Electron paramagnetic resonance imaging acceleration method and system

Abstract

The invention relates to the technical field of biomedical imaging, in particular to an electron paramagnetic resonance imaging acceleration method and system, and the method comprises the steps: reconstructing an EPRI image from a projection collected from a sparse view through employing an FBP algorithm, and obtaining a first reconstructed image; the first reconstructed image is input into a pre-trained double-attention Transform network model, and a second reconstructed image is obtained; and combining the first reconstructed image and the second reconstructed image, taking the combined image as a guide condition of a reverse generation process of a conditional diffusion model, and obtaining a final reconstructed EPRI image by utilizing the reverse generation process of the conditional diffusion model. According to the scheme, the strip artifacts in the EPRI image sparsely reconstructed by the FBP algorithm can be suppressed to the maximum extent.

Description

Technical Field

[0001] The invention relates to the technical field of biomedical imaging, and in particular to an electron paramagnetic resonance imaging acceleration method and system. Background Art

[0002] Electron paramagnetic resonance imaging (EPRI) is a new metabolic information imaging modality that is suitable for imaging information such as oxygen concentration, pH value, and redox state in vivo. Its main application is precision radiotherapy based on oxygen image navigation. Studies have shown that in contrast radiotherapy experiments of mouse tumors, the survival rate of conventional radiotherapy is only 25%, while the survival rate of this precision radiotherapy is as high as 65%. At present, the commercial EPRI reconstruction algorithm is the 3D filtered back projection (FBP) algorithm, which requires a large number of projections (i.e., complete projections) to be collected over the entire angle range to reconstruct the image, which makes the scanning process very time-consuming and limits the clinical application of EPRI. An effective acceleration method is to collect only sparse projections to reconstruct the image, which is called "sparse reconstruction". However, the FBP algorithm is not suitable for sparse reconstruction because the images reconstructed from sparse projections using the algorithm usually contain severe stripe artifacts. Therefore, it is necessary to develop advanced EPRI sparse view reconstruction methods to minimize these stripe artifacts. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide an electron paramagnetic resonance imaging acceleration method and system for addressing the deficiencies in the above-mentioned prior art, so as to solve the technical problem that the current EPRI reconstruction method cannot suppress sparse artifacts.

[0004] The objective of the present invention is achieved by the following technical solutions: In a first aspect, the present invention provides an electron paramagnetic resonance imaging acceleration method, comprising: The EPRI image is reconstructed from the projections collected from the sparse views using the FBP algorithm to obtain a first reconstructed image; Inputting the first reconstructed image into a pre-trained dual-attention Transformer network model to obtain a second reconstructed image; the dual-attention Transformer network model is used to capture in-plane regional, in-plane global and cross-plane image features; The first reconstructed image and the second reconstructed image are combined and used as a guiding condition for a reverse generation process of a conditional diffusion model. The reverse generation process of the conditional diffusion model is used to obtain a final reconstructed EPRI image.

[0005] As a further improvement of the present invention, the dual-attention Transformer network model includes a pixel embedding stage, an encoding stage, a bottleneck stage, a decoding stage and a pixel de-embedding stage; the pixel embedding stage is used to embed each pixel in the first reconstructed image into a label token to obtain a label map; the encoding stage is used to capture the image features of the plane region, the plane global and the cross-plane according to the label map and perform downsampling operations; the bottleneck stage is used to further extract the global features of the feature map; the decoding stage is used to reconstruct the features of the image and perform upsampling operations. In this stage, the decoder feature map is spliced ​​with the feature map of the encoder at the same level in the channel dimension to make up for the information loss caused by downsampling layer by layer; the pixel de-embedding stage is used to embed the label map output by the decoder into a pixel map to form a residual image, and output the second reconstructed image based on the residual image and the first reconstructed image.

[0006] As a further improvement of the present invention, the dual attention Transformer network model includes a DA-Transformer block; the DA-Transformer block is respectively set in the encoding stage, the bottleneck stage, and the decoding stage to capture the image features of the in-plane region, the in-plane global and the cross-plane.

[0007] As a further improvement of the present invention, the DA-Transformer block includes a layer normalization module, a window-based in-plane self-attention module, a window-based cross-plane self-attention module and a region-enhanced feedforward network module; the feature map is normalized by the layer normalization module and then input in parallel to the window-based in-plane self-attention module and the window-based cross-plane self-attention module, and then input into the region-enhanced feedforward network module after being processed by another layer normalization module; the window-based in-plane self-attention module is used to calculate the pixel self-attention within non-overlapping windows, and use the shift window mechanism for cross-window connection; the window-based cross-plane self-attention module is used to take each non-overlapping window in the plane as a marking token, and calculate the cross-plane self-attention of the longitudinal dimension; the region-enhanced feedforward network module is used to introduce nonlinear capabilities while enhancing the ability of the feedforward network to capture multi-scale regional features.

[0008] As a further improvement of the present invention, the specific processing flow of the window-based in-plane self-attention module is as follows: In the window segmentation layer, the plane corresponding to the resolution of the feature map is segmented into a non-overlapping window of a first set size, segmenting the input feature map according to the non-overlapping window of the first set size; calculating the in-plane self-attention of the feature map after the current segmentation; In the window merging layer, the resolution of the feature map is restored according to the merging window; The output of the self-attention in the plane is:

[0009] In the formula, is a linear projection with the same input and output dimensions; It is a flattening operation on the feature space; Yes and The factor by which the dot product size is scaled.

[0010] As a further improvement of the present invention, the specific processing flow of the window-based in-plane self-attention module is as follows: In the window segmentation layer, the plane corresponding to the resolution of the feature map is segmented into a non-overlapping window of a first set size, segmenting the input feature map according to the non-overlapping window of the first set size; calculating the in-plane self-attention of the feature map after the current segmentation; In the window merging layer, the resolution of the feature map is restored according to the merging window; The output of the self-attention in the plane is:

[0011] In the formula, is a linear projection with the same input and output dimensions; It is a flattening operation on the feature space; Yes and The factor by which the dot product size is scaled.

[0012] As a further improvement of the present invention, the region enhancement feedforward network module has the following specific processing steps:

[0013] In the formula, is a linear projection operation with an output dimension four times the input dimension, is a linear projection operation with an input dimension four times the output dimension, GELU(⋅) is the GELU activation function, and HDC(⋅) represents a hybrid dilated convolution module, which consists of three deep dilated convolution cascades to capture image features in multi-scale regions within a plane.

[0014] As a further improvement of the present invention, the conditional diffusion model includes a trained Gaussian noise prediction network module, and the training process of the Gaussian noise prediction network module is: The Gaussian noise prediction network module is trained by using the first and second reconstructed images together. The RAdam optimizer with a learning rate of 1e-4 is used for optimization. The loss function is updated.

[0015] As a further improvement of the present invention, the reverse generation process of the conditional diffusion model is:

[0016] In the formula, The first reconstructed image and the second reconstructed image are used as conditional inputs. The reverse denoising process starting from ; is the first reconstructed image; is the second reconstructed image; , , is a set of pre-set variance sequences, represents the standard normal distribution, is the variance, is the time step; is the trained Gaussian noise prediction network.

[0017] In a second aspect, the present invention provides an electron paramagnetic resonance imaging acceleration system for implementing the above-mentioned electron paramagnetic resonance imaging acceleration method, comprising: A first image reconstruction module, used to reconstruct the EPRI image from the projections collected by the sparse view using the FBP algorithm to obtain a first reconstructed image; A second image reconstruction module is used to input the first reconstructed image into a pre-trained dual-attention Transformer network model to obtain a second reconstructed image; the dual-attention Transformer network model is used to capture in-plane regional, in-plane global and cross-plane image features; The reverse generation module is used to combine the first reconstructed image and the second reconstructed image, and the combined image is used as a guiding condition for the reverse generation process of the conditional diffusion model, and the reverse generation process of the conditional diffusion model is used to obtain the final reconstructed EPRI image.

[0018] In a third aspect, the present invention provides a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to perform the above-mentioned electron paramagnetic resonance imaging acceleration method.

[0019] In a fourth aspect, the present invention provides a computing device, comprising: One or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include steps for executing the above-mentioned electron paramagnetic resonance imaging acceleration method.

[0020] The beneficial effect of the present invention is that: the electron paramagnetic resonance imaging acceleration method provided by the present invention integrates the self-attention mechanism within the plane and across the plane, while enhancing the ability of the feedforward network to capture multi-scale regional features, so as to effectively capture the image features of the plane region, the plane global and across the plane. The dual-attention Transformer network model is coupled with the conditional diffusion model, and the image feature extraction and modeling capabilities of the dual-attention Transformer network model and the data distribution learning capabilities of the conditional diffusion model are integrated, taking into account the advantages of the dual-attention Transformer network model and the conditional diffusion model, and having a strong sparse artifact suppression capability. The output of the pre-trained dual-attention Transformer network is combined with the image reconstructed using FBP sparse projection to jointly guide the reverse process of the conditional diffusion model. This coupling method enables the proposed model to integrate the image feature extraction and modeling capabilities of the dual-attention Transformer network and the data distribution learning capabilities of the conditional diffusion model, thereby showing a more powerful artifact suppression and image feature retention capability in the sparse reconstruction task. The coupling of the dual-attention Transformer network and the conditional diffusion model can effectively suppress the stripe artifacts in the FBP reconstructed image under sparse view conditions.

[0021] Furthermore, the dual-attention Transformer network can help the model distinguish between real signals and noise, suppress the influence of artifacts and noise, and improve the clarity of the image. Through the pixel embedding and de-embedding stages, the multi-level encoder and decoder structure, and the combined effect of the hybrid dilated convolution and dual-attention mechanism, the model can effectively extract the local details and global semantic information of the image. The model can accurately map the extracted feature information back to the pixel space and generate a high-quality second reconstructed image.

[0022] Furthermore, the in-plane self-attention module can capture the global correlation information within the image plane, and effectively extract the 2D feature information of the image; the cross-plane self-attention module can capture the global correlation information between different image planes, and effectively extract the cross-plane 3D feature information of the image; the regional enhancement feedforward network module embeds the hybrid hole convolution module with the ability to expand the receptive field into the traditional Transformer feedforward network, effectively enhancing its ability to capture multi-scale regional features. In addition, compared with the global self-attention calculation, the window-based self-attention calculation has less computational complexity and is more suitable for processing high-resolution images. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 It is an overall structural diagram of the electron paramagnetic resonance imaging acceleration method in an embodiment of the present invention.

[0025] Figure 2 2 is a network structure diagram of a dual-attention Transformer network model in an embodiment of the present invention.

[0026] Figure 3 is a schematic diagram of the structure within the dual-attention Transformer network model in an embodiment of the present invention; (a) is a network structure diagram of the basic module DA-Transformer block in the dual-attention Transformer, (b) is a network structure diagram of the window-based in-plane self-attention module, (c) is a network structure diagram of the window-based cross-plane self-attention module, and (d) is a structure diagram of the regional enhanced feedforward network.

[0027] Figure 4 These are representative slice images reconstructed by different methods in the embodiments of the present invention.

[0028] Figure 4 In the figure, the image marked with "full view reconstruction" is reconstructed by the FBP algorithm from 208 view projections, and the rest are the results of processing the images sparsely reconstructed by the FBP algorithm from 50 views using different methods. The display window is [0, 0.0012].

[0029] Figure 5 It is a representative slice image reconstructed by the FBP algorithm, the method of the present invention and its ablation model in the embodiment of the present invention.

[0030] Figure 5 In the figure, the image marked with "complete view reconstruction" is reconstructed by the FBP algorithm from 208 view projections, and the others are the processing results of the model after the method of the present invention and the coupling mode ablation and network module ablation. The display window is [0, 0.0012]. DETAILED DESCRIPTION

[0031] In order to make the purpose and technical solution of the present invention clearer and easier to understand, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0032] Terminology explanation: FBP: Filtered back projection algorithm is a classic image reconstruction algorithm.

[0033] EPRI: electron paramagnetic resonance imaging.

[0034] DA-T: Dual Attention Transformer Network.

[0035] ReFF: Region-enhanced Feedforward Networks.

[0036] WISA: A window-based in-plane self-attention module.

[0037] WTSA: Window-based Cross-Plane Self-Attention Module.

[0038] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments, wherein the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0039] Example 1 like Figure 1-Figure 5 As shown, this embodiment provides an electron paramagnetic resonance imaging acceleration method for maximally suppressing stripe artifacts in the EPRI image sparsely reconstructed by the FBP algorithm. The specific implementation method is as follows.

[0040] First, the FBP algorithm is used to reconstruct the EPRI image from the projections collected from the sparse view to obtain the first reconstructed image .

[0041] Secondly, the first reconstructed image is input into the pre-trained dual attention Transformer network (DA-T) model to obtain the second reconstructed image ; The dual attention Transformer network model is used to capture in-plane regional, in-plane global and cross-plane image features.

[0042] The dual attention Transformer (DA-T) network in this embodiment utilizes hybrid dilated convolution and integrates in-plane and cross-plane self-attention mechanisms to effectively capture in-plane regional, in-plane global and cross-plane image features, making up for the shortcoming of the conditional diffusion model that cannot capture image features.

[0043] like Figure 2As shown, the input of the DA-T network model is the first reconstructed image x. In this embodiment, 8 adjacent slices are combined into a 3D EPRI image. The shape of the input image is 8×80×80, where the height (H) and width (W) are both 80, and the longitudinal depth (D) is 8. The entire DA-T network module consists of five stages, namely, the pixel embedding stage, the encoding stage, the bottleneck stage, the decoding stage, and the pixel de-embedding stage. The pixel embedding stage is used to embed each pixel in the first reconstructed image into a token to obtain a token map; the encoding stage is used to capture the image features of the plane region, the plane global and the cross-plane and perform downsampling operations on the token map; the bottleneck stage is used to further extract the global features and high-level features of the feature map; the decoding stage is used to perform feature reconstruction and upsampling operations on the image. In this stage, the decoder feature map is spliced ​​with the feature map of the same level encoder in the channel dimension to compensate for the information loss caused by downsampling layer by layer; the pixel de-embedding stage is used to "de-embed" the token map output by the decoder into a pixel map to form a residual image, and output a second reconstructed image based on the residual image and the first reconstructed image.

[0044] The specific operations of each stage are as follows: The pixel embedding stage uses a 2D convolution with a kernel size of 3×3 to embed each pixel into a label of dimension C (C=64), and then flattens the label map to D×(H*W)×C.

[0045] The encoding stage processes feature maps at four different levels. Each level consists of two consecutive DA-Transformer blocks. The DA-Transformer block is a basic block designed to capture 2D multi-scale regional features and implement efficient 3D self-attention to model long-range dependencies, and its input and output have the same shape. It is worth noting that the downsampling operation is limited to the H×W plane, while the longitudinal depth D remains unchanged. Therefore, each downsampling operation uses a 2D convolutional layer with a kernel size of 4×4 to reduce the plane size of the feature map and expand its channels.

[0046] The bottleneck stage also consists of two consecutive DA-Transformer blocks to further extract high-level features and model pixel-level global dependencies. To avoid spatial information loss, there is no downsampling and upsampling before and after the bottleneck stage, and only two 1×1 kernel size convolutions are used to double and halve the channel size, respectively.

[0047] The decoding stage is similar to the encoding stage, and also processes feature maps at four different levels to recover and improve features. Each level consists of two consecutive DA-Transformer blocks. Similar to downsampling, upsampling operations are also limited to the H×W plane. Each upsampling operation uses a transposed convolution layer with a 2×2 kernel size to expand the plane size of the feature map and reduce its channels. At each decoder stage, the feature map must be spliced ​​with the output of the encoder at the same level in the channel dimension to compensate for the loss of spatial information.

[0048] In the pixel de-embedding stage, the feature map is reshaped to D×2C×H×W to restore the image resolution, and the channel dimension is reduced from 2C to 1 through a 2D convolution with a kernel size of 3×3 to form a residual image (1×D×H×W). The final output image is the second reconstructed image. , composed of the residual image and the original input image (the first reconstructed image ) are added together.

[0049] The DA-Transformer block includes a layer normalization module, a window-based within-plane self-attention module, a window-based cross-plane self-attention module, and a region-augmented feedforward network module.

[0050] As shown in Figure 3 (a), the feature map is normalized by the layer normalization (LayerNorm) module and then input into the window-based in-plane self-attention module and the window-based cross-plane self-attention module in parallel; the window-based in-plane self-attention (WISA) module is used to calculate the pixel self-attention in non-overlapping windows and use the shift window mechanism for cross-window connection, so as to effectively model the long-range dependency in the plane. The window-based cross-plane self-attention (WTSA) module is used to take each non-overlapping window in the plane as a marker and calculate the cross-plane self-attention in the longitudinal dimension to model the common similarity between slices in medical images. The dual self-attention composed of WISA and WTSA in parallel realizes efficient 3D self-attention for modeling long-range dependencies in 3D space. After that, the feature map is processed again by another LayerNorm and sent to the region-enhanced feedforward network (ReFF) to capture multi-scale regional features. Among them, Figure 3 (b) shows the network structure of the WISA module. The feature map is processed in the WISA module according to the following process: In the window segmentation (Win_Partition) layer, the WISA module converts each resolution The plane is divided into The size is non-overlapping windows, so that the shape of the feature map changes from Became Then, the output of the in-plane self-attention is calculated as follows:

[0051] in, is a linear projection with the same input and output dimensions, is the 2D space of features ( ) to perform a flattening operation, Yes and The factor by which the dot product size is scaled, The calculation of self-attention in the plane divides the feature channel into In the window merging (Win_Merge) layer, the WISA module merges windows to restore the resolution of the feature map, thereby transforming the shape from becomes .

[0052] Figure 3 (c) shows the network structure of the WTSA module. The feature map is processed in the WTSA module according to the following process: In the window segmentation (Win_Partition) layer, the WTSA module divides each resolution The plane is divided into The size is non-overlapping windows, so that the shape of the feature map changes from Became Then, the output of cross-plane self-attention is calculated as follows:

[0053] in, is a linear projection with the same input and output dimensions, is the global average pooling (GAP) of pixels within the window, Is used for zooming and A learnable factor for the dot product size. The calculation of cross-plane self-attention also divides the feature channels into Each head independently computes self-attention in parallel. Then, The dimension is expanded to In the Win_Merge layer, WTSA merges windows to restore the resolution of the feature map, thereby transforming the shape from becomes .

[0054] Figure 3 (d) shows the network structure of the ReFF module, which embeds a hybrid atrous convolution module with an extended receptive field into a traditional Transformer feedforward network. The feature map is processed in the ReFF module according to the following steps:

[0055] In the formula, is a linear projection operation with an output dimension four times the input dimension, is a linear projection operation with an input dimension four times the output dimension, GELU(⋅) is the GELU activation function, and HDC(⋅) represents the hybrid dilated convolution module (HDC). HDC consists of three deep dilated convolution cascades to capture image features of multi-scale regions in a plane. In this embodiment, the dilation rates of the hybrid dilated convolution module are 1, 2, and 4, respectively. For simplicity, the corresponding formula of the ReFF module omits the reshape operation of the feature map before and after the HDC module.

[0056] Finally, the first reconstructed image and the second reconstructed image are combined and used as the guiding condition for the reverse generation process of the conditional diffusion model (CDM). The final reconstructed EPRI image is obtained by using the reverse generation process of the conditional diffusion model. .

[0057] The reverse generation process of the conditional diffusion model is:

[0058] In the formula, The first reconstructed image and the second reconstructed image are used as conditional inputs. The reverse denoising process starting from ; is the first reconstructed image; is the second reconstructed image; , , is a set of pre-set variance sequences, represents the standard normal distribution, is the variance, is the time step; is the trained Gaussian noise prediction network. It can be seen that the reverse denoising process set uses the trained Gaussian noise prediction network Output forward diffusion process at any time step The predicted value of the added Gaussian noise is then used to calculate the expected value of the distribution , and then according to the set variance Determine the data distribution.

[0059] In addition, this embodiment also provides pseudo code for generating high-quality sparse reconstructed images by conditional diffusion model sampling. As shown in Table 1, in the actual sampling process, the present invention refers to the method for generating sampling trajectories proposed in DDIM to speed up the sampling process.

[0060] Table 1: Pseudo code for generating high-quality sparse reconstructed images

[0061] This embodiment provides a Gaussian noise prediction network for training a conditional diffusion model The pseudo code is shown in Table 2. The loss function adopts loss.

[0062] Table 2: Gaussian noise prediction network training pseudo code

[0063] In this embodiment, the method of the present invention is further verified by experiments. Specifically, the physical phantom is imaged using a JIVA-25 electron paramagnetic resonance oxygen imager, with a main magnetic field strength of 250 G, a resonance frequency of 720 MHz, and a resonator volume of φ25 mm×15 mm. The physical phantom used for electron paramagnetic resonance imaging in the experiment consists of ampoules containing two concentrations of trityl (OX063-D24) solution of 1 mmol / L and 2 mmol / L, and the specifications of the ampoules used are 1 mL, 2 mL, 5 mL and 10 mL. The phantoms are divided into single-bottle phantoms, double-bottle phantoms and triple-bottle phantoms, and a total of 38 phantoms were made. The resolution of the original 3D EPRI image is , the side length of each identical cubic voxel is 0.0875 cm. Using the maximum spacing projection sorting method (MSPS), sparse view projections consisting of 50 views are extracted from the full view projections of 208 views. Pairs of 3D images are reconstructed from the full view (208 views) and sparse view (50 views) projections using the FBP algorithm. In this embodiment, by reconfiguring the software parameters of the imager, pairs of images reconstructed from the full view and sparse view projections are obtained after the full view projection scan. By slicing 38 pairs of 3D EPRI images along three spatial axes, a total of 9,120 pairs of 2D images are obtained. Subsequently, every 8 adjacent 2D slices are recombined into a small 3D EPRI image ( ), thus generating 1140 pairs of small 3D images for experiments.

[0064] Since the method of the present invention requires pre-training the DA-T network module before integrating it into the CDM for training the Gaussian noise prediction network, the 1140 pairs of small 3D EPRI images are divided into two datasets. The first dataset (380 pairs) is used for pre-training DA-T, while the second dataset (760 pairs) is further divided into a training set (646 pairs), a validation set (57 pairs), and a test set (57 pairs), which are used for training, validating, and testing the Gaussian noise prediction network in the conditional diffusion model, respectively. It is worth noting that the effective input image size of the Gaussian noise prediction network is , so the training set, validation set, and test set have 5168, 456, and 456 pairs of 2D images, respectively.

[0065] The training process of the dual attention Transformer network model and the conditional diffusion model in this embodiment is divided into two stages. In the first stage, DA-T (i.e. ) for pre-training. Specifically, this embodiment uses the AdamW optimizer, and its momentum parameter is set to , The weight decay is set to 1e-9. The training process is set to 200 epochs, and the learning rate adopts a cosine decay strategy, gradually decaying from the initial value 1e-4 to 1e-6. The input 3D image size is , the batch size is set to 1, and the window size is set to , which is set in WTSA as , the dimension of each head in the multi-head self-attention is set to 64.

[0066] In the second stage, the Gaussian noise prediction network in the conditional diffusion model (i.e., ) in the FBP reconstructed image (first reconstructed image) through sparse view projection ) and the output of the pre-trained DA-T (the second reconstructed image ). Specifically, we use the RAdam optimizer with a learning rate of 1e-4. The number of training iterations is set to 300,000, and the training batch size of the input small 3D image (8 × 80 × 80) is set to 2. It is worth noting that for the Gaussian noise prediction network, the effective image size is , so the effective 2D image training batch size is 16. Coefficient table Follow t It increases linearly from 1e-4 to 2e-2, the time step (T) is set to 1000, and the sampling step is set to 2.

[0067] In this embodiment, the performance of the method of the present invention and five deep learning methods in sparse view reconstruction in EPRI is compared. Among them, FBPConvNet is a representative based on convolutional neural network (CNN), Uformer, Restormer and LIT-Former are representatives of CNN and Transformer coupled networks, and SR3 is a representative of the existing conditional diffusion model. The single-step regression network models, including FBPConvNet, Uformer, Restormer and LIT-Former, are trained using the same strategy as DA-T. For SR3, this embodiment sets the number of training iterations to 1 million times, the number of time steps (T) to 2000, and the number of sampling steps to 100. All five comparison models are trained, verified and tested on the same data set used to train, verify and test the Gaussian noise prediction network in the conditional diffusion model.

[0068] All program codes in the experiment were run on the same computer (2 Intel Xeon E5-2620 v4CPUs, clocked at 2.10GHz, 64.0GB memory, NVIDIA GeForce RTX 3090, Windows 10), and were written in Python using the Pytorch library. All models were trained to convergence.

[0069] In order to study the difference in model performance, this embodiment selected three case images reconstructed by different models for qualitative evaluation (visual inspection), and used the Jet algorithm built into OpenCV to convert grayscale images into pseudo-color images to observe the differences more clearly. In Figure 4, "Complete View Reconstruction" represents the image reconstructed by the FBP algorithm from the complete view (208 views), and the rest are the results of processing the images sparsely reconstructed by the FBP algorithm from 50 views using different methods. In the qualitative evaluation, it can be seen that compared with the comparison model, the method of the present invention can better preserve the structural features of the reconstructed image (such as Figure 4 The white arrow in Figure 4 shows the area) and texture details (the green arrow in Figure 4 shows the area), while effectively suppressing stripe artifacts (the red arrow in Figure 4 shows the area).

[0070] In addition, this embodiment uses the normalized object error (NOE) and the Pearson correlation coefficient (PCC) to quantitatively evaluate the quality of the reconstructed image, and the formulas are defined as follows:

[0071]

[0072] in, is the image reconstructed from sparse view projection, is the image reconstructed by the FBP algorithm based on the full view projection. In this formula, express and The covariance between express The standard deviation of As shown in Table 3, the proposed method outperforms all five compared models in terms of NOE and PCC, which is consistent with our previous visual inspection results.

[0073] Table 3: Quantitative evaluation of reconstruction effects of different methods under 50 views

[0074] In this embodiment, two sets of ablation experiments are also carried out to verify the advancedness of the coupling method of the dual attention Transformer network (DA-T) and the conditional diffusion model (CDM) in the method of the present invention and the effectiveness of the core components in DA-T.

[0075] In this embodiment, three DA-T and CDM coupling modes are designed: "only pre-trained DA-T" means using only the pre-trained DA-T network without coupling CDM; "only CDM" means using only CDM without coupling the DA-T network; "the method of the present invention" refers to the DA-T and CDM coupling mode proposed by the present method. As can be seen from Table 4, the DA-T and CDM coupling mode proposed by the present method shows better sparse reconstruction performance than both the single DA-T and CDM. In terms of qualitative evaluation, compared with the images reconstructed by the ablation model, the images reconstructed by the method of the present invention have obvious advantages over the single DA-T in terms of image structure preservation (as shown in the first row of Figure 5), and have obvious advantages over the single DA-T and CDM in terms of image texture detail preservation (as shown in the second row of Figure 5).

[0076] Table 4: Quantitative evaluation of DA-T and CDM coupling ablation

[0077] This embodiment designs three ablation networks of DA-T: "deleting WTSA in DA-T" refers to a coupling model formed by using DA-T without WTSA module, "deleting WISA in DA-T" refers to a coupling model formed by using DA-T without WISA module, and "replacing ReFF with LeFF in DA-T" refers to a coupling model formed by using DA-T with LeFF module in Uformer instead of ReFF module. Except for the network structure, these three ablation models are the same as the method of the present invention. As shown in Table 5.

[0078] Table 5: Quantitative evaluation of DA-T network structure ablation

[0079] The quantitative performance of the proposed method in terms of NOE and PCC is better than that of all three ablation models, which shows the effectiveness of each core component in the pre-trained DA-T. In terms of qualitative evaluation, compared with the images reconstructed by the ablation model, the images reconstructed by the proposed method have obvious advantages in terms of image structure preservation (as shown in the third row of Figure 5) and image texture detail preservation (as shown in the fourth row of Figure 5).

[0080] The EPRI acceleration method of coupling Transformer and conditional diffusion model proposed in this embodiment improves the performance of Transformer network in sparse reconstruction tasks by designing a new dual attention Transformer (DA-T), and integrates the image feature extraction and modeling capabilities of DA-T and the data distribution learning capabilities of the conditional diffusion model by designing the coupling method of DA-T and conditional diffusion model, taking into account the advantages of DA-T network and conditional diffusion model, and has powerful sparse reconstruction capabilities. The EPRI sparse configuration experiment conducted under 50 view projections verified the significant advantages of the method proposed in the present invention in sparse artifact suppression and structure and detail preservation.

[0081] Example 2 This embodiment provides an electron paramagnetic resonance imaging acceleration system, which is used to implement the electron paramagnetic resonance imaging acceleration method in embodiment 1. The system includes: A first image reconstruction module, used to reconstruct the EPRI image from the projections collected by the sparse view using the FBP algorithm to obtain a first reconstructed image; A second image reconstruction module is used to input the first reconstructed image into a pre-trained dual-attention Transformer network model to obtain a second reconstructed image; the dual-attention Transformer network model is used to capture in-plane regional, in-plane global and cross-plane image features; The reverse generation module is used to combine the first reconstructed image and the second reconstructed image, and the combined image is used as a guiding condition for the reverse generation process of the conditional diffusion model, and the reverse generation process of the conditional diffusion model is used to obtain the final reconstructed EPRI image.

[0082] The electron paramagnetic resonance imaging acceleration system of this embodiment corresponds to the electron paramagnetic resonance imaging acceleration method of the above-mentioned embodiment 1, and thus will not be described in detail here.

[0083] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, all of which fall within the scope of protection of the present invention.

Claims

1. An electron paramagnetic resonance imaging acceleration method, characterized in that: include: The EPRI image is reconstructed from the projections collected from the sparse views using the FBP algorithm to obtain a first reconstructed image; Inputting the first reconstructed image into a pre-trained dual-attention Transformer network model to obtain a second reconstructed image; the dual-attention Transformer network model is used to capture in-plane regional, in-plane global and cross-plane image features; The first reconstructed image and the second reconstructed image are combined and used as a guiding condition for a reverse generation process of a conditional diffusion model. The reverse generation process of the conditional diffusion model is used to obtain a final reconstructed EPRI image.

2. The electron paramagnetic resonance imaging acceleration method according to claim 1, characterized in that: The dual-attention Transformer network model includes a pixel embedding stage, an encoding stage, a bottleneck stage, a decoding stage and a pixel de-embedding stage; the pixel embedding stage is used to embed each pixel in the first reconstructed image into a label token to obtain a label map; the encoding stage is used to capture the image features of the plane region, the plane global and the cross-plane according to the label map and perform downsampling operations; the bottleneck stage is used to further extract the global features of the feature map; The decoding stage is used to reconstruct image features and perform upsampling operations. In this stage, the decoder feature map is spliced ​​with the encoder feature map at the same level in the channel dimension to make up for the information loss caused by downsampling layer by layer; The pixel de-embedding stage is used to embed the label image output by the decoder into a pixel image to form a residual image, and output a second reconstructed image according to the residual image and the first reconstructed image.

3. The electron paramagnetic resonance imaging acceleration method according to claim 2, characterized in that: The dual attention Transformer network model includes DA-Transformer blocks; the DA-Transformer blocks are respectively set in the encoding stage, the bottleneck stage, and the decoding stage to capture the image features of the in-plane region, the in-plane global and the cross-plane.

4. The electron paramagnetic resonance imaging acceleration method according to claim 3, characterized in that: The DA-Transformer block includes a layer normalization module, a window-based in-plane self-attention module, a window-based cross-plane self-attention module, and a region-enhanced feedforward network module; the feature map is processed by the layer normalization module, and then input into the window-based in-plane self-attention module and the window-based cross-plane self-attention module in parallel, and then processed by another layer normalization module and input into the region-enhanced feedforward network module; the window-based in-plane self-attention module is used to calculate the pixel self-attention within non-overlapping windows, and use the shift window mechanism for cross-window connection; the window-based cross-plane self-attention module is used to take each non-overlapping window in the plane as a tag token, and calculate the cross-plane self-attention in the longitudinal dimension; The region-enhanced feedforward network module is used to introduce nonlinear capabilities while enhancing the ability of the feedforward network to capture multi-scale regional features.

5. The electron paramagnetic resonance imaging acceleration method according to claim 4, characterized in that: The specific processing flow of the window-based in-plane self-attention module is as follows: In the window segmentation layer, the plane corresponding to the resolution of the feature map is segmented into a non-overlapping window of a first set size, segmenting the input feature map according to the non-overlapping window of the first set size; calculating the in-plane self-attention of the feature map after the current segmentation; In the window merging layer, the resolution of the feature map is restored according to the merging window; The output of the self-attention in the plane is: In the formula, is a linear projection with the same input and output dimensions; It is a flattening operation on the feature space; Yes and The factor by which the dot product size is scaled.

6. The electron paramagnetic resonance imaging acceleration method according to claim 4, characterized in that: The specific processing flow of the window-based cross-plane self-attention module is as follows; In the window segmentation layer, the plane corresponding to the resolution of the feature map is divided into a non-overlapping window of a second set size, segmenting the input feature map according to the non-overlapping window of the second set size; calculating the cross-plane self-attention of the feature map after the current segmentation; In the window merging layer, the resolution of the feature map is restored according to the merging window; The output of cross-plane self-attention is: In the formula, is a linear projection with the same input and output dimensions, is the global average pooling (GAP) of pixels within the window, Is used for zooming and Learnable factor for the dot product size.

7. The electron paramagnetic resonance imaging acceleration method according to claim 4, characterized in that: The specific processing flow of the regional enhanced feedforward network module is as follows; In the formula, is a linear projection operation with an output dimension four times the input dimension, is a linear projection operation with an input dimension four times the output dimension, GELU(⋅) is the GELU activation function, and HDC(⋅) represents a hybrid dilated convolution module, which consists of three deep dilated convolution cascades to capture image features in multi-scale regions within a plane.

8. The electron paramagnetic resonance imaging acceleration method according to claim 1, characterized in that: The conditional diffusion model includes a trained Gaussian noise prediction network module, and the training process of the Gaussian noise prediction network module is: The Gaussian noise prediction network module is trained by using the first and second reconstructed images together. The RAdam optimizer with a learning rate of 1e-4 is used for optimization. The loss function is updated.

9. The electron paramagnetic resonance imaging acceleration method according to claim 8, characterized in that: The reverse generation process of the conditional diffusion model is: In the formula, The first reconstructed image and the second reconstructed image are used as conditional inputs. The reverse denoising process starting from ; is the first reconstructed image; is the second reconstructed image; , , is a set of pre-set variance sequences, represents the standard normal distribution, is the variance, is the time step; is the trained Gaussian noise prediction network.

10. An electron paramagnetic resonance imaging acceleration system, used to implement the electron paramagnetic resonance imaging acceleration method according to any one of claims 1 to 9, characterized in that: include: A first image reconstruction module, used to reconstruct the EPRI image from the projections collected by the sparse view using the FBP algorithm to obtain a first reconstructed image; A second image reconstruction module is used to input the first reconstructed image into a pre-trained dual-attention Transformer network model to obtain a second reconstructed image; the dual-attention Transformer network model is used to capture in-plane regional, in-plane global and cross-plane image features; The reverse generation module is used to combine the first reconstructed image and the second reconstructed image, and the combined image is used as a guiding condition for the reverse generation process of the conditional diffusion model, and the reverse generation process of the conditional diffusion model is used to obtain the final reconstructed EPRI image.

Images