Zero-sample low-dose CT image denoising method and device based on strip diffusion model
Through the zero-sample low-dose CT image denoising method based on the band diffusion model, the limitation of the denoising of low-dose CT image in the prior art requires pairing normal dose CT images, and effective denoising under different scanning conditions is achieved. Significantly improves image quality and detail retention.
Patent Information
- Application Number
- CN202411245470.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-09-06
AI Technical Summary
The existing low-dose CT image denoising methods require paired normal-dose CT images for model training, and ignore imaging differences under different scanning protocols, resulting in image quality reduction and loss of details, making it difficult to meet the requirements of medical diagnosis.
Using the zero-sample low-dose CT image denoising method based on the band diffusion model, the forward diffusion structure and reverse inference structure are built, and the noise is gradually added and removed, and the UNet model is used to segment and use the UNet model to By splicing out the noise denoising band, training the model only requires a normal dose CT image, reducing the difficulty of model training and retaining image details.
Denoising low-dose CT images under different thicknesses and dose conditions is achieved, reducing the complexity of model training, retaining the complex structure and details of the image, and achieving a denoising effect close to the normal dose CT images, which is better than the existing ones. method.
Smart Images

Figure CN118761929B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical CT image processing, and in particular to a zero-sample low-dose CT image denoising method and device based on a strip diffusion model. Background Art
[0002] Low-dose computed tomography (CT) can reduce the risk of patient exposure to radiation and extend the service life of x-ray tubes. However, artifacts and noise in low-dose CT may reduce image quality and potentially affect subsequent diagnosis. In recent years, supervised image post-processing methods have been widely studied due to their effectiveness in noise reduction. Low-dose CT image denoising methods and systems based on the rapid diffusion model require paired low-dose and normal-dose CT images to complete model training. However, clinical conditions for obtaining paired normal-dose and low-dose CT images are very harsh, which limits the practical application and widespread adoption of supervised methods.
[0003] In addition, the clinical environment requires CT images to be scanned at different thicknesses, doses, or on different devices, which places higher demands on the versatility of the model. Existing methods only consider a single scanning scheme, ignoring the imaging differences under different scanning protocols, and there is a problem of data limitation. In addition, interpolation will lead to the loss of key details, which will have an adverse effect on subsequent diagnosis. Figure 1 The original image, the interpolated image, and the image processed by the VAE (Variational Autoencoder) model are shown respectively. After encoding and decoding, the VAE model introduces additional honeycomb artifacts and blurred details. Medical images have more high-frequency information. The processing method of the VAE latent space destroys a large number of details, making it difficult to meet the requirements for detail preservation in the CT noise artifact suppression task. In addition, the direct prediction of the entire image by the diffusion model significantly increases the complexity of model training and poses a challenge to obtaining results consistent with visual perception. Summary of the invention
[0004] In view of the problems existing in the prior art, the present invention provides a zero-sample low-dose CT image denoising method and device based on a strip diffusion model.
[0005] The technical solution adopted by the present invention specifically includes:
[0006] The first aspect discloses a zero-sample low-dose CT image denoising method based on a strip diffusion model, comprising:
[0007] Acquire a normal-dose CT image under conventional scanning conditions, and based on the CTDI value under the conventional scanning conditions, acquire a low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image, and obtain low-dose CT images of different thicknesses;
[0008] Build a strip diffusion model, which includes a forward diffusion structure and a reverse reasoning structure. The forward diffusion structure is used to gradually add noise to the input normal dose image in a Markov chain manner to obtain a pure noise image. The reverse reasoning structure uses the UNet model to divide the input CT image into different strips. At the same time, reverse reasoning gradually removes the noise of each strip to generate denoised strips in turn, and the generated denoised strips are spliced into the denoising result. A certain overlap rate is set between each strip.
[0009] Inputting the acquired normal dose CT image into the strip diffusion model, and calculating the difference between the noise added by the forward diffusion structure and the noise predicted by the UNet model through the mean square loss function to train the entire strip diffusion model;
[0010] A low-dose CT image whose CTDI value is a preset ratio of the CTDI value of a normal-dose CT image is input as test data into the trained strip diffusion model to obtain a denoised low-dose CT image.
[0011] In one embodiment, the probability distribution of the noise image at step t obtained by the forward diffusion structure is:
[0012] ;
[0013] in is the variance of the Gaussian noise added at step t, represents a Gaussian distribution, is the noise image obtained in step t, The noise image is obtained for the t-1th step, For known Case The probability distribution of .
[0014] In one embodiment, the UNet model includes an encoder, an intermediate layer, a decoder, a time encoder and a position encoder, the encoder includes a 3x3 convolution block and an upsampling convolution module, the intermediate layer includes a convolution layer, the decoder includes a 3x3 convolution block and a downsampling convolution module, the time encoder is used to encode the current time, and the position encoder is used to encode the current strip, wherein the features after time and position encoding are fused and added to the intermediate layer of the UNet model, the last layer of the encoder and the last layer of the decoder.
[0015] In one embodiment, the noise predicted by the UNet model is calculated as:
[0016]
[0017]
[0018] in, is a null value, is the noise predicted for the first strip, k is the number of strips divided, a represents the overlap rate, is the content of the k-1th strip, is the noise predicted for the kth strip.
[0019] In one embodiment, the 3x3 convolution block is a residual network block, and the input of the encoder is the noise image obtained in the tth step. After the residual network block calculates the rich feature information, the downsampling convolution module obtains a smaller feature map; then the intermediate layer learns a more advanced feature representation. , The input decoder calculates the feature map through the residual network block, and then the upsampling convolution module restores the image detail structure to obtain the final output feature map.
[0020] In one embodiment, the encoder and the decoder add the corresponding encoder block and decoder feature block through a skip connection.
[0021] In one embodiment, a low-dose CT image with a CTDI value of 25% or 50% of a normal-dose CT image is input as test data into a trained strip diffusion model to obtain a denoised low-dose CT image, including:
[0022] The low-dose CT images in the test data are sampled in strips, where the first strip is conditioned on the null value and the low-dose CT image content to obtain the denoising result, and the subsequent strips additionally introduce part of the information of the previous strip, and then each denoised strip is spliced in sequence to obtain the denoised low-dose CT image.
[0023] Based on the same inventive concept, the second aspect of the present invention discloses a zero-sample low-dose CT image denoising device based on a strip diffusion model, comprising:
[0024] A data acquisition module, used to acquire a normal-dose CT image under conventional scanning conditions, and based on the CTDI value under the conventional scanning conditions, acquire a low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image, and obtain low-dose CT images of different thicknesses;
[0025] A model building module is used to build a strip diffusion model. The strip diffusion model includes a forward diffusion structure and a reverse reasoning structure. The forward diffusion structure is used to gradually add noise to the input normal dose image in a Markov chain manner to obtain a pure noise image. The reverse reasoning structure uses a UNet model to divide the input CT image into different strips. At the same time, reverse reasoning gradually removes the noise of each strip to generate denoised strips in turn, and the generated denoised strips are spliced into denoising results. A certain overlap rate is set between each strip.
[0026] A model training module, used for inputting the acquired normal dose CT image into the strip diffusion model, and calculating the difference between the noise added by the forward diffusion structure and the noise predicted by the UNet model through a mean square loss function to train the entire strip diffusion model;
[0027] The test module is used to input the acquired low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image as test data into the trained strip diffusion model to obtain a denoised low-dose CT image.
[0028] Based on the same inventive concept, the third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the zero-sample low-dose CT image denoising method based on the strip diffusion model described in the first aspect.
[0029] Based on the same inventive concept, the fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the zero-sample low-dose CT image denoising method based on the strip diffusion model described in the first aspect is implemented.
[0030] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:
[0031] The present invention discloses a zero-sample low-dose CT image denoising method based on a strip diffusion model. The method obtains low-dose CT images of different thicknesses for testing, collects related normal-dose CT images for model training, and does not require complete pairing; constructs a strip diffusion model; gradually diffuses noise on the normal-dose image pairs during the training process to generate a noise target image; based on the low-dose CT image as a condition, the low-dose CT image is processed in blocks, and the noise is gradually removed by reverse reasoning to generate denoised strips in sequence, and finally spliced into a denoising result. The present invention uses a strip diffusion model for low-dose CT image reconstruction, which reduces the difficulty of model training. In the prior art, it is usually difficult to obtain CT images under paired doses due to ethical restrictions. The method of the present invention only requires normal-dose CT images to train the entire model. In addition, the diffusion model has great advantages in protecting the original details. While capturing the complex structure and details of the image, the image noise and artifacts are better removed. In addition, the processing results of the present invention and the existing advanced methods are submitted to radiologists for blind evaluation, and the method of the present invention achieves the best results. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] 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.
[0033] Figure 1 This is a comparison diagram of low-dose images in the prior art, images after interpolation and images processed by the VAE model;
[0034] Figure 2 It is the overall framework of the zero-sample low-dose CT image denoising method based on the strip diffusion model in the embodiment of the present invention;
[0035] Figure 3 It is a structural diagram of the Unet model in an embodiment of the present invention;
[0036] Figure 4 Schematic diagram of representative slice denoising results in an embodiment of the present invention. DETAILED DESCRIPTION
[0037] The present invention constructs a novel zero-sample strip scanning diffusion model, which uses the high fidelity of the diffusion model to complete end-to-end denoising of low-dose CT images of different doses, thicknesses or devices. In particular, the training process only requires normal-dose CT images, which greatly reduces the data dependence of the model, and only uses one scene data to train the model across dose and thickness conditions. The sampling process adopts a strip scanning strategy, combining overlapping strip information and input low-dose CT images to solve the maximum a posteriori problem, thereby generating denoising results in turn. The present invention only uses simple convolution and attention architectures, and conducts extensive experiments on datasets involving different doses and thicknesses. The results show that the method of the present invention outperforms the supervised method in most cases, and visualization and blind evaluation show that the method of the present invention is very close to normal-dose CT images.
[0038] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] Embodiment 1
[0040] The present invention discloses a zero-sample low-dose CT image denoising method based on a strip diffusion model, comprising:
[0041] S1: obtaining a normal-dose CT image under conventional scanning conditions, and based on the CTDI value under the conventional scanning conditions, obtaining a low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image, and obtaining low-dose CT images of different thicknesses;
[0042] S2: Build a strip diffusion model, which includes a forward diffusion structure and a reverse reasoning structure. The forward diffusion structure is used to gradually add noise to the input normal dose image in a Markov chain manner to obtain a pure noise image. The reverse reasoning structure uses the UNet model to divide the input CT image into different strips. At the same time, the reverse reasoning gradually removes the noise of each strip to generate denoised strips in turn, and the generated denoised strips are spliced into the denoising result. A certain overlap rate is set between each strip.
[0043] S3: inputting the acquired normal dose CT image into the strip diffusion model, and calculating the difference between the noise added by the forward diffusion structure and the noise predicted by the UNet model through the mean square loss function to train the entire strip diffusion model;
[0044] S4: A low-dose CT image whose CTDI value is a preset ratio of the CTDI value of a normal-dose CT image is input as test data into the trained strip diffusion model to obtain a denoised low-dose CT image.
[0045] Specifically, the CTDI value refers to CT Dose Index, which refers to the radiation dose received by the subject in the plane of the ray during the CT examination. The preset ratio means that the CTDI value of the low-dose CT image is the percentage value of the CTDI value of the normal-dose CT image. The preset ratio can be set according to the actual situation. For example, the CTDI value under conventional scanning conditions can be used as the basis to obtain CT images with CTDI values of 25% and 50% of the CTDI value of the normal-dose CT image as low-dose CT images. Different thicknesses can also be selected according to actual conditions, such as 1mm, 2mm and 3mm. The model training process only uses conventional dose (normal dose) 3mm CT images (slices), and the test process uses 3mm thick 25% and 50% dose CT image slices, and 1mm thick 25% dose and 50% dose CT image slices.
[0046] In the specific implementation process, in S1, the public data set Mayo2016 is used. The Mayo data set provides CT data of two different thicknesses, 1 mm and 3 mm. The data set provides corresponding low-dose CT and normal-dose CT data, where the low dose is 25% of the normal dose, and contains a total of 10 patient scan data. This embodiment randomly selects 9 patient data to train the model and one patient for testing.
[0047] In order to verify the effect of the model on different dose data, the data of real dead piglets were selected. The data was scanned by GE equipment, and each scan contained 850 CT images. Normal dose CT, 25% low dose CT and 10% low dose CT data were selected respectively. In order to further verify the model on different devices, a chest prosthesis was scanned using GE (General Electric Company) and Siemens (Siemens) equipment. This embodiment uses a chest prosthesis provided by GE. The prosthesis shows consistent volume, size and density in different anatomical structures such as lungs, vertebrae, chest wall, mediastinum, etc., which is very close to the soft tissue and bone density of the corresponding parts of the human body. During the CT scan, the phantom shows the same x-ray attenuation characteristics as the human body, so that the acquisition of CT values is very close to the actual human anatomical structure. We configured the GE equipment to 120 kV and 300 mA to obtain normal dose chest data, and then adjusted the tube current to obtain low dose data of 25% of the normal dose. Subsequently, the normal dose and 25% dose CT data were configured on the Siemens equipment, corresponding to the CTDI value, based on the CTDI value provided by the GE equipment. Both devices were scanned with a layer thickness of 0.625 mm and repeated 5 times, 4 times for training and 1 time for testing.
[0048] In step S2, a certain overlap rate is set between each stripe, and the value of the overlap rate can be set according to the situation, for example, to 25%, 32%, etc.
[0049] Step S3 is the training process of the model. In one embodiment, the training process cuts the image into five strips, and sets a 25% overlap rate between each strip. The first strip is generated unconditionally, and the noise corresponding to the strip is predicted only by a null value. The following four strips use the information of the previous strip as a condition to assist in generating noise. Finally, the difference between the predicted noise and the noise added in the forward process is calculated by the mean square loss function, thereby completing the training of the entire model.
[0050] This paper designs a novel strip scanning zero sampling denoising method. The entire framework only needs normal dose CT images for training to achieve end-to-end low dose CT image denoising. Specifically, we divide the entire CT image into different strips and set a certain overlap rate between the strips. Each strip is gradually generated by the overlapping information of the previous strip, which cleverly solves the training problem of the entire image and improves the retention of original details without using scaling and dimensionality reduction.
[0051] In one embodiment, the probability distribution of the noise image at step t obtained by the forward diffusion structure is:
[0052] ;
[0053] in is the variance of the Gaussian noise added at step t, represents a Gaussian distribution, is the noise image obtained in step t, The noise image is obtained for the t-1th step, For known Case The probability distribution of .
[0054] Specifically, the forward diffusion structure only inputs the normal dose image and gradually adds noise in a Markov chain manner. For example, if the total number of iterations is set to T, the pure noise image is .
[0055] In one embodiment, the UNet model includes an encoder, an intermediate layer, a decoder, a time encoder and a position encoder, the encoder includes a 3x3 convolution block and an upsampling convolution module, the intermediate layer includes a convolution layer, the decoder includes a 3x3 convolution block and a downsampling convolution module, the time encoder is used to encode the current time, and the position encoder is used to encode the current strip, wherein the features after time and position encoding are fused and added to the intermediate layer of the UNet model, the last layer of the encoder and the last layer of the decoder.
[0056] In one embodiment, the noise predicted by the UNet model is calculated as:
[0057]
[0058]
[0059] in, is a null value, is the noise predicted for the first strip, k is the number of strips divided, a represents the overlap rate, is the content of the k-1th strip, is the noise predicted for the kth strip.
[0060] Specifically, if the input image is divided into 5 strips, the input condition of the first strip is a null value. , the input conditions for the second to fifth strips are the overlapping information of the previous strip The overlap rate a is set to 25%. The training process is calculated by MSELoss (mean square loss function) and Difference. Noise added to the forward pass.
[0061] In one embodiment, the 3x3 convolution block is a residual network block, and the input of the encoder is the noise image obtained in the tth step. After the residual network block calculates the rich feature information, the downsampling convolution module obtains a smaller feature map; then the intermediate layer learns a more advanced feature representation. , The input decoder calculates the feature map through the residual network block, and then the upsampling convolution module restores the image detail structure to obtain the final output feature map.
[0062] In one embodiment, the encoder and the decoder add the corresponding encoder block and decoder feature block through a skip connection.
[0063] See also Figure 3 , is the structure diagram of the Unet model. The four blocks on the left are encoders, the middle block is the middle layer, and the four blocks on the right are decoders. Skip connections are introduced in UNet. The four encoder blocks correspond to the four decoder blocks respectively, and the corresponding encoder blocks are added to the feature blocks of the decoder through skip connections. This enables the network to focus on local details and global context information at the same time.
[0064] In one embodiment, a low-dose CT image with a CTDI value of 25% or 50% of a normal-dose CT image is input as test data into a trained strip diffusion model to obtain a denoised low-dose CT image, including:
[0065] The low-dose CT images in the test data are sampled in strips, where the first strip is conditioned on the null value and the low-dose CT image content to obtain the denoising result, and the subsequent strips additionally introduce part of the information of the previous strip, and then each denoised strip is spliced in sequence to obtain the denoised low-dose CT image.
[0066] Specifically, the testing process is the sampling process, which removes noise from the strips of the low-dose CT image in turn and finally splices them into a complete CT image. The predicted strips can be expressed as , where k is 5, indicating that there are 5 stripes in total. represents the first prediction strip, Indicates the second prediction strip, Denotes the kth predicted strip. Remove noise from the kth strip of the low-dose CT image After getting . Finally, the denoised image is spliced together .
[0067] The method of the present invention is described below by means of specific examples.
[0068] like Figure 2As shown, the overall framework of the zero-sample low-dose CT image denoising method based on the strip diffusion model is carried out in steps. First, the first strip is obtained only by relying on the null value. The subsequent strip generation will enhance the controllability of the content based on the 25% (overlap rate) information of the previous strip. Finally, different strips are spliced from top to bottom to obtain the final result.
[0069] The CT scanning scheme was designed to obtain low-dose images of different equipment, different doses and different slice thicknesses for testing, and the normal-dose CT images were used to train the entire model. The entire process was image domain post-processing, and the strip diffusion model was constructed based on this data.
[0070] This embodiment selects two public data sets and scans the phantom data under different devices separately. The public data set Mayo2016 is used. The Mayo data set provides CT data of two different thicknesses, 1mm and 3mm. The data set provides corresponding low-dose CT and normal-dose CT data, with the low dose being 25% of the normal dose, and contains a total of 10 patient scan data. This embodiment randomly selects 9 patient data to train the model and one patient for testing. In order to verify the effect of the model on different dose data, the data of real dead piglets were selected. The data was scanned by GE equipment, and each scan contained 850 CT images. Normal-dose CT, 25% low-dose CT and 10% low-dose CT data were selected respectively. In order to further verify the model on different devices, this embodiment uses GE and Siemens equipment to scan a chest prosthesis. This study uses a chest prosthesis provided by General Electric (GE) of the United States. The prosthesis shows consistent volume, size and density in different anatomical structures such as the lungs, vertebrae, chest wall, mediastinum, etc., which is very close to the soft tissue and bone density of the corresponding parts of the human body. During the CT scan, the phantom shows the same x-ray attenuation characteristics as the human body, making the acquisition of CT values very close to the actual human anatomy. In this embodiment, the GE device is configured to 120 kV and 300 mA to obtain normal dose chest data, and then the tube current is adjusted to obtain low dose data of 25% of the normal dose. Subsequently, the normal dose and 25% dose CT data are configured on the Siemens device, corresponding to the CTDI value, based on the CTDI value provided by the GE device. Both devices are scanned with a scanning layer thickness of 0.625 mm, and each is repeated 5 times, 4 of which are used for training and 1 for testing.
[0071] Figure 3 The entire UNet framework is shown. The input of the encoder is the noise image of the current t steps , ResNetblock (residual network block) calculates rich feature information; a smaller feature map is obtained through the downsampling convolution module to reduce the spatial resolution; then Midblock learns a more advanced feature representation The input of the decoder is the features processed by the intermediate layer , after calculating the feature map through the ResNet block, the upsampling convolution layer restores the image detail structure. Skip connections are introduced in UNet to connect the feature map of a certain layer of the encoder with the feature map of the corresponding layer of the decoder. This enables the network to focus on local details and global context information at the same time. The additionally introduced time encoder and position encoder encode the time t and the strip order into one-dimensional features and are added to the middle layer, encoder, and the last layer of the decoder of the UNet model.
[0072] The test is to remove noise from the strips in turn during the sampling process, and finally splice them into a complete CT image. The predicted strips can be expressed as . Finally, the denoised image is spliced together .
[0073] Mayo data, piglet data and phantom data slices were randomly selected in the test set and read blindly by radiologists. Each group of images contains informed low-dose CT images LDCT and normal-dose images NDCT. The denoising results of the other methods are blind, including NLM (non-local mean filtering), RED-CNN (residual encoder-decoder convolutional neural network), EDCNN (edge-enhanced composite loss dense connection network), CTformer (CT deformation network) and the method of the present invention. Referring to the standard EU CT evaluation guidelines, the noise, structural clarity, artifacts and diagnosability of CT images were evaluated respectively, using a 5-point evaluation method (1 is the worst and 5 is the highest). Table 1 shows the results of the doctor's blind evaluation. The method of the present invention has the highest diagnosability and the highest degree of image detail structure retention. As a supervised learning method, CTformer needs to train paired low-dose and normal-dose CT images at the same time. Although this method has good denoising and artifact removal effects, it has the worst final diagnosability due to excessive blurring. Figure 4 The denoising results of the method on representative slices of different data sets are shown. In order to show the image details more clearly, the region of interest is selected to enlarge and show the denoising details. NDCT is a normal-dose CT image, and LDCT is a low-dose CT image. Ours represents the denoising result of the present invention. The denoising result of the region of interest is very close to NDCT, and the details of the image are well preserved, and there is no problem of over-smoothing at the edges. The image noise is significantly reduced, and the visual effect is significantly improved.
[0074] Table 1 Results of blind evaluation by doctors
[0075]
[0076] Embodiment 2
[0077] Based on the same inventive concept, this embodiment discloses a zero-sample low-dose CT image denoising device based on a strip diffusion model, comprising:
[0078] A data acquisition module, used to acquire a normal-dose CT image under conventional scanning conditions, and based on the CTDI value under the conventional scanning conditions, acquire a low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image, and obtain low-dose CT images of different thicknesses;
[0079] A model building module is used to build a strip diffusion model. The strip diffusion model includes a forward diffusion structure and a reverse reasoning structure. The forward diffusion structure is used to gradually add noise to the input normal dose image in a Markov chain manner to obtain a pure noise image. The reverse reasoning structure uses a UNet model to divide the input CT image into different strips. At the same time, reverse reasoning gradually removes the noise of each strip to generate denoised strips in turn, and the generated denoised strips are spliced into denoising results. A certain overlap rate is set between each strip.
[0080] A model training module, used for inputting the acquired normal dose CT image into the strip diffusion model, and calculating the difference between the noise added by the forward diffusion structure and the noise predicted by the UNet model through a mean square loss function to train the entire strip diffusion model;
[0081] The test module is used to input the acquired low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image as test data into the trained strip diffusion model to obtain a denoised low-dose CT image.
[0082] Since the device introduced in the second embodiment of the present invention is a device used to implement the zero-sample low-dose CT image denoising method based on the strip diffusion model in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, the person skilled in the art can understand the specific structure and deformation of the device, so it is not repeated here. All devices used in the method in the first embodiment of the present invention belong to the scope of protection of the present invention.
[0083] Embodiment 3
[0084] Based on the same inventive concept, the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method described in the first embodiment is implemented.
[0085] Since the computer-readable storage medium introduced in the third embodiment of the present invention is a computer-readable storage medium used to implement the zero-sample low-dose CT image denoising method based on the strip diffusion model in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, those skilled in the art can understand the specific structure and deformation of the computer-readable storage medium, so it is not repeated here. All computer-readable storage media used in the method of the first embodiment of the present invention belong to the scope of protection of the present invention.
[0086] Embodiment 4
[0087] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in Embodiment 1 when executing the program.
[0088] Since the computer device introduced in the fourth embodiment of the present invention is a computer device used to implement the zero-sample low-dose CT image denoising method based on the strip diffusion model in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, the person skilled in the art can understand the specific structure and deformation of the computer device, so it is not repeated here. All computer devices used in the method of the first embodiment of the present invention belong to the scope of protection of the present invention.
[0089] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0090] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0091] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic creative concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A zero-sample low-dose CT image denoising method based on a strip diffusion model, characterized in that: include: Acquire a normal-dose CT image under conventional scanning conditions, and based on the CTDI value of the normal-dose CT image under the conventional scanning conditions, acquire a low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image, and obtain low-dose CT images of different thicknesses; Build a strip diffusion model, which includes a forward diffusion structure and a reverse reasoning structure. The forward diffusion structure is used to gradually add noise to the input normal dose image in a Markov chain manner to obtain a pure noise image. The reverse reasoning structure uses the UNet model to divide the input CT image into different strips. At the same time, reverse reasoning gradually removes the noise of each strip to generate denoised strips in turn, and the generated denoised strips are spliced into the denoising result. A certain overlap rate is set between each strip. Inputting the acquired normal dose CT image into the strip diffusion model, and calculating the difference between the noise added by the forward diffusion structure and the noise predicted by the UNet model through the mean square loss function to train the entire strip diffusion model; A low-dose CT image whose CTDI value is a preset ratio of the CTDI value of a normal-dose CT image is input as test data into the trained strip diffusion model to obtain a denoised low-dose CT image.
2. The zero-sample low-dose CT image denoising method based on the strip diffusion model according to claim 1, characterized in that: The probability distribution of the noise image at step t obtained by the forward diffusion structure is: ; in is the variance of the Gaussian noise added at step t, represents a Gaussian distribution, is the noise image obtained in the tth step, The noise image is obtained for the t-1th step, Known In case The probability distribution of .
3. The zero-sample low-dose CT image denoising method based on the strip diffusion model according to claim 1, characterized in that: The UNet model includes an encoder, an intermediate layer, a decoder, a time encoder and a position encoder. The encoder includes a 3x3 convolution block and an upsampling convolution module, the intermediate layer includes a convolution layer, the decoder includes a 3x3 convolution block and a downsampling convolution module, the time encoder is used to encode the current time, and the position encoder is used to encode the current strip. The features after time and position encoding are fused and added to the intermediate layer of the UNet model, the last layer of the encoder and the last layer of the decoder.
4. The zero-sample low-dose CT image denoising method based on the strip diffusion model according to claim 1, characterized in that: The noise predicted by the UNet model is calculated as: in, is a null value, is the noise predicted for the first strip, k is the number of strips divided, a represents the overlap rate, is the content of the k-1th strip, is the noise predicted for the kth strip.
5. The zero-sample low-dose CT image denoising method based on the strip diffusion model as claimed in claim 3, characterized in that: The 3x3 convolution block is a residual network block, and the input of the encoder is the noise image obtained in the tth step. After the residual network block calculates the rich feature information, the downsampling convolution module obtains a smaller feature map; then the intermediate layer learns a more advanced feature representation. , The input decoder calculates the feature map through the residual network block, and then the upsampling convolution module restores the image detail structure to obtain the final output feature map.
6. The zero-sample low-dose CT image denoising method based on the strip diffusion model according to claim 5, characterized in that: The encoder and decoder add the corresponding encoder blocks and decoder feature blocks through skip connections.
7. The zero-sample low-dose CT image denoising method based on the strip diffusion model according to claim 1, characterized in that: The low-dose CT images with CTDI values of 25% and 50% of the normal-dose CT images are used as test data to input the trained strip diffusion model to obtain the denoised low-dose CT images, including: The low-dose CT images in the test data are sampled in strips, where the first strip is conditioned on the null value and the low-dose CT image content to obtain the denoising result, and the subsequent strips additionally introduce part of the information of the previous strip, and then each denoised strip is spliced in sequence to obtain the denoised low-dose CT image.
8. A zero-sample low-dose CT image denoising device based on a strip diffusion model, characterized in that: include: A data acquisition module, used to acquire a normal-dose CT image under conventional scanning conditions, and based on the CTDI value of the normal-dose CT image under the conventional scanning conditions, acquire a low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image, and obtain low-dose CT images of different thicknesses; A model building module is used to build a strip diffusion model. The strip diffusion model includes a forward diffusion structure and a reverse reasoning structure. The forward diffusion structure is used to gradually add noise to the input normal dose image in a Markov chain manner to obtain a pure noise image. The reverse reasoning structure uses a UNet model to divide the input CT image into different strips. At the same time, reverse reasoning gradually removes the noise of each strip to generate denoised strips in turn, and the generated denoised strips are spliced into denoising results. A certain overlap rate is set between each strip. A model training module, used for inputting the acquired normal dose CT image into the strip diffusion model, and calculating the difference between the noise added by the forward diffusion structure and the noise predicted by the UNet model through a mean square loss function to train the entire strip diffusion model; The test module is used to input the acquired low-dose CT image whose CTDI value is a preset ratio of the CTDI value of the normal-dose CT image as test data into the trained strip diffusion model to obtain a denoised low-dose CT image.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the zero-sample low-dose CT image denoising method based on a strip diffusion model as claimed in any one of claims 1 to 7 is implemented.
10. A computer device 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 program, the zero-sample low-dose CT image denoising method based on the strip diffusion model as claimed in any one of claims 1 to 7 is implemented.
