Double-domain prior enhanced efficient low-dose CT reconstruction method and device

Abstract

The invention discloses a double-domain prior enhanced high-efficiency low-dose CT reconstruction method and device which are applied to a high-efficiency low-dose CT reconstruction model, and the method comprises the steps: converting a degenerated chordal graph collected through CT scanning into a degenerated CT; independently training an image recovery module, and processing the degraded CT through the image recovery module to obtain an enhanced CT; performing front projection on the enhanced CT to obtain an enhanced chordal graph, and calculating the enhanced chordal graph and the degraded chordal graph to obtain a mean value result; only a projection domain recovery module is trained, the mean value result is processed through the projection domain recovery module, a reconstructed chordal graph is obtained and reconstructed, and a reconstructed CT is obtained. According to the method provided by the invention, the CT reconstruction priori knowledge of the image restoration module and the data characteristics of the projection domain are utilized, and accurate reconstruction is realized under various degradation types and different degradation degrees; and then, only a projection domain recovery module is trained, rich feature representation learned in the image domain is reserved, the feature extraction capability of the projection domain is optimized, and the model performance is improved while the training cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to an efficient low-dose CT reconstruction method and device with dual-domain prior enhancement. Background Art

[0002] As one of the modern mainstream medical images, the detection method of Computed Tomography (CT) has been widely used in clinical diagnosis in various clinical fields. In CT examinations, the main goal of reducing radiation dose is to reduce the radiation dose received by patients, thereby reducing the health risks brought by radiation exposure. However, the reduction of radiation dose will inevitably affect the image quality, which may lead to an increase in image noise, loss of details, and a decline in diagnostic accuracy.

[0003] However, most current deep learning methods in the field of CT reconstruction are mainly designed for specific degradation models and are usually trained separately for a specific degree of degradation. In addition, many CT image reconstruction methods mainly focus on learning in the image domain and ignore the importance of the projection domain. Although the reconstruction methods in the image domain have achieved good results in some cases, they often have limitations when dealing with complex degradations (such as noise, undersampling, etc.), especially when the projection data cannot be fully utilized, the learning ability in the image domain is easily restricted.

[0004] Therefore, simply relying on the image domain for training and reconstruction may not be able to fully recover high-quality images. In addition, some dual-domain methods train the image domain and the projection domain simultaneously. In sparse-view and limited-view scenarios, the sparsity of the projection data may limit the model by the quality of the projection data, resulting in a decline in the reconstruction result. Summary of the Invention

[0005] For this reason, the purpose of the present invention is to solve at least to some extent the deficiencies in the prior art, and thus propose an efficient low-dose CT reconstruction method with dual-domain prior enhancement.

[0006] In a first aspect, the present invention provides an efficient low-dose CT reconstruction method with dual-domain prior enhancement, which is applied to an efficient low-dose CT reconstruction model. The efficient low-dose CT reconstruction model includes an image restoration module and a projection domain restoration module. The method includes:

[0007] Converting the degraded chord diagram collected by CT scanning into a degraded CT image;

[0008] Train the image restoration module alone, and perform preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image. Wherein, the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features;

[0009] Perform forward projection on the enhanced CT image to obtain an enhanced chordogram, and calculate the mean value after adding the pixel values of the enhanced chordogram and the degraded chordogram element by element to obtain a mean result;

[0010] Freeze the image restoration module and only train the projection domain restoration module. Process the mean result through the projection domain restoration module to obtain a reconstructed chordogram, and then reconstruct the reconstructed chordogram to obtain a reconstructed CT.

[0011] In a second aspect, the present invention provides a high-efficiency low-dose CT reconstruction device with dual-domain prior enhancement, which is applied to a high-efficiency low-dose CT reconstruction model. The high-efficiency low-dose CT reconstruction model includes an image restoration module and a projection domain restoration module. The device includes:

[0012] A conversion module: used to convert the degraded chordogram collected by CT scanning into a degraded CT image;

[0013] An image restoration module: used to train the image restoration module alone, and perform preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image. Wherein, the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features;

[0014] A combination module: used to perform forward projection on the enhanced CT image to obtain an enhanced chordogram, and calculate the mean value after adding the pixel values of the enhanced chordogram and the degraded chordogram element by element to obtain a mean result;

[0015] A projection domain restoration module: used to freeze the image restoration module and only train the projection domain restoration module. Process the mean result through the projection domain restoration module to obtain a reconstructed chordogram, and then reconstruct the reconstructed chordogram to obtain a reconstructed CT.

[0016] In a third aspect, the present invention provides an efficient low-dose CT reconstruction device with dual-domain prior enhancement, including a memory, a processor, and a computer program stored in the memory and executable on the processor. It is characterized in that when the processor executes the computer program, each step in the efficient low-dose CT reconstruction method with dual-domain prior enhancement as described in the first aspect is implemented.

[0017] In a fourth aspect, the present invention further provides a storage medium with a computer program stored thereon. When the computer program is executed, each step in the efficient low-dose CT reconstruction method with dual-domain prior enhancement as described in the first aspect is implemented.

[0018] The present invention provides an efficient low-dose CT reconstruction method and device, which are applied to an efficient low-dose CT reconstruction model. The efficient low-dose CT reconstruction model includes an image restoration module and a projection domain restoration module. The method includes: converting a degraded chordogram collected by CT scanning into a degraded CT image; separately training the image restoration module, and performing preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image. Among them, the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features; performing forward projection on the enhanced CT image to obtain an enhanced chordogram, and calculating the mean value after adding the pixel values of the enhanced chordogram and the degraded chordogram element by element to obtain a mean result; freezing the image restoration module and only training the projection domain restoration module, processing the mean result through the projection domain restoration module to obtain a reconstructed chordogram, and then reconstructing the reconstructed chordogram to obtain a reconstructed CT. Through the method provided by the present invention, by using the CT reconstruction prior knowledge of the trained image restoration module and the data characteristics of the projection domain, more accurate reconstruction can be achieved under various degradation types and different degradation degrees. Moreover, by decoupling the high and low frequency features of the degraded CT image to extract features, the efficient low-dose CT model can selectively increase the weight of the information-rich frequency band according to different degradation types and intensities, thereby enhancing the reconstruction effect and accuracy; subsequently, only training the projection domain restoration module can retain the rich feature representations learned in the image domain while focusing on optimizing the feature extraction ability of the projection domain, thereby improving the model performance while reducing the training cost. Description of the Drawings

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.

[0020] Figure 1 Schematic flowchart of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0021] Figure 2 Network framework diagram of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0022] Figure 3 Schematic sub-flowchart of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0023] Figure 4 In (a) is the network framework diagram of the image restoration module of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0024] Figure 4 In (b) is the network framework diagram of the adaptive frequency learning layer of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0025] Figure 5 Another schematic sub-flowchart of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0026] Figure 6 Another schematic sub-flowchart of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0027] Figure 7 Another schematic sub-flowchart of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0028] Figure 8 Network framework diagram of the projection domain restoration module of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0029] Figure 9 Another schematic sub-flowchart of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention;

[0030] Figure 10 Schematic diagram of the program modules of the efficient low-dose CT reconstruction device with dual-domain prior enhancement of the present invention. Detailed implementation manners

[0031] To make the objectives, features, and advantages of the present invention more apparent and understandable, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present invention.

[0032] Please refer to Figure 1 and Figure 2 , Figure 1 which is a schematic flowchart of the efficient low-dose CT reconstruction method with dual-domain prior enhancement in the embodiments of the present application. Figure 2 is a network framework diagram of the efficient low-dose CT reconstruction method with dual-domain prior enhancement of the present invention. In this embodiment, the above-mentioned efficient low-dose CT reconstruction method with dual-domain prior enhancement is applied to an efficient low-dose CT reconstruction model, and the efficient low-dose CT reconstruction model includes an image restoration module and a projection domain restoration module, including:

[0033] Step 101: Convert the degraded chordogram collected by CT scanning into a degraded CT image.

[0034] In this embodiment, the degraded chordogram is the projection data collected by CT scanning. First, the collected projection data - the degraded chordogram is obtained through a CT reconstruction algorithm to get a degraded CT image, and then the efficient low-dose CT reconstruction model can perform subsequent CT reconstruction based on the degraded CT image.

[0035] Step 102: Independently train the image restoration module, and perform preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image. Among them, the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features.

[0036] In this embodiment, first independently train the image restoration module, and perform preliminary image enhancement on the degraded CT image through the image restoration module. And in the process of performing preliminary image enhancement processing on the degraded CT image by the image restoration module, it also includes decoupling the high-frequency features and low-frequency features of the degraded CT image and then performing feature extraction respectively, and finally fusing the high-frequency features and low-frequency features. In this way, the efficient low-dose CT reconstruction model can selectively increase the weight of the information-rich frequency band according to the different degradation types and intensities, thereby enhancing the reconstruction effect and accuracy.

[0037] Further, please refer to Figure 3 and Figure 4 ,Figure 3 It is a schematic diagram of a sub - process of the efficient low - dose CT reconstruction method with dual - domain prior enhancement in the embodiments of the present application. Figure 4 Among them, (a) is a network framework diagram of the image restoration module of the efficient low - dose CT reconstruction method with dual - domain prior enhancement in the embodiments of the present application. In this embodiment, the degraded CT image is preliminarily enhanced by the image restoration module to obtain the enhanced CT image. The image restoration module consists of four - layer encoders and decoders. Among them, each layer of the encoder adopts an adaptive frequency learning layer and multiple Transformer blocks, and the decoder includes multiple Transformer blocks. Specifically as follows:

[0038] Step 201: Process the degraded CT image through the adaptive frequency learning layer to obtain a fused feature.

[0039] Step 202: Then input the fused feature into the Transformer block for feature extraction, and add a down - sampling operation in each layer of the encoder to convert the fused feature into a lower - resolution latent representation.

[0040] Step 203: Input the lower - resolution latent representation into the decoder, and gradually restore the lower - resolution latent representation through the Transformer block, up - sampling operation and convolution operation in sequence to obtain the enhanced CT image.

[0041] In this embodiment, the image restoration module consists of a four - level encoder and decoder network. In each layer of the encoder, an adaptive frequency learning layer (AFLB) and multiple Transformer blocks (TB) are adopted to gradually capture the local and global features of the degraded CT image through the local self - attention mechanism, and convert the high - resolution features into lower - resolution latent representations step by step. The decoder layer also consists of multiple Transformer blocks, and the lower - resolution latent representation is gradually restored through the up - sampling operation and convolution operation to obtain the enhanced CT image. Among them, the latent representation refers to compressing the image data into a low - dimensional latent space through a certain method for more efficient processing and analysis. For example, a standard 256x256 - pixel color image contains approximately 65,000 data points, and the computational cost of these data points for a deep - learning model is very high. To improve computational efficiency, the generative model converts these original data into a low - dimensional latent space representation (latent image) through the encoding process, which makes subsequent generation, modification or other operations more efficient.

[0042] Specifically, the specific processing steps of the image restoration module are as follows:

[0043] The degraded CT image is input into the adaptive frequency learning layer for processing, where the high-frequency features and low-frequency features of the degraded CT image can be decoupled. Then, different features of the high-frequency features and low-frequency features are mined, and finally, the high-frequency features and low-frequency features are fused to obtain fused features. The obtained fused features are then input into the Transformer block in the encoder for feature extraction, and downsampling operations are added in each layer to convert the fused features into a lower-resolution latent representation. Finally, the lower-resolution latent representation is input into the decoder, and the lower-resolution latent representation is gradually restored through the TB block, upsampling operation, and convolution operation in the decoder to obtain the enhanced CT image.

[0044] Further, please refer to Figure 4 and Figure 5 , Figure 4 where (b) in Figure 5 is the network framework diagram of the adaptive frequency learning layer of the dual-domain prior enhanced high-efficiency low-dose CT reconstruction method in the embodiment of the present application.

[0045] Step 301: Extract the first shallow-layer features from the degraded CT image through a 3×3 convolutional layer.

[0046] Step 302: Perform Fourier transform and learnable Gaussian band-pass filtering on the degraded CT image after convolutional processing, and multiply the Fourier-transformed degraded CT image by the mask elements generated by the learnable Gaussian band-pass filtering to separate the high-frequency features and the low-frequency features.

[0047] Step 303: Perform inverse Fourier transform on the high-frequency features and the low-frequency features respectively, and then perform cross-attention calculation on the inverse Fourier-transformed high-frequency features and low-frequency features with the first shallow-layer features respectively to obtain multiple first features.

[0048] Step 304: Modulate the high-frequency features and the low-frequency features that have undergone cross-attention calculation with the first shallow-layer features respectively through spatial attention and channel attention, and add the modulated high-frequency features and low-frequency features element by element to fuse the high-frequency features and the low-frequency features.

[0049] Step 305: Generate the first fused feature from the fused high-frequency features and low-frequency features through a 1×1 convolutional layer, and perform cross-attention calculation on the first fused feature and the first shallow-layer features to obtain the fused features.

[0050] In this embodiment, this step is the processing process of the degraded CT image by the adaptive frequency learning layer, and the specific process is as follows:

[0051] Extract the first shallow features of the degraded CT image through a convolutional layer with a size of 3, which can convert the degraded CT image into a feature layer with a higher dimension to obtain more information; then perform a Fourier transform on the convolved degraded CT image, and multiply it by the mask elements generated by the learnable Gaussian band-pass filter to separate the high-frequency features and low-frequency features, and perform inverse Fourier transforms on the high-frequency features and low-frequency features respectively.

[0052] Then, perform cross-attention calculations on the high-frequency features and low-frequency features after the inverse Fourier transform with the first shallow features respectively to guide the excavation of different first features; the low-frequency features and high-frequency features are respectively modulated through channel attention and spatial attention, and then added element by element to fuse the high-frequency features and low-frequency features, and then generate the first fusion feature through a convolutional layer with a size of 1; in order to further optimize the feature fusion process, perform cross-attention calculations on the first fusion feature and the first shallow features to obtain the fusion feature.

[0053] Through the above steps, the high-efficiency low-dose CT reconstruction model can adaptively adjust the content of high-frequency and low-frequency features according to the degradation type in the input degraded CT image, so as to realize the reconstruction of degraded CT images of multiple degradation types.

[0054] Among them, Figure 4 Upsample in is upsampling, Downsample is downsampling, matrix multiplication is matrix multiplication, Element-wise Addition is element-wise addition, Element-wise subtraction is element-wise subtraction, and Element-wise Multiplication is element-wise multiplication.

[0055] Further, please refer to Figure 6 , Figure 6 is another sub-process schematic diagram of the high-efficiency low-dose CT reconstruction method with dual-domain prior enhancement in the embodiment of the present application. In this embodiment, the step of performing preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image further includes:

[0056] Step 401: Calculate the mean square error between the enhanced CT image and a preset restored image to obtain a first calculation result, and use the first calculation result as the image domain reconstruction supervision loss;

[0057] Step 402: Iteratively optimize the high-efficiency low-dose CT reconstruction model using the image domain reconstruction supervision loss.

[0058] In this embodiment, the image restoration module is trained separately. Therefore, the degraded CT images to be trained and the preset restored images are used as samples. The mean square error is calculated between the enhanced CT image generated by the degraded CT image passing through the image restoration module and the preset restored image. The obtained first calculation result is used as the image and reconstruction supervision loss, that is, the mean square error is calculated between the actually obtained enhanced CT image and the preset restored image that can be obtained, so as to obtain the actually generated error. Then, the obtained image domain reconstruction supervision loss is input into the image restoration module in the high-efficiency low-dose CT reconstruction model for optimization, so as to iteratively optimize the high-efficiency low-dose CT reconstruction model, and thus the enhanced CT image obtained in the image restoration module of the high-efficiency low-dose CT reconstruction model can be closer to the preset restored image sampled as the training sample.

[0059] Step 103: Perform forward projection on the enhanced CT image to obtain an enhanced chordogram, and add the pixel values of the enhanced chordogram and the degraded chordogram element by element and then calculate the mean to obtain a mean result.

[0060] In this embodiment, forward projection (FP) of the enhanced CT image can obtain an enhanced chordogram, and then add the pixel values of the obtained enhanced chordogram and the degraded chordogram Figure 2 and calculate the mean of the two to obtain a mean result. The obtained mean result here will be input into the projection domain restoration module in Step 104 for processing, so as to retain the rich feature representations already learned in the image restoration module, and at the same time focus on optimizing the feature extraction ability of the projection domain restoration module, thereby improving the performance of the high-efficiency low-dose CT reconstruction model while reducing the training cost.

[0061] Step 104: Freeze the image restoration module and only train the projection domain restoration module. Process the mean result through the projection domain restoration module to obtain a reconstructed chordogram, and then reconstruct the reconstructed chordogram to obtain a reconstructed CT.

[0062] In this embodiment, freeze the weights of the image restoration module and only train the projection domain restoration module. Input the mean result obtained in Step 103 into the projection domain restoration module for processing to obtain a reconstructed chordogram, and then reconstruct the reconstructed chordogram through the filtered back-projection algorithm to obtain a reconstructed CT.

[0063] Further, please refer to Figure 7 and Figure 8 , Figure 7 which is another sub-process schematic diagram of the dual-domain prior enhanced high-efficiency low-dose CT reconstruction method in the embodiments of the present application.Figure 8 This is the network framework diagram of the projection domain recovery module of the dual-domain prior enhanced efficient low-dose CT reconstruction method in the embodiments of the present application. In this embodiment, the mean result is processed by the projection domain recovery module to obtain a reconstructed chordogram. The projection domain recovery module consists of three layers of encoders and decoders, and each layer of encoder and decoder includes a Transformer block, specifically as follows:

[0064] Step 501: Extract the second shallow feature in the mean result through a 3×3 convolutional layer;

[0065] Step 502: Pass the second shallow feature through three Transformer blocks in the encoder respectively to extract the second feature, the third feature, and the fourth feature, and add a downsampling operation to each layer of the encoder;

[0066] Step 503: Perform an upsampling operation on the fourth feature, splice it with the third feature to obtain a first spliced feature, process the first spliced feature through a 1×1 convolutional layer, and then input it into the Transformer block in the decoder to extract the fifth feature;

[0067] Step 504: Perform an upsampling operation on the fifth feature, splice it with the second feature to obtain a second spliced feature, then pass the second spliced feature through two Transformer blocks to extract features, and then generate a sixth feature through a 3×3 convolutional layer convolution. After adding the sixth feature to the mean result element by element, the reconstructed chordogram is obtained.

[0068] In this embodiment, the above steps are the network structure of the projection domain recovery module, which consists of three layers of encoder and decoder networks. On the premise of ensuring the feature extraction ability, it effectively reduces the computational complexity of the efficient low-dose CT reconstruction model and makes it more adaptable to the special requirements of the projection data reconstruction task.

[0069] Specifically, the mean result obtained by adding the pixel values of the enhanced CT image and the degraded chord Figure 2 and then taking the mean is input into the projection domain recovery module, which can retain the rich feature representations learned in the image recovery module while focusing on optimizing the feature extraction ability in the projection domain, thereby improving the model performance while reducing the training cost. The second shallow feature is extracted through a convolutional layer of size 3, and then the encoder composed of three TB blocks is used to learn multi-scale features and local-global features. The decoder gradually restores the image details during the process of upsampling step by step, so as to obtain the reconstructed chordogram.

[0070] Among them, Figure 8The Depth-wise Convclution in it is depth convolution, Upsample is upsampling, Downsample is downsampling, Element-wise Addition is element-wise addition, and Concatenation is concatenation.

[0071] Further, refer to Figure 9 , Figure 9 Figure 9 is a schematic diagram of another sub-process of the efficient low-dose CT reconstruction method with dual-domain prior enhancement in the embodiments of the present application. In this embodiment, the mean result is processed by the projection domain recovery module to obtain a reconstructed chordogram, and then the reconstructed chordogram is reconstructed to obtain a reconstructed CT, which further includes:

[0072] Step 601: Calculate the mean square error of the reconstructed chordogram, the reconstructed CT, the preset recovered reconstructed chordogram, and the preset recovered reconstructed CT respectively to obtain a second calculation result, and use the second calculation result as the projection domain reconstruction supervision loss;

[0073] Step 602: Iteratively optimize the efficient low-dose CT reconstruction model with the projection domain reconstruction supervision loss.

[0074] In this embodiment, the projection domain recovery module is also trained separately. Therefore, the mean square error of the actually obtained reconstructed grand chart and the preset recovered reconstructed grand chart as the training sample is calculated. For the entire efficient low-dose reconstructed CT model, the mean square error of the actually obtained reconstructed CT and the preset recovered reconstructed CT as the training sample also needs to be calculated. The error result between the two is used as the projection domain reconstruction supervision loss, and the obtained projection domain reconstruction supervision loss is input into the projection domain recovery module in the efficient low-dose CT reconstruction model for optimization, so as to iteratively optimize the efficient low-dose CT reconstruction model.

[0075] Further, the specific implementation steps of the efficient low-dose CT reconstruction method with dual-domain prior enhancement in the embodiments of the present application are as follows:

[0076] 1. Train the image recovery module separately, and perform preliminary image enhancement on the degraded CT image through the image recovery module. Among them, the image recovery module consists of four-level encoder-decoder.

[0077] 1-1: Extract the first shallow features of the degraded CT image through the convolutional layer;

[0078] 1-2: Input the convolutional degraded CT image into the adaptive frequency learning module in the encoding layer for processing, and perform combined calculation with the first shallow features during the processing to obtain combined features;

[0079] 1-3. Input the fused features into the TB block in the encoder for further feature extraction, and add a downsampling operation to obtain a lower resolution potential representation;

[0080] 1-4. The lower-resolution potential representation is input into the decoder, and the low-resolution potential representation is gradually restored through TB blocks, upsampling, and convolution operations to obtain an enhanced CT image.

[0081] 2. Frozen image recovery module, train the projection domain recovery module separately, forward project the enhanced CT image to obtain the enhanced chord graph, and compare the enhanced chord graph with the degraded chord Figure 2 The pixel values of the two pixels are added together and the average result is input into the projection domain recovery module. The projection domain recovery module includes a three-layer encoder and a decoder.

[0082] 2-1. Extract the second shallow feature of the mean result through the convolution layer;

[0083] 2-2. The encoder composed of three layers of TB blocks learns multi-scale features and local-global features. The decoder gradually restores image details in the process of step-by-step upsampling to obtain the reconstructed chord diagram.

[0084] 3. Reconstruct the reconstructed chordogram into reconstructed CT using a filtered back-projection algorithm.

[0085] In summary, this application proposes a novel adaptive frequency learning module and a high-efficiency low-dose reconstruction CT module for pre-trained image restoration. The model combines dual-domain prior enhancement of image domain features and projection domain features, and the prior knowledge of the trained image restoration module assists the subsequent image restoration in the projection domain. By designing a model that can learn high- and low-frequency decoupling, feature mining, and fusion, it is possible to solve CT reconstruction tasks of various degradation types and different degrees of degradation, reduce computing and storage costs, and enable the model to adaptively adjust the weights of different frequency sub-bands, enhance the model's ability to distinguish various forms of degraded images, and utilize the CT reconstruction prior knowledge of the trained image restoration module and the data characteristics of the projection domain, effectively overcoming the shortcomings of the single-domain model, thereby improving the reconstruction quality, and significantly improving the performance with only a small increase in computing costs.

[0086] An embodiment of the present application provides an efficient low-dose CT reconstruction method enhanced by dual-domain priors, which is applied to an efficient low-dose CT reconstruction model. The efficient low-dose CT reconstruction model includes an image restoration module and a projection domain restoration module. The method includes: converting a degraded chord diagram collected by CT scanning into a degraded CT image; separately training the image restoration module, and performing preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image. Wherein, the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features; performing forward projection on the enhanced CT image to obtain an enhanced chord diagram, and calculating the mean value after adding the pixel values of the enhanced chord diagram and the degraded chord diagram element by element to obtain a mean result; freezing the image restoration module, only training the projection domain restoration module, processing the mean result through the projection domain restoration module to obtain a reconstructed chord diagram, and then reconstructing the reconstructed chord diagram to obtain a reconstructed CT. Through the method provided by the present invention, by using the CT reconstruction prior knowledge of the trained image restoration module and the data characteristics of the projection domain, more accurate reconstruction can be achieved under various degradation types and different degradation degrees. And by decoupling the high and low frequency features of the degraded CT image to extract features, the efficient low-dose CT model can selectively increase the weight of the information-rich frequency band according to different degradation types and intensities, thereby enhancing the reconstruction effect and accuracy; subsequently, only training the projection domain restoration module can retain the rich feature representations learned in the image domain, while focusing on optimizing the feature extraction ability of the projection domain, thereby improving the model performance while reducing the training cost.

[0087] Furthermore, the present application also provides a dual-domain prior enhanced efficient low-dose CT reconstruction device 700. Figure 10 It is a schematic diagram of the program modules of the dual-domain prior enhanced efficient low-dose CT reconstruction device in the embodiment of the present application. In this embodiment, the above-mentioned dual-domain prior enhanced efficient low-dose CT reconstruction device 700 includes:

[0088] A conversion module 701: configured to convert a degraded chord diagram collected by CT scanning into a degraded CT image;

[0089] An image restoration module 702: configured to separately train the image restoration module, and perform preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image. Wherein, the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features;

[0090] Combined module 703: It is used to perform forward projection on the enhanced CT image to obtain an enhanced chordogram, add the pixel values of the enhanced chordogram and the degraded chordogram element by element, and then calculate the mean value to obtain a mean result;

[0091] Projection domain recovery module 704: It is used to freeze the image recovery module, only train the projection domain recovery module, process the mean result through the projection domain recovery module to obtain a reconstructed chordogram, and then reconstruct the reconstructed chordogram to obtain a reconstructed CT.

[0092] The embodiment of the present application provides a high-efficiency low-dose CT reconstruction device 700 with dual-domain prior enhancement, which is applied to a high-efficiency low-dose CT reconstruction model. The high-efficiency low-dose CT reconstruction model includes an image recovery module and a projection domain recovery module, and can achieve: converting a degraded chordogram collected by CT scanning into a degraded CT image; separately training the image recovery module, and performing preliminary image enhancement processing on the degraded CT image through the image recovery module to obtain the enhanced CT image. Among them, the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features; performing forward projection on the enhanced CT image to obtain an enhanced chordogram, adding the pixel values of the enhanced chordogram and the degraded chordogram element by element, and then calculating the mean value to obtain a mean result; freezing the image recovery module, only training the projection domain recovery module, processing the mean result through the projection domain recovery module to obtain a reconstructed chordogram, and then reconstructing the reconstructed chordogram to obtain a reconstructed CT. Through the method provided by the present invention, by using the CT reconstruction prior knowledge of the trained image recovery module and the data characteristics of the projection domain, more accurate reconstruction can be achieved under various degradation types and different degradation degrees. Moreover, by decoupling the high and low frequency features of the degraded CT image to extract features, the high-efficiency low-dose CT model can selectively increase the weight of the information-rich frequency band according to different degradation types and intensities, thereby enhancing the reconstruction effect and accuracy; subsequently, only training the projection domain recovery module can retain the rich feature representations learned in the image domain, while focusing on optimizing the feature extraction ability of the projection domain, thereby improving the model performance while reducing the training cost.

[0093] Furthermore, the present application also provides a high-efficiency low-dose CT reconstruction device with dual-domain prior enhancement, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it realizes each step in the above-mentioned high-efficiency low-dose CT reconstruction method with dual-domain prior enhancement.

[0094] Furthermore, the present application also provides a storage medium, on which a computer program is stored. When the computer program is executed by a processor, each step in the above-mentioned efficient low-dose CT reconstruction method with dual-domain prior enhancement is implemented.

[0095] In each embodiment of the present invention, the various functional modules can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The above-mentioned integrated module can be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0096] Based on such an understanding, the technical solution of the specification of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0097] It should be noted that for the foregoing method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention. In the above embodiments, the descriptions of the various embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0098] For those skilled in the art, according to the idea of the embodiments of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. An efficient low-dose CT reconstruction method enhanced by dual-domain prior, characterized in that Applied to an efficient low-dose CT reconstruction model, the efficient low-dose CT reconstruction model includes an image restoration module and a projection domain restoration module, and the method includes: Converting the degraded chordogram collected by CT scanning into a degraded CT image; Individually training the image restoration module, and performing preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image. Wherein, the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features; Performing forward projection on the enhanced CT image to obtain an enhanced chordogram, adding the pixel values of the enhanced chordogram and the degraded chordogram element by element, and obtaining a mean result; Freezing the image restoration module and only training the projection domain restoration module, processing the mean result through the projection domain restoration module to obtain a reconstructed chordogram, and then reconstructing the reconstructed chordogram to obtain a reconstructed CT.

2. The method according to claim 1, wherein For the preliminary image enhancement processing of the degraded CT image through the image restoration module to obtain the enhanced CT image, the image restoration module is composed of four layers of encoders and decoders. Among them, each layer of encoder adopts an adaptive frequency learning layer and multiple Transformer blocks, and the decoder includes multiple Transformer blocks, specifically as follows: Processing the degraded CT image through the adaptive frequency learning layer to obtain a fused feature; Then inputting the fused feature into the Transformer block for feature extraction, and adding a downsampling operation in each layer of the encoder to convert the fused feature into a lower-resolution latent representation; Inputting the lower-resolution latent representation into the decoder, and gradually restoring the lower-resolution latent representation through the Transformer block, upsampling operation and convolution operation in sequence to obtain the enhanced CT image.

3. According to the method described in claim 2, the processing of the degraded CT image after convolution processing through the adaptive frequency learning layer to obtain a fused feature includes: Extracting first shallow features from the degraded CT image through a 3×3 convolutional layer; Performing Fourier transform and learnable Gaussian band-pass filtering processing on the degraded CT image after convolution processing respectively, and multiplying the Fourier-transformed degraded CT image by the mask elements generated by learnable Gaussian band-pass filtering to separate the high-frequency features and the low-frequency features; Performing inverse Fourier transform on the high-frequency features and the low-frequency features respectively, and then performing cross-attention calculation on the inverse Fourier-transformed high-frequency features and low-frequency features and the first shallow features respectively to obtain multiple first features; The high-frequency features and the low-frequency features after cross-attention calculation with the first shallow features are respectively modulated by spatial attention and channel attention, and the modulated high-frequency features and low-frequency features are added element-wise to fuse the high-frequency features and the low-frequency features; The fused high-frequency features and low-frequency features are passed through a 1×1 convolutional layer to generate first fused features, and cross-attention calculation is performed on the first fused features and the first shallow features to obtain the fused features.

4. The method according to claim 1, wherein The mean result is processed by the projection domain recovery module to obtain a reconstructed chord diagram. The projection domain recovery module consists of three layers of encoders and decoders, and each layer of encoder and decoder includes a Transformer block, specifically as follows: Extract second shallow features in the mean result through a 3×3 convolutional layer; The second shallow features are successively passed through three Transformer blocks in the encoder to extract second features, third features, and fourth features respectively, and downsampling operations are added in each layer of the encoder; An upsampling operation is performed on the fourth features, and they are concatenated with the third features to obtain first concatenated features. After the first concatenated features are processed through a 1×1 convolutional layer, they are input into the Transformer block in the decoder to extract fifth features; An upsampling operation is performed on the fifth features, and they are concatenated with the second features to obtain second concatenated features. Then the second concatenated features are successively passed through two Transformer blocks to extract features, and then convolved through a 3×3 convolutional layer to generate sixth features. After the sixth features are added element-wise to the mean result, the reconstructed chord diagram is obtained.

5. The method according to claim 1, wherein The reconstructed chord diagram is reconstructed through a filtered back-projection algorithm to obtain a reconstructed CT.

6. The method according to claim 1, characterized in that, Applying the degraded CT image to the image restoration module for preliminary image enhancement processing to obtain the enhanced CT image further includes: Calculating the mean square error between the enhanced CT image and a preset restored image to obtain a first calculation result, and using the first calculation result as the image domain reconstruction supervision loss; Iteratively optimizing the high-efficiency low-dose CT reconstruction model with the image domain reconstruction supervision loss.

7. The method according to claim 1, wherein Processing the mean result through the projection domain recovery module to obtain a reconstructed chord diagram, and then reconstructing the reconstructed chord diagram to obtain a reconstructed CT further includes: Calculating the mean square error between the reconstructed chord diagram, the reconstructed CT and a preset restored reconstructed chord diagram, a preset restored reconstructed CT respectively to obtain a second calculation result, and using the second calculation result as the projection domain reconstruction supervision loss; Iteratively optimizing the high-efficiency low-dose CT reconstruction model with the projection domain reconstruction supervision loss.

8. An efficient low-dose CT reconstruction device enhanced by dual-domain prior, characterized in that, Applied to a high-efficiency low-dose CT reconstruction model, the high-efficiency low-dose CT reconstruction model includes an image restoration module and a projection domain recovery module, and the device includes: A conversion module: used to convert the degraded chord diagram collected by CT scanning into a degraded CT image; Image restoration module: used to train the image restoration module separately, perform preliminary image enhancement processing on the degraded CT image through the image restoration module to obtain the enhanced CT image, wherein the preliminary image enhancement processing process includes decoupling the high-frequency features and low-frequency features of the degraded CT image, respectively extracting multiple first features of the high-frequency features and the low-frequency features, and then fusing the high-frequency features and the low-frequency features; Combination module: used to perform forward projection on the enhanced CT image to obtain an enhanced chordogram, add the pixel values of the enhanced chordogram and the degraded chordogram element by element and then take the mean to obtain a mean result; Projection domain restoration module: used to freeze the image restoration module and only train the projection domain restoration module, process the mean result through the projection domain restoration module to obtain a reconstructed chordogram, and then reconstruct the reconstructed chordogram to obtain a reconstructed CT.

9. An efficient low-dose CT reconstruction device with dual-domain prior enhancement, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step in the efficient low-dose CT reconstruction method with dual-domain prior enhancement as described in any one of claims 1-7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step in the efficient low-dose CT reconstruction method with dual-domain prior enhancement as described in any one of claims 1-7.

Images