A method of X-ray thermoacoustic imaging based on multidimensional information fusion network

Abstract

The present invention discloses an X-ray thermoacoustic imaging method based on a multidimensional information fusion network, and relates to the field of artificial intelligence medical imaging technology. The present invention can adapt to the rapid reconstruction of high-quality X-ray thermoacoustic images under different sampling conditions, effectively extract and utilize the information of signals and images, and complement the characteristic information in the signals and images. While ensuring the quality of reconstruction, the time cost is reduced, the defects of traditional X-ray thermoacoustic imaging methods are avoided, and the development process of X-ray thermoacoustic imaging is accelerated.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence medical imaging technology, and in particular to an X-ray thermoacoustic imaging method based on a multidimensional information fusion network. Background Art

[0002] With the development of information science and technology, computer science and technology have also been widely used in the medical field. Modern medicine can be said to be inseparable from computer science and technology, among which computer-aided diagnosis technology and computer-guided cancer radiotherapy technology are increasingly attracting people's attention. They are cutting-edge topics that have emerged in the field of multidisciplinary intersection in recent years.

[0003] X-ray induced acoustic computed tomography (XACT) is a hybrid imaging technology based on the thermoacoustic effect stimulated by X-rays. Specifically, XACT uses the thermoacoustic effect. When the target tissue is irradiated by an X-ray pulse, part of the energy is deposited in the tissue. This part of energy will cause the local temperature to rise, causing thermoelastic expansion, thereby generating acoustic pressure signals within the ultrasonic range. These signals are collected by an ultrasonic transducer and converted into images through a reconstruction algorithm. The low scattering characteristics of XACT give it a deeper imaging depth. At the same time, the intensity of the acoustic pressure signal induced by X-rays is proportional to the radiation dose, so XACT can achieve real-time monitoring of the in vivo dose. In addition, the X-ray induced acoustic wave (XA) propagates in 3D, so XACT can achieve three-dimensional reconstruction of the target through a single X-ray pulse, thereby greatly reducing the radiation dose required for in vivo 3D imaging. These characteristics make XACT show a wide range of application prospects in many fields of clinical medicine, such as bone density mapping, cancer auxiliary diagnosis and dose monitoring.

[0004] At present, X-ray thermoacoustic image reconstruction mainly uses the back projection (BP) algorithm, the time reversal (TR) algorithm, and the model-based (MB) algorithm.

[0005] Due to the simplicity and high efficiency of the BP algorithm, the BP algorithm is usually used to reconstruct the XA signal into the initial sound pressure distribution. However, under the condition of XA signal with low signal-to-noise ratio, the XACT image reconstructed by BP may have more noise artifacts and perform poorly in reconstructing the sound pressure distribution. TR is a method based on the wave equation, which can obtain higher reconstruction quality than TR, but performs poorly when there are fewer detectors or the detector array structure is incomplete. In addition, TR reconstruction requires the evaluation of wave propagation at each step, which leads to its large time cost. The model-based XACT reconstruction method relies on the discretization of the photoacoustic forward model, which can significantly improve the reconstruction quality by combining inherent properties and greatly speed up the XACT reconstruction while avoiding the storage of the measurement matrix. However, the model-based XACT reconstruction method requires accurate prior knowledge and model matrix.

[0006] Existing X-ray thermoacoustic image reconstruction methods cannot obtain high-quality X-ray thermoacoustic images, and are prone to noise artifacts and other problems during the reconstruction process. At the same time, due to the limitations of the imaging algorithm itself, the improvement of X-ray thermoacoustic imaging image quality is often at the expense of time, or requires other prior knowledge, model matrices and other harsh conditions. In short, the main problem with the existing X-ray thermoacoustic image reconstruction methods is that it is difficult to take into account both the reconstruction resolution and reconstruction speed of X-ray thermoacoustic images.

[0007] Therefore, a new solution to the above problems needs to be proposed. Summary of the invention

[0008] The purpose of the present invention is to provide an X-ray thermoacoustic imaging method based on a multidimensional information fusion network to solve the existing problems, improve the resolution of X-ray reconstructed images, reduce the time required for reconstruction, and accelerate the application of X-ray thermoacoustic imaging in clinical practice.

[0009] To achieve the above object, the present invention provides the following technical solution: an X-ray thermoacoustic imaging method based on a multidimensional information fusion network, comprising at least the following steps:

[0010] S1: Simulates the initial sound pressure data set after X-ray radiotherapy, and uses K-Wave to simulate the X-ray-induced ultrasonic signal simulation data under different sampling conditions;

[0011] S2: Build an end-to-end multi-dimensional information fusion network model MFF-NET module by module and output the final X-ray thermoacoustic image;

[0012] S3: Using a supervised learning method, MFF-NET is trained on the initial sound pressure data set and the ultrasonic signal simulation data set, and the training task of reconstructing ultrasonic signals into high-quality X-ray thermoacoustic images under different sampling environments is completed in turn to obtain a trained X-ray thermoacoustic image reconstruction model;

[0013] S4: Input the ultrasound signal induced by X-rays under radiotherapy into the trained X-ray thermoacoustic image reconstruction model, and output the reconstructed high-quality X-ray thermoacoustic image.

[0014] S5: The predicted sound pressure distribution is restored to generate the corresponding dose distribution, and the sound pressure and dose map slices are stacked into 3D form for 3D dose verification.

[0015] Furthermore, the S1 at least includes the following steps:

[0016] Differentiate different tissues according to HU values ​​in CT slices, and obtain tissue-specific parameters, wherein the specific parameters at least include tissue density and Grunisson coefficient;

[0017] The dose coefficient was calculated by tissue-specific parameters;

[0018] There is a conversion relationship between the sound pressure generated in X-ray radiotherapy and the given dose. According to the conversion relationship, the radiation dose distribution in the radiotherapy plan is converted into the initial sound pressure distribution generated after X-ray radiotherapy to obtain the initial sound pressure data set;

[0019] Preset image resolution and simulate the sampling process under different sensor numbers and sensor array shapes in k-Wave;

[0020] Presetting an initial sound pressure distribution within a scanning area of ​​the sensor array, and setting parameters of the ultrasonic sensor, the parameters including at least frequency and bandwidth;

[0021] The sound pressure propagation process is simulated by k-Wave, random noise is added, and a simulated ultrasonic simulation signal data set is collected.

[0022] Furthermore, the end-to-end multi-dimensional information fusion network model MFF-NET in S2 includes a signal feature extraction module on the left, an image feature extraction module on the right, and a feature fusion module in the middle.

[0023] Furthermore, the signal feature extraction module includes a bidirectional long short-term memory neural network BiLSTM and a codec U-net, the BiLSTM contains a total of 512 units, its input is a simulated ultrasonic signal sequence, the input dimension is the number of ultrasonic transducers i×the number of sampling points w, and the extracted features are restored to a feature map of size 128×128 through a fully connected layer;

[0024] The downsampling block of the codec U-net includes two consecutive convolution blocks and a pooling layer with a size of 2×2 and a stride of 2, wherein the convolution block is composed of a convolution layer with a size of 3×3 and a stride of 1, a batch normalization layer and a ReLu activation function, and the upsampling block of the codec U-net includes two convolution layers consistent with those in the downsampling block and a deconvolution layer with a size of 2×2 and a stride of 2.

[0025] Further, the image feature extraction module includes a main body and a cross attention module;

[0026] The main body is a Unet architecture, and a cross attention module is integrated on the basis of the Unet architecture to replace the residual connection of the Unet;

[0027] The cross-attention module first extracts multiple features from the encoder stage, applies average pooling to extract patches, and uses 1×1 convolution to implement the projection operation, and finally uses continuous channel cross-attention and spatial cross-attention mechanisms.

[0028] Furthermore, the feature fusion module introduces a multi-level feature information fusion module, which takes the feature maps of different sizes outputted from the image post-processing module in the image feature extraction module as input, firstly unifies the number of channels of features of different scales through 1×1 convolution, then applies spatial and channel attention mechanisms, and then applies pooling, identity mapping and interpolation operations to features of different sizes to unify them to feature sizes corresponding to image post-processing, and finally applies Hadamard multiplication to all resampled feature maps to enhance features;

[0029] The features after the image feature fusion module are connected with the feature map of the same size of the decoder U-net in the signal extraction module, and finally the final X-ray thermoacoustic image is output.

[0030] Furthermore, the S3 at least includes the following steps:

[0031] Taking the ultrasonic signal in the ultrasonic simulation signal data set and the X-ray thermoacoustic image reconstructed by FBP as input;

[0032] Using the initial sound pressure distribution corresponding to the initial sound pressure data set as a label, and adopting a supervised learning method for training;

[0033] Test MFF-Net and obtain the trained X-ray thermoacoustic image reconstruction model.

[0034] Furthermore, the supervised learning method training includes at least the following steps:

[0035] Based on the back-propagation Adam algorithm, input the ultrasonic simulation signal data set and the initial sound pressure data set of batch size;

[0036] The root mean square error is used as a loss function to supervise the training of the network. The root mean square error is used to supervise the final output X-ray thermoacoustic image and the output of the image feature extraction module, and the weight is set to balance the supervised learning strength of the two.

[0037] The final loss function is used as the evaluation index. On the basis of reducing overfitting as much as possible, the best set of parameters is selected as the best parameter model obtained by training.

[0038] Furthermore, the S4 at least includes the following steps:

[0039] Obtaining X-ray-induced ultrasound signals through simulation experiments or real radiotherapy;

[0040] Input into the trained X-ray thermoacoustic image reconstruction model;

[0041] The reconstructed target X-ray thermoacoustic image is output as a reconstructed high-quality X-ray thermoacoustic image.

[0042] Furthermore, the S5 at least includes the following steps:

[0043] According to the relationship between acoustic pressure and dose, the acoustic pressure distribution predicted by the model is converted into the deposited dose distribution. The acoustic pressure and dose distribution slices are stacked and restored to 3D form. The effectiveness of the proposed method in 3D dose verification function is verified.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention first obtains the initial sound pressure data set after simulating radiotherapy through attributes such as HU value, and then uses K-wave to simulate the X-ray induced ultrasonic simulation signal under different sampling conditions, constructs an end-to-end multidimensional information fusion network model MFF-NET, and adopts a supervised learning method to train MFF-NET on the data set, and sequentially completes the training task of reconstructing the ultrasonic signal to high-quality X-ray thermoacoustic image under different sampling environments, and obtains a trained X-ray thermoacoustic image reconstruction model. The X-ray induced ultrasonic signal under radiotherapy is input into the trained X-ray thermoacoustic image reconstruction model, and the reconstructed high-quality X-ray thermoacoustic image is output. As a result, it can adapt to the rapid reconstruction of high-quality X-ray thermoacoustic images under different sampling conditions, effectively extract and utilize the information of signals and images, and complement the characteristic information in the signals and images. While ensuring the quality of reconstruction, it reduces the time cost and avoids the defects of traditional X-ray thermoacoustic imaging methods, and accelerates the development of X-ray thermoacoustic imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0047] Figure 1 Schematic diagram for obtaining the initial sound pressure data set;

[0048] Figure 2 The K-Wave toolbox of the present invention simulates different numbers of sensors and sensor array shapes;

[0049] Figure 3 It is the end-to-end multi-dimensional information fusion network model MFF-NET diagram of the present invention;

[0050] Figure 4 Schematic diagram of the cross attention module of the present invention;

[0051] Figure 5 It is a schematic diagram of a multi-level feature information fusion module of the present invention;

[0052] Figure 6 X-ray thermoacoustic image reconstructed for the present invention.

[0053] Figure 7 This is a three-dimensional schematic diagram of sound pressure and dose reconstructed by the present invention. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0055] An X-ray thermoacoustic imaging method based on a multidimensional information fusion network comprises at least the following steps:

[0056] S1: Simulates the initial sound pressure data set after X-ray radiotherapy, and collects the X-ray induced ultrasonic signal simulation data obtained by K-Wave simulation under different sampling conditions;

[0057] S2: Build an end-to-end multi-dimensional information fusion network model MFF-NET module by module and output the final X-ray thermoacoustic image;

[0058] S3: Using a supervised learning method, MFF-NET is trained on the initial sound pressure data set and the ultrasonic signal simulation data set, and the training task of reconstructing ultrasonic signals into high-quality X-ray thermoacoustic images under different sampling environments is completed in turn to obtain a trained X-ray thermoacoustic image reconstruction model;

[0059] S4: Input the ultrasound signal induced by X-rays under radiotherapy into the trained X-ray thermoacoustic image reconstruction model, and output the reconstructed high-quality X-ray thermoacoustic image.

[0060] S5: The predicted sound pressure distribution is restored to generate the corresponding dose distribution, and the sound pressure and dose map slices are stacked into 3D form for 3D dose verification.

[0061] Specifically, the schematic diagram of obtaining the initial sound pressure data set in S1 is as follows: Figure 1 As shown, the tissue is divided into different tissue types according to the HU value in the specific CT slice, and then specific attribute parameter values ​​are assigned according to different tissue types. The relationship between the acoustic pressure generated in X-ray radiotherapy and the given dose is as follows:

[0062] p0(r)=ρΓη th D(r)

[0063] where p0(r) is the initial sound pressure distribution, D(r) is the given dose, ρ is the tissue density, and Γ is the Grunisson coefficient.

[0064] According to the relationship between the initial sound pressure and the given dose, the radiation dose in the radiotherapy plan is multiplied by the dose coefficient of the corresponding area to obtain the initial sound pressure distribution after X-ray radiotherapy, and the obtained initial sound pressure distribution is used as the label in the data set.

[0065] Specifically, the K-Wave toolbox simulates different numbers of sensors and sensor array shapes such as Figure 2 In order to simulate the ultrasonic simulation signal under different sampling conditions, the initial sound pressure distribution obtained in S1 is preset in Figure 2 In the scanning area of ​​the sensor array, the center frequency of the ultrasonic sensor is set to 1MHZ, the bandwidth is 100%, and the sampling rate is 8MHZ; k-Wave is used to simulate the sound pressure propagation process. At the same time, in order to be as close to the actual sampling as possible, Gaussian noise of random intensity is added in the process of signal acquisition, and the simulated ultrasonic simulation signal data is collected. The corresponding initial sound pressure image is matched to form the required data set, and the obtained data set is split into a training set and a test set in a ratio of 7:3.

[0066] Specifically, the end-to-end multidimensional information fusion network model MFF-NET in S2 is as follows: Figure 3 As shown in the figure, the overall multi-dimensional information fusion network model is mainly composed of three independent modules, namely, signal feature extraction module, image feature extraction module and feature fusion module. The signal feature extraction module is used to extract features from X-ray induced ultrasonic signals, the image feature extraction module is used to extract features from the initial X-ray thermoacoustic image reconstructed by FBP, and the feature fusion module narrows the semantic gap between the features extracted from the signal and the features from the image.

[0067] The signal feature extraction module consists of a bidirectional long short-term memory neural network BiLSTM and a codec U-net. The BiLSTM contains a total of 512 units. Its input is a simulated ultrasonic signal sequence with an input dimension of the number of ultrasonic transducers i×the number of sampling points w. The extracted features are restored to a feature map of size 128×128 through a fully connected layer. The codec downsampling block of contains two consecutive convolution blocks and a pooling layer of size 2×2 and stride 2. The convolution block consists of a convolution layer of size 3×3 and stride 1, a batch normalization layer, and a ReLu activation function. The codec upsampling block of contains two convolution layers consistent with the downsampling block and a deconvolution layer of size 2×2 and stride 2.

[0068] The image feature extraction module uses FBP to reconstruct the initial X-ray thermoacoustic image induced by the X-ray ultrasonic signal as input. The main body of the image feature extraction module is the Unet architecture, and the cross attention module is integrated on this basis to replace the residual connection of Unet. The cross attention module is as follows Figure 4 As shown in the figure, the cross-attention module first extracts multiple features from the encoder stage, applies average pooling to extract patches, and uses 1×1 convolution to perform projection operations to obtain encoder tokens. This process can be described as:

[0069] Ti=DConv1D(Reshape(AvgPool2D(Ei)))

[0070] Where Ti represents the flattened patch of the i-th encoder stage, and Ei represents n encoder stages of different scales.

[0071] Continuous channel-wise and spatial-wise attention mechanisms are used on these tokens to capture long-range dependencies. Finally, these tokens are serialized and upsampled using layer normalization and GeLU and connected to their decoder counterparts.

[0072] Feature fusion module, introduces a multi-level feature information fusion module, such as Figure 5 As shown, the feature maps of different sizes output from the image post-processing module in the feature extraction module are used as input. First, the feature maps of each level of the input are i 0 A 1×1 convolution operation is performed to unify the number of channels of features of different scales, and then the spatial and channel attention mechanism is applied. The specific process can be described as:

[0073]

[0074] Here f i 1 Represented as the feature map after the i-th level, and Represents the spatial and channel attention mechanism, and the processed feature map obtained is f i 2 , where the feature map f i 2 The width, height and number of channels are H i , W i , c.

[0075] In order to unify the input feature maps of different sizes to the same resolution, according to the size relationship between the target feature map and the current feature map, average pooling, identity mapping, and bilinear interpolation are used to adjust the feature map of each level to shrink, keep unchanged, and enlarge. The process is:

[0076]

[0077] Among them, Pool, Map, and Interpol represent average pooling, identity mapping, and bilinear interpolation, respectively.

[0078] After scaling all feature maps to a uniform size, Hadamard multiplication is applied to all resized feature maps to enhance features.

[0079] The features finally outputted after the feature fusion module are connected with the feature map of the same size of the decoder in the signal extraction module, and finally the final X-ray thermoacoustic image is outputted.

[0080] Specifically, a supervised learning method is used to train MFF-NET on the dataset, and the training task of reconstructing high-quality X-ray thermoacoustic images from ultrasonic signals under different sampling environments is completed in turn, and a trained X-ray thermoacoustic image reconstruction model is obtained, including

[0081] The ultrasonic signal in the ultrasonic simulation signal dataset is used as the input of the signal feature extraction module. The ultrasonic signal is reconstructed into the initial X-ray thermoacoustic image through the FBP algorithm as the input of the image feature extraction module. The corresponding initial sound pressure distribution in the initial sound pressure dataset is used as the label. The supervised learning method is used to train and test the MFF-Net to obtain the trained X-ray thermoacoustic image reconstruction model.

[0082] Specifically, supervised learning methods include

[0083] Based on the back-propagation Adam algorithm, a training set of batch size is input, and the root mean square error is used as the loss function to supervise the training of the network, that is:

[0084]

[0085] where Y I Represents the training label, y i Represents the output of the network.

[0086] The root mean square error is used to supervise the final output of the X-ray thermoacoustic image and the output of the image feature extraction module, and the weights are set to balance the supervised learning strength of the two. The final loss function is used as the evaluation indicator, which can be described here as:

[0087] L=αL recon+(1-α)L post

[0088] Here L recon is the loss of the final output image, L post is the loss of the output image of the post-processing module, α is the weight of the loss function, and here we set α to 0.75.

[0089] On the basis of reducing overfitting as much as possible, a set of parameters with the best effect is selected as the best parameter model obtained through training.

[0090] Specifically, the ultrasound signal induced by X-rays under radiotherapy is input into the trained X-ray thermoacoustic image reconstruction model, and the reconstructed high-quality X-ray thermoacoustic image is output. Figure 2 The reconstructed X-ray thermoacoustic images under the three sensor array settings are shown in Figure 2. Figure 6 As shown, Figure 6 The X-ray thermoacoustic images reconstructed by several traditional methods under different sampling conditions are also included. The specific reconstruction process of MFF-Net is as follows:

[0091] X-ray induced ultrasound signals are obtained through simulation experiments or real radiotherapy, and the ultrasound signals are reconstructed into X-ray thermoacoustic initial images through FBP. The signals and the initial images are simultaneously input into the trained X-ray thermoacoustic image reconstruction model, and the reconstructed target X-ray thermoacoustic image is output as the reconstructed high-quality X-ray thermoacoustic image.

[0092] Specifically, the predicted acoustic pressure distribution is restored to generate the corresponding dose distribution, and the acoustic pressure and dose map slices are stacked into 3D form for 3D dose verification, including:

[0093] According to the relationship between sound pressure and dose, the sound pressure distribution predicted by the model is converted into the corresponding deposition dose distribution, so that the actual dose deposition distribution is obtained. Then, according to the relative position of the slices in the anatomical information, the sound pressure and dose distribution slices are stacked and restored to a 3D form to obtain the actual 3D sound pressure distribution and 3D dose distribution (such as Figure 7 As shown in the figure, the planned dose distribution was compared with the actual dose distribution, and the gamma index was used to quantitatively evaluate the predicted 3D dose distribution. The gamma index analysis was performed with threshold standards of 3% / 3mm, 2% / 3mm, 3% / 5mm, and 2% / 5mm to verify the effectiveness of the proposed method in 3D dose verification function.

[0094] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An X-ray thermoacoustic imaging method based on a multidimensional information fusion network, characterized in that: At least the following steps are included: S1: Simulates the initial sound pressure data set after X-ray radiotherapy, and uses K-Wave to simulate the X-ray-induced ultrasonic signal simulation data under different sampling conditions; S2: Build an end-to-end multi-dimensional information fusion network model MFF-NET module by module and output the final X-ray thermoacoustic image; The end-to-end multidimensional information fusion network model MFF-NET in S2 includes a signal feature extraction module on the left, an image feature extraction module on the right, and a feature fusion module in the middle; The signal feature extraction module includes a bidirectional long short-term memory neural network BiLSTM and a codec U-net. The BiLSTM contains a total of 512 units, and its input is a simulated ultrasonic signal sequence. The input dimension is the number of ultrasonic transducers i×the number of sampling points w. The extracted features are restored to a feature map of size 128×128 through a fully connected layer; The downsampling block of the codec U-net comprises two consecutive convolution blocks and a pooling layer of size 2×2 and stride 2, wherein the convolution block comprises a convolution layer of size 3×3 and stride 1, a batch normalization layer and a ReLu activation function, and the upsampling block of the codec U-net comprises two convolution layers consistent with those in the downsampling block and a deconvolution layer of size 2×2 and stride 2; The image feature extraction module includes a main body and a cross attention module; The main body is a Unet architecture, and a cross attention module is integrated on the basis of the Unet architecture to replace the residual connection of the Unet; The cross-attention module first extracts multiple features from the encoder stage, applies average pooling to extract patches, and uses 1×1 convolution to implement projection operations, and finally uses continuous channel cross-attention and spatial cross-attention mechanisms; The feature fusion module introduces a multi-level feature information fusion module, which takes the feature maps of different sizes output from the image post-processing module in the image feature extraction module as input, firstly unifies the number of channels of features of different scales through 1×1 convolution, then applies spatial and channel attention mechanisms, and then applies pooling, identity mapping and interpolation operations to features of different sizes to unify them to feature sizes corresponding to image post-processing, and finally applies Hadamard multiplication to all resampled feature maps to enhance features; The features after the image feature fusion module are connected with the feature map of the same size of the decoder U-net in the signal extraction module, and finally the final X-ray thermoacoustic image is output; S3: Using a supervised learning method, MFF-NET is trained on the initial sound pressure data set and the ultrasonic signal simulation data set, and the training task of reconstructing ultrasonic signals into high-quality X-ray thermoacoustic images under different sampling environments is completed in turn to obtain a trained X-ray thermoacoustic image reconstruction model; S4: Input the ultrasound signal induced by X-rays under radiotherapy into the trained X-ray thermoacoustic image reconstruction model, and output the reconstructed high-quality X-ray thermoacoustic image; S5: The predicted sound pressure distribution is restored to generate the corresponding dose distribution, and the sound pressure and dose map slices are stacked into 3D form for 3D dose verification.

2. The X-ray thermoacoustic imaging method based on a multidimensional information fusion network according to claim 1, characterized in that: The S1 at least comprises the following steps: Differentiate different tissues according to HU values ​​in CT slices, and obtain tissue-specific parameters, wherein the specific parameters at least include tissue density and Grunisson coefficient; The dose coefficient was calculated by tissue-specific parameters; There is a conversion relationship between the sound pressure generated in X-ray radiotherapy and the given dose. According to the conversion relationship, the radiation dose distribution in the radiotherapy plan is converted into the initial sound pressure distribution generated after X-ray radiotherapy to obtain the initial sound pressure data set; Preset image resolution and simulate the sampling process under different sensor numbers and sensor array shapes in k-Wave; Presetting an initial sound pressure distribution within a scanning area of ​​the sensor array, and setting parameters of the ultrasonic sensor, the parameters including at least frequency and bandwidth; The sound pressure propagation process is simulated by k-Wave, random noise is added, and a simulated ultrasonic simulation signal data set is collected.

3. The X-ray thermoacoustic imaging method based on a multidimensional information fusion network according to claim 1, characterized in that: The S3 at least includes the following steps: Taking the ultrasonic signal in the ultrasonic simulation signal data set and the X-ray thermoacoustic image reconstructed by FBP as input; Using the initial sound pressure distribution corresponding to the initial sound pressure data set as a label, and adopting a supervised learning method for training; Test MFF-Net and obtain the trained X-ray thermoacoustic image reconstruction model.

4. The X-ray thermoacoustic imaging method based on a multidimensional information fusion network according to claim 3, characterized in that: The supervised learning method training includes at least the following steps: Based on the back-propagation Adam algorithm, input the ultrasonic simulation signal data set and the initial sound pressure data set of batch size; The root mean square error is used as a loss function to supervise the training of the network. The root mean square error is used to supervise the final output X-ray thermoacoustic image and the output of the image feature extraction module, and the weight is set to balance the supervised learning strength of the two. The final loss function is used as the evaluation index. On the basis of reducing overfitting as much as possible, the best set of parameters is selected as the best parameter model obtained by training.

5. The X-ray thermoacoustic imaging method based on a multidimensional information fusion network according to claim 1, characterized in that: The S4 at least comprises the following steps: Obtaining X-ray-induced ultrasound signals through simulation experiments or real radiotherapy; Input into the trained X-ray thermoacoustic image reconstruction model; The reconstructed target X-ray thermoacoustic image is output as a reconstructed high-quality X-ray thermoacoustic image.

6. The X-ray thermoacoustic imaging method based on a multidimensional information fusion network according to claim 1, characterized in that: The S5 at least comprises the following steps: According to the relationship between acoustic pressure and dose, the acoustic pressure distribution predicted by the model is converted into the deposited dose distribution. The acoustic pressure and dose distribution slices are stacked and restored to 3D form. The effectiveness of the proposed method in 3D dose verification function is verified.

Images