A method for improving the imaging quality of Compton camera

Abstract

The present invention discloses a method for improving the imaging quality of a Compton camera, which belongs to the field of Compton camera imaging. The method comprises: collecting scattering and absorption data obtained after Compton scattering and photoelectric effect of a radioactive source on a Compton camera, and then screening the data; reconstructing the screened data into a two-dimensional image using a reconstruction algorithm, and then making a data set based on the reconstructed two-dimensional image, and then performing data enhancement on the data set, and finally obtaining an enhanced data set, and dividing it into a training set and a test set according to a 7:3 ratio; constructing a Compton camera imaging method based on a pix2pix model, wherein U‑Net is used as the main structure of a generator in the pix2pix model, a multi-scale feature extraction module is added to an encoder and a decoder, and an RS attention module is designed, and a Transformer architecture is used in a discriminator. The present invention greatly improves the problem of poor imaging quality of Compton cameras.

Description

Technical Field

[0001] The present invention relates to the field of Compton camera imaging, and in particular to a method for improving the imaging quality of a Compton camera. Background Art

[0002] A Compton camera is an imaging device used to detect gamma rays. It mainly calculates the spatial distribution of gamma sources by measuring the energy deposition and position of Compton scattering between gamma rays and detectors, and can directly locate the radiation source without a collimator. Compton cameras are based on Compton scattering imaging. A typical Compton camera consists of two layers of detectors, the first layer of absorption detectors and the second layer of scattering detectors. At present, the traditional image reconstruction algorithms of Compton cameras are mainly divided into two categories: analytical methods and iterative methods. In the analytical method, the direct back-projection method directly reconstructs the image by solving the relationship equation between the image and the projection data. It is the simplest and fastest reconstruction method. However, due to the problems of low scattering angle resolution, strong noise interference, and difficulty in image reconstruction during the imaging process of Compton cameras, improving their imaging quality has always been a hot topic and difficulty in research. The inventors of this application have found that post-processing the output image of the Compton camera using a deep learning method can effectively eliminate noise and enhance details. Summary of the invention

[0003] The purpose of the present invention is to propose a method for improving the imaging quality of Compton cameras. Firstly, a dataset of Compton camera imaging is produced to solve the problem of the lack of public datasets for Compton cameras. The Pix2Pix model is determined as the basic network, and its original generator and discriminator networks are reconstructed to complete high-quality image reconstruction.

[0004] The technical solution adopted by the present invention is:

[0005] A method for improving the imaging quality of a Compton camera, comprising:

[0006] Step 1, collecting scattering and absorption data obtained after Compton scattering and photoelectric effect of the radiation source on the Compton camera, and then screening the data;

[0007] Step 2: Use the reconstruction algorithm to reconstruct the filtered data into a two-dimensional image, then make a data set based on the reconstructed two-dimensional image, and then perform data enhancement on the data set to finally obtain the enhanced data set, which is divided into a training set and a test set in a ratio of 7:3;

[0008] Step three, construct a Compton camera imaging method based on the pix2pix model. In the pix2pix model, U-Net is used as the main structure of the generator, a multi-scale feature extraction module is added to the encoder and decoder, and an RS attention module is designed. The Transformer architecture is used in the discriminator.

[0009] Compared with the prior art, the present invention has the following beneficial effects:

[0010] First, the present invention creates a Compton camera imaging dataset, which solves the problem that there is no public dataset and cannot be trained. The dataset is preprocessed using SRGAN, and the dataset is enhanced to improve the training effect.

[0011] Secondly, the present invention adds a multi-scale feature extraction module to the encoder and decoder of the generator to increase the receptive field and enhance the extraction of image information.

[0012] Thirdly, the present invention designs an RS attention module and places it at each skip connection, merging the features processed by the attention module with the corresponding layer of the decoder part. The RS attention module can improve the feature selectivity of the skip connection, thereby improving the effect of information fusion.

[0013] Fourthly, the present invention uses the Transformer architecture in the discriminator instead of the traditional convolutional network, which can better capture the global structure and detail information of the image. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention is a flowchart of a method for improving the imaging quality of a Compton camera according to an embodiment of the present invention.

[0015] Figure 2 A generator structure diagram for improving pix2pix according to an embodiment of the present invention.

[0016] Figure 3 4 is a structural diagram of a multi-scale feature fusion module according to an embodiment of the present invention.

[0017] Figure 4 2 is a structural diagram of the RS attention module according to an embodiment of the present invention.

[0018] Figure 5 The discriminator structure diagram of pix2pix is ​​improved according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The technical solutions and advantages of the present invention will be further described below in conjunction with the accompanying drawings in the present invention.

[0020] The present invention provides a method for improving the imaging quality of a Compton camera, the flow chart of the method is as follows: Figure 1 As shown, specifically including:

[0021] Step 1: Collect the scattering and absorption data obtained after the Compton scattering and photoelectric effect of the radiation source on the Compton camera, and then screen the data.

[0022] The data detected by the Compton camera is screened and processed, including filtering the data according to the energy of the detected radiation source to obtain data that meets the Compton scattering conditions, including:

[0023] Compton camera detects the Compton scattering of the radiation source on the first layer of scattering detectors. The scattering position and deposited energy , a photoelectric effect occurs on the second layer of absorption detector, and the absorption position and energy , the relationship between the energies is as follows:

[0024] ,

[0025] in, is the energy of the incident photon. The error in the above formula can fluctuate within 5 keV. Data that do not satisfy the above formula will be cleaned up.

[0026] Step 2: Use the reconstruction algorithm to reconstruct the filtered data into a two-dimensional image, and then create a data set based on the reconstructed two-dimensional image. Then, perform data enhancement on the data set, and finally obtain the enhanced data set, which is divided into a training set and a test set in a ratio of 7:3.

[0027] Two reconstruction algorithms are used to reconstruct two-dimensional images, including:

[0028] First, the imaging principle of Simple Back Projection (SBP) is used for analysis: the imaging space is first divided into m rows and n columns of voxels. When imaging, combined with the imaging principle of Compton camera, it is only necessary to calculate the total number of times each voxel in the imaging space is passed through by the reconstructed cone surface. The more times, the more likely this position is to be the source of the radiation source; then the Maximum Likelihood Expectation Maximization (MLEM) method is used for iteration: the statistical iterative algorithm is used to calculate the maximum probability image value under a given projection result. Each iteration will update the image estimate value. One iteration consists of projection operation, comparison of the projection operation result with the measurement data, back projection operation, and image update. The formula is as follows:

[0029] ,

[0030] Among them, j is the number of voxel grids in the specified imaging space, r is the number of iterations, M is the valid imaging event, and N represents the discrete grid. is the intensity of pixel grid j in the imaging space after r iterations, represents the probability that a photon is emitted in pixel grid j and detected as a valid event i, is the probability that a photon is emitted in pixel grid j and detected as a valid event.

[0031] The data set preparation includes: using the Simple Back Projection (SBP) method in the analytical method to image the filtered Compton camera data according to 1000, 4000, 8000, 12000, 40000, 100000 data respectively, which are low-statistic images. Then using the Maximum Likelihood Expectation Maximization (MLEM) method in the iterative method to image 300000 and 500000 data to obtain high-statistic images, all these images are introduced into SRGAN for image enhancement, and the enhanced low-statistic images and high-statistic images are made into sample pairs to prepare the initial data set.

[0032] Perform data augmentation on the dataset, including:

[0033] The data set is expanded by using data enhancement techniques such as horizontal flipping, rotation, cropping, adding noise, occlusion, and brightness adjustment. This solves the problem of insufficient data and enhances the generalization of the model. The following is a key analysis and use of several enhancement techniques:

[0034] 1) The specific process of brightness adjustment is as follows: first set different brightness factors, then enhance the brightness of the original image by calling the ImagEnhance class in the PIL module in Python, and finally obtain the image after brightness adjustment.

[0035] 2) The specific process of Gaussian noise processing is as follows: first, a Gaussian noise matrix is ​​generated, and then the generated Gaussian noise matrix is ​​added to the original image array to obtain a new image array with noise added.

[0036] 3) The specific process of occlusion processing is as follows: first, create a grayscale image and fill it with a more obvious white as an occlusion layer. Secondly, invert the RGB value of each pixel of the image to obtain an inverted image. Then, combine the original image, the inverted image and the occlusion layer together to obtain the image after occlusion processing.

[0037] Finally, the expanded data set is divided into training set and test set in a ratio of 7:3.

[0038] Step three, construct a Compton camera imaging method based on the pix2pix model.

[0039] The pix2pix model uses U-Net as the main structure of the generator, adds multi-scale feature extraction modules to the encoder and decoder, designs the RS attention module, and uses the Transformer architecture to replace the traditional convolutional network in the discriminator. The model diagram is as follows Figure 2 shown.

[0040] The generator structure has a total of five layers, and the number of feature channels from input to the bottom layer is 3, 64, 128, 256, 512, 512 respectively.

[0041] The encoder includes a convolution module, a downsampling module, a multi-scale feature extraction module, and an RS attention mechanism module. The convolution module extracts features from the current input, and then the output is connected to the multi-scale feature extraction module, which is then downsampled by the downsampling module. After each downsampling operation, a multi-scale feature extraction module is applied. Each module contains 3×3 and 5×5 convolution kernels, and uses different step sizes and pooling strategies. At the deepest level of the encoder, a larger convolution kernel 7×7 is used to capture the global context while maintaining the consistency of the multi-scale feature extraction module. Each layer of the decoder first performs an upsampling operation, then performs the operation of the RS attention module, and then merges the output of the RS attention module with the skip connection of the corresponding encoder layer, and then adds a multi-scale feature extraction module to enhance the decoder's ability to generate details.

[0042] The downsampling operation of the encoder includes convolution layer, ReLU and pooling layer; the upsampling operation of the decoder includes deconvolution, convolution and ReLU.

[0043] The encoder includes: the first layer is a 3×3 convolution layer, followed by a multi-scale feature extraction module, and then the second, third, and fourth layers, which are used to first perform downsampling operations and then extract multi-scale features. The deepest fifth layer uses a 7×7 convolution operation to capture global information. After the fifth layer is a multi-scale feature extraction module. At each jump connection of the encoder, it first passes through an RS attention module and then merges with the decoder part; in the decoder, the first, second, third, and fourth layers are first upsampled, and the output of the upsampling operation passes through the RS attention module, and then the output is merged with the jump connection of the corresponding encoder layer. A multi-scale feature extraction module is added after the RS attention module, and finally the feature map is converted into the final output image through a 1×1 convolution.

[0044] The multi-scale feature extraction module is as follows: Figure 3 As shown in the figure, it includes parallel convolution layers, batch normalization (BatchNormalization, BN) layers, ReLU layers, feature map splicing layers, and finally the spliced ​​feature maps are compressed and integrated through 1×1 convolution layers.

[0045] The parallel convolutional layer specifically includes:

[0046] Each branch uses a different convolution kernel size (3×3, 5×5, 7×7), and sets the expansion rate to 2 for the 5×5 convolution and 3 for the 7×7 convolution to further enrich the receptive field. Smaller convolution kernels extract detailed information of the image, while larger convolution kernels are used to capture a more holistic picture. A BN layer is added after each parallel convolution layer to ensure that the output of each convolution is normalized before entering the activation function and splicing, which helps to keep the model stable during training. The output feature map of each convolution path is weighted fused after splicing to combine multi-scale information. Finally, after 1×1 convolution, the ReLU activation function is applied to increase the nonlinear expression ability of the network.

[0047] The RS attention module is as follows Figure 4 As shown in the figure, it includes 3×3 convolutional layer, BN, Leaky ReLU activation function, SimAM and residual module, including:

[0048] The RS attention module first undergoes a convolution operation to extract local features, then uses batch normalization to increase the training speed, and then uses the Leaky ReLU activation function to improve the robustness of the model. Then, a lightweight attention mechanism module SimAM is introduced to enable the model to better focus on important areas. Finally, the residual module is connected to allow features to be better transmitted and learned in the deep network.

[0049] In the network model, the discriminator uses the Transformer architecture instead of the traditional convolutional network. Figure 5 This is a diagram of the discriminator structure of improving pix2pix according to an embodiment of the present invention. Figure 5As shown in the figure, the discriminator specifically includes: convolution layer, image block embedding (Patch Embedding) layer, position encoding layer, Transformer encoder layer, global average pooling layer, and fully connected layer. Among them, the input first passes through several convolution layers to extract local features, and then the convolutional feature map is divided into small blocks (patch) through the Patch Embedding layer, and each small block is mapped to a fixed embedding dimension and converted into an embedding vector. The position information is added to each small block through the position encoding layer to ensure that the spatial dependency is not lost; then through multiple Transformer encoder layers, each encoder layer includes a multi-head self-attention mechanism, LayerNorm, residual connection and feedforward neural network, multiple Transformer encoder layers process these embedded vectors to capture global dependencies; finally, the outputs of multiple Transformer encoder layers are aggregated into a global feature vector through the global average pooling layer, and then a scalar is output through the fully connected layer (FC) to judge the authenticity of the image.

[0050] In one embodiment, in model training, the training process includes 200 training epochs and the batch size is 4.

[0051] Step 4: Input the Compton camera imaging picture to be processed into the Compton camera imaging quality improvement model to obtain a high-quality Compton camera imaging picture. The Compton camera imaging quality improvement model refers to the trained model obtained using the above steps.

[0052] The first three steps can be used to obtain a trained model for improving the image quality of the Compton camera and the best trained weight file. In the subsequent use process, you only need to input the Compton camera image to be improved as the input image to obtain a high-resolution image of the radiation source point, and analyze the distribution of the radiation source, which is beneficial to subsequent operations. For example, after obtaining the distribution of gamma rays, the distribution of boron dose can be determined.

[0053] In summary, first, the radiation source generates Compton scattering and photoelectric effect on the Compton camera to generate data, and then two traditional reconstruction algorithms are used to reconstruct images respectively to obtain high-statistic images and low-statistic images, and the images are enhanced by SRGAN, and the enhanced high- and low-statistic images are merged into sample pairs, and the data enhancement technology is used to obtain the data set required for the subsequent steps; then a Compton camera imaging method based on the pix2pix model is constructed, and U-Net is used as the main structure of the generator, a multi-scale feature extraction module is added, and an attention module is designed to better generate images that fit the high statistics in the model, improve the performance of the generator, and use the Transformer architecture in the discriminator to replace the traditional convolutional network, so as to better capture the global structure and detail information of the image; finally, the trained network model is used to process the Compton camera to-be-processed image. The present invention combines the traditional reconstruction algorithm of the Compton camera with the deep learning method, expands the application scope of Compton camera imaging, and makes up for the shortcomings of the traditional algorithm. At the same time, the improved pix2pix model designed improves the quality of the reconstructed image of the Compton camera.

[0054] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope required by the present invention.

Claims

1. A method for improving the imaging quality of a Compton camera, characterized in that: The method comprises: Step 1, collecting scattering and absorption data obtained after Compton scattering and photoelectric effect of the radiation source on the Compton camera, and then screening the data; Step 2: Use the reconstruction algorithm to reconstruct the filtered data into a two-dimensional image, then make a data set based on the reconstructed two-dimensional image, and then perform data enhancement on the data set to finally obtain the enhanced data set, which is divided into a training set and a test set in a ratio of 7:3; Step 3: Construct a Compton camera imaging method based on the pix2pix model. The pix2pix model uses U-Net as the main structure of the generator, adds a multi-scale feature extraction module to the encoder and decoder, designs an RS attention module, and uses a Transformer architecture in the discriminator. The encoder includes: the first layer is a 3×3 convolution layer, followed by a multi-scale feature extraction module, followed by the second, third, and fourth layers, which are used to first perform downsampling operations and then extract multi-scale features. The deepest fifth layer uses a 7×7 convolution operation to capture global information. After the fifth layer, there is a multi-scale feature extraction module. At each jump connection of the encoder, it first passes through an RS attention module and then merges with the decoder part; in the decoder, the first, second, third, and fourth layers are first upsampled, and the output of the upsampling operation passes through the RS attention module, and then the output is merged with the jump connection of the corresponding encoder layer. A multi-scale feature extraction module is added after the RS attention module, and finally the feature map is converted into the final output image through a 1×1 convolution; The multi-scale feature extraction module includes a parallel convolution layer, a batch normalization layer, a ReLU layer, and a feature map splicing layer. Finally, the spliced ​​feature map is compressed and integrated through a 1×1 convolution layer. Each branch uses a different convolution kernel size. A smaller convolution kernel extracts detailed information of the image, while a larger convolution kernel is used to capture a more holistic picture. The RS attention module includes a 3×3 convolutional layer, a batch normalization layer, a Leaky ReLU activation function, a SimAM, and a residual module; The discriminator structure includes: a convolution layer, a small block embedding layer, a position encoding layer, a Transformer encoder layer, a global average pooling layer, and a fully connected layer. The input image is passed through several convolution layers to extract local features, and then the convolutional feature map is divided into small blocks through the small block embedding layer, and each small block is mapped to a fixed embedding dimension; then it enters the position encoding layer to add position information for each small block; then it passes through multiple Transformer encoder layers, each Transformer encoder layer includes a multi-head self-attention mechanism, layer normalization, residual connection and feedforward neural network; finally, the output of the Transformer encoder layer is aggregated into a global feature vector through a global average pooling layer, and then a scalar is output through a fully connected layer to judge the authenticity of the image.

2. The method for improving the imaging quality of a Compton camera as claimed in claim 1, characterized in that: The screening process includes filtering the data according to the detected radiation source energy to obtain data that meets the Compton scattering condition.

3. The method for improving the imaging quality of a Compton camera as claimed in claim 1, characterized in that: Step 2 includes: using analytical and iterative methods to reconstruct images of the filtered data respectively to obtain high-statistic images and low-statistic images, putting the high-statistic images and low-statistic images into SRGAN for image enhancement, and then merging the enhanced high-statistic images and low-statistic images into sample pairs, and using data enhancement technology to enhance the data set.

4. The method for improving the imaging quality of a Compton camera as claimed in claim 3, characterized in that: Step 2 includes: using the direct back projection method in the analytical method to image the filtered Compton camera data according to 1000, 4000, 8000, 12000, 40000, and 100000 data respectively to obtain low-statistic images; then using the maximum likelihood expectation maximization method in the iterative method to image with 300000 and 500000 data to obtain high-statistic images; the data enhancement technology includes at least one of the following: horizontal flipping, rotation, cropping, noise addition, occlusion, and brightness adjustment of the data.

5. The method for improving the imaging quality of a Compton camera as claimed in claim 1, characterized in that: The generator structure has a total of five layers, and the number of feature channels from input to the bottom layer is 3, 64, 128, 256, 512, 512 respectively.

6. The method for improving the imaging quality of a Compton camera as claimed in claim 1, characterized in that: The downsampling operation of the encoder includes convolution layer, ReLU and pooling layer; the upsampling operation of the decoder includes deconvolution, convolution and ReLU.

7. The method for improving the imaging quality of a Compton camera as claimed in claim 1, characterized in that: Step three also includes: training the pix2pix model. In the model training, the training process includes 200 training cycles with a batch size of 4.

Images