion; in order to comprehensively evaluate its performance, we compared the prediction accuracy and error of the RNN-ARMB model with other advanced temperature prediction models in the current field; the comparison results show that the RNN-ARMB model exhibits superior performance with a mean square error of 0.1856 and a prediction accuracy of 0.3℃; in contrast, the random forest model of Xiong's team has a mean square error of 1.1400 and a prediction accuracy of 1℃; the Hadamard-GLCM model of Hu's team has a mean square error of 0.8475 and a prediction accuracy of 0.5℃; the temperature field reconstruction model of Wang's team has a mean square error of 0.4200 and a prediction accuracy of 0.7℃; the empirical mode decomposition model of Chen's team has a mean square error of 1.0300 and a prediction accuracy of 1℃; while the sound attenuation model of Fu's team has a mean square error of 0.9142, but its prediction accuracy reaches 2℃, which may be due to the high sensitivity of the model under certain specific conditions; in summary, the RNN-ARMB model performs excellently in terms of mean square error and prediction accuracy, fully proving its effectiveness and advancement in the task of multi-modal data fusion temperature prediction; the comparison table is shown in Table 2.

[0146] Table 2 Error and prediction accuracy comparison of different research team models

[0147]

[0148] Example:

[0149] like Figure 13 As shown, taking the treatment of breast cancer using the microwave hyperthermia ultrasound non-invasive temperature measurement system of this application as an example, firstly, a B-mode ultrasound machine, a microwave hyperthermia machine, and a computer are placed, and the B-mode ultrasound machine is connected to the computer, and the microwave hyperthermia machine is connected to the antenna; then, the antenna and ultrasound probe are fixed; the antenna and ultrasound probe are perpendicular; next, the tumor site of the patient's breast is placed in the central area of ​​the antenna energy radiation and directly below the B-mode ultrasound probe to achieve precise treatment of the tumor area; then, the power switch of the microwave hyperthermia machine is turned on, the time is set to 10 minutes, and the power is set to 6 watts, the B-mode ultrasound device is turned on and switched to small organ measurement mode, and the gain is adjusted to achieve balance, the interface of the microwave hyperthermia ultrasound non-invasive temperature measurement system is opened on the computer; the working switch of the microwave hyperthermia machine is turned on; the power and time of the microwave hyperthermia machine are dynamically adjusted according to the temperature value of the computer temperature measurement interface, and the temperature of the hyperthermia area is controlled by controlling the microwave heating power, thereby achieving the maximum treatment effect; after the treatment is completed, the microwave hyperthermia machine and the B-mode ultrasound device are turned off.

[0150] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein; this application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein; the specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0151] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A microwave hyperthermia ultrasonic non-invasive temperature measurement system based on an RNN-ARMB model, characterized in that, The system comprises: a microwave hyperthermia instrument; a waveguide antenna connected to the output end of the microwave hyperthermia instrument, the microwave hyperthermia instrument feeding energy to the waveguide antenna, and the waveguide antenna generating heat; a B-mode ultrasound device for collecting ultrasound video images; a computer connected to the B-mode ultrasound device for receiving the ultrasound video images collected by the B-mode ultrasound device and converting them into ultrasound image data, and inputting the ultrasound image data into a trained RNN-ARMB model to obtain a predicted temperature.

2. The microwave hyperthermia ultrasonic non-invasive temperature measurement system based on the RNN-ARMB model according to claim 1, characterized in that, The computer is provided with: a conversion module configured to convert the ultrasound video images into ultrasound image data to be detected; a preprocessing module configured to perform frame average filtering and median filtering on the ultrasound image data to be detected to obtain filtered ultrasound image data; a region of interest positioning module configured to position the filtered ultrasound image data to obtain image data after positioning of the region of interest; an extraction module configured to obtain the mean gray value, standard deviation of gray value, average gradient, and hybrid entropy of the image data after positioning of the region of interest; a signal-to-noise ratio extraction module configured to obtain the signal-to-noise ratio of the image data after positioning of the region of interest; a k parameter calculation module configured to obtain an improved k parameter from the signal-to-noise ratio of the image data after positioning of the region of interest by using an improved zero-difference K distribution model; a prediction module configured to input the improved k parameter, mean gray value, standard deviation of gray value, average gradient, and hybrid entropy of the image data after positioning of the region of interest into the trained RNN-ARMB model to obtain a predicted temperature.

3. The microwave hyperthermia ultrasonic non-invasive temperature measurement system based on the RNN-ARMB model according to claim 2, characterized in that: The positioning method of the filtered ultrasound image data is as follows: taking the position of the temperature probe as the center, a square region with a side length of 256 pixels is defined as the image after positioning of the region of interest.

4. The microwave hyperthermia ultrasound non-invasive temperature measurement system based on the RNN-ARMB model according to claim 2, wherein: the mean gray value, standard deviation of gray value, average gradient, and hybrid entropy are calculated by the following formulas: gray mean value: wherein M and N are the number of rows and columns of the image respectively, and f(i,j) is the gray value of the image at coordinate (i,j); Gray scale standard deviation: wherein M and N are the number of rows and columns of the image respectively, and f(i,j) is the gray value of the image at coordinate (i,j); Average Gradient: where M and N are the number of rows and columns of the image, respectively, and are the gradients of the image in the x and y directions, respectively. Mixing entropy: where p ijk is the probability of the simultaneous occurrence of the gray values i,j,k.

5. The microwave hyperthermia ultrasound non-invasive temperature measurement system based on the RNN-ARMB model according to claim 2, wherein: the signal-to-noise ratio is obtained by the following method: step 501, calculating the ultrasound backscattering envelope amplitude A by using Hilbert transform; the definition of Hilbert transform is the convolution of signal x(t) and the impulse response h(t) of the system: wherein denotes a Hilbert transform in the time domain; Constructing the analysis signal: Compute envelope amplitude: step 502, calculating the signal-to-noise ratio by using the envelope amplitude by the following formula: where P signal is the signal power, P noise is the noise power, and is calculated as follows: where μ E is the mean value of the signal envelope amplitude, calculated as 6. The microwave hyperthermia ultrasonic non-invasive temperature measurement system based on RNN-ARMB model according to claim 2, characterized in that: the improved zero-difference K distribution model is as follows: the improved zero-difference K distribution model is represented by the following formula: where Rvis the signal-to-noise ratio, μ is the number of effective scatterers, v is a positive real number, Γ(z) is the Gamma function, pFq(a1,...,ap; b1,...,bq; c) denotes the generalized hypergeometric series; csc(·) is the cosecant function; η is defined as: Scaling factor calculation formula: The new signal-to-noise ratio R' calculation formula: R' = (R - R min )*β R of formula (1) v is equal to R` of formula (4); Gamma function substitution formula:

7. The microwave hyperthermia ultrasonic non-invasive temperature measurement system based on RNN-ARMB model according to claim 2, characterized in that: The RNN-ARMB model comprises: an RNN module: the RNN module adopts a standard RNN layer, the input feature dimension is 5, the number of hidden layer units is 10, the structure is single-layer, and batch_first is set to True to match the shape of the input data. The attention mechanism module is arranged between the hidden layer of the RNN module and the output layer of the RNN module, is used for weighting the output of the RNN, and extracts key information; The memory unit module is arranged after the output layer of the RNN module, introduces a learnable memory matrix with a size of 100x10, and interacts with the RNN output through the attention mechanism; The residual connection module is arranged after the output layer of the RNN module, is used for adding the input of the RNN to the output of the last time step of the RNN after linear layer transformation, and enhances the retention ability of the model to the input information; The batch normalization module is arranged after the output layer of the RNN module, so as to accelerate the training process and improve the stability of the model; The full connection layer module is arranged after the residual connection module, the batch normalization module and the memory unit module, combines the RNN output, the residual connection result and the memory information, and outputs the final temperature prediction value through the full connection layer.

Images