Signal noise reduction model training method, signal noise reduction method, CT signal noise reduction model training method and CT signal noise reduction method

Abstract

The invention relates to the technical field of signal processing, and provides a signal noise reduction model training method, a signal noise reduction method, a CT signal noise reduction model training method and a CT signal noise reduction method, which can improve the signal retention capability of a signal noise reduction model. The method comprises the following steps: acquiring multi-scale training input signal data; acquiring a first network; the first network comprises a scale-by-scale coding path and a scale-by-scale decoding path; and inputting the training input signal data of each scale into a coding node of a corresponding scale in the coding path, and / or inputting the training input signal data of each scale into a decoding node of a corresponding scale in the decoding path, so as to train the first network and obtain a signal noise reduction model.

Description

Technical Field

[0001] The present application relates to the field of signal processing technology, and in particular to a signal denoising model training method and apparatus thereof, a signal denoising method and apparatus thereof, a CT signal denoising model training method and apparatus thereof, a CT signal denoising method and apparatus thereof, a computer device, a storage medium, and a computer program product. Background Art

[0002] A signal is a physical quantity that carries information. Most signals carry a certain amount of noise, which is considered interference. Denoising a signal can reduce the noise, improve the signal-to-noise ratio, and prevent the signal from being overwhelmed by excessive noise.

[0003] One of the noise reduction technologies is to train a signal noise reduction model through deep learning. However, in general training, the network input and output data are all of a single scale, and the signal noise reduction model has poor ability to retain signals. Summary of the Invention

[0004] Based on this, it is necessary to provide a signal denoising model training method and device, a signal denoising method and device, a CT signal denoising model training method and device, a CT signal denoising method and device, a computer device, a storage medium and a computer program product to address the above technical problems.

[0005] The present application provides a method for training a signal denoising model, the method comprising:

[0006] Obtain multi-scale training input signal data;

[0007] Acquire a first network; the first network includes a scale-by-scale encoding path and a scale-by-scale decoding path;

[0008] Inputting the training input signal data of each scale into the encoding node of the corresponding scale in the encoding path, and / or inputting the training input signal data of each scale into the decoding node of the corresponding scale in the decoding path, so as to train the first network and obtain a signal denoising model.

[0009] The present application provides a training device for a signal denoising model, the device comprising:

[0010] A multi-scale signal acquisition module is used to obtain multi-scale training input signal data;

[0011] A first network acquisition module, configured to acquire a first network; the first network includes a scale-by-scale encoding path and a scale-by-scale decoding path;

[0012] The first network training module is used to input the training input signal data of each scale into the encoding node of the corresponding scale in the encoding path, and / or input the training input signal data of each scale into the decoding node of the corresponding scale in the decoding path, so as to train the first network and obtain a signal denoising model.

[0013] The present application provides a signal noise reduction method, the method comprising:

[0014] Obtaining a signal to be denoised;

[0015] A denoised signal is obtained according to the signal to be denoised and the signal denoising model; the signal denoising model is obtained by training according to the signal denoising model training method introduced in the above embodiment.

[0016] The present application provides a signal noise reduction device, comprising:

[0017] A signal acquisition module for noise reduction, used to obtain the signal for noise reduction;

[0018] The signal denoising module is used to obtain a denoised signal based on the signal to be denoised and a signal denoising model; the signal denoising model is trained according to the signal denoising model training method introduced in the above embodiment.

[0019] The present application provides a training method for a CT signal noise reduction model, the method comprising:

[0020] Obtaining at least one scale of training input CT raw data and multi-scale training input CT images based on the detected CT raw data;

[0021] Obtaining a second network; the second network includes a scale-by-scale raw data encoding path, a scale-by-scale image encoding path, a scale-by-scale image decoding path, and a scale-by-scale raw data decoding path; an end of the raw data encoding path and a starting end of the image encoding path are connected via a backprojection unit, and an end of the image decoding path and a starting end of the raw data decoding path are connected via a forward projection unit;

[0022] Inputting the training input CT raw data of at least one scale into the raw data encoding node of the corresponding scale in the raw data encoding path, and / or inputting the training input CT raw data of at least one scale into the raw data decoding node of the corresponding scale in the raw data decoding path;

[0023] Inputting the training input CT image of each scale into the image encoding node of the corresponding scale in the image encoding path, and / or inputting the training input CT image of each scale into the image decoding node of the corresponding scale in the image decoding path;

[0024] After inputting at least one scale of training input CT raw data and multi-scale training input CT images, the second network is trained to obtain a CT signal denoising model.

[0025] The present application provides a training device for a CT signal noise reduction model, the device comprising:

[0026] A CT signal acquisition module is used to obtain at least one scale of training input CT raw data and multi-scale training input CT images based on the CT raw data obtained by detection;

[0027] a second network module, configured to obtain a second network; the second network comprising a scale-by-scale raw data encoding path, a scale-by-scale image encoding path, a scale-by-scale image decoding path, and a scale-by-scale raw data decoding path; an end of the raw data encoding path and a starting end of the image encoding path are connected via a backprojection unit, and an end of the image decoding path and a starting end of the raw data decoding path are connected via a forward projection unit;

[0028] a raw data input module, configured to input the training input CT raw data of at least one scale into the raw data encoding node of the corresponding scale in the raw data encoding path, and / or input the training input CT raw data of at least one scale into the raw data decoding node of the corresponding scale in the raw data decoding path;

[0029] an image input module, configured to input the training input CT image of each scale into the image encoding node of the corresponding scale in the image encoding path, and / or input the training input CT image of each scale into the image decoding node of the corresponding scale in the image decoding path;

[0030] The second network training module is used to train the second network after inputting training input CT raw data of at least one scale and training input CT images of multiple scales to obtain a CT signal denoising model.

[0031] The present application provides a CT signal noise reduction method, the method comprising:

[0032] Acquire the CT signal to be de-noised;

[0033] A CT signal after noise reduction is obtained according to the CT signal to be noise reduced and the CT signal noise reduction model; the CT signal noise reduction model is obtained by training according to the training method of the CT signal noise reduction model introduced in the above embodiment.

[0034] The present application provides a CT signal noise reduction device, the device comprising:

[0035] A CT signal acquisition module for noise reduction, used to acquire the CT signal for noise reduction;

[0036] The CT signal denoising module is used to obtain a denoised CT signal based on the CT signal to be denoised and a CT signal denoising model; the CT signal denoising model is trained according to the CT signal denoising model training method introduced in the above embodiment.

[0037] The present application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the above method.

[0038] The present application provides a computer-readable storage medium having a computer program stored thereon, and the computer program is used by a processor to execute the above method.

[0039] The present application provides a computer program product having a computer program stored thereon, wherein the computer program is used by a processor to execute the above method.

[0040] In the solution provided in the present application, when training the signal denoising model, multi-scale training input signal data is used. The training input signal data of each scale can be input into the encoding node of the corresponding scale in the encoding path, and / or, the training input signal data of each scale can be input into the decoding node of the corresponding scale in the decoding path. During the network processing process, the original information of the signal can be better retained, thereby improving the ability of the signal denoising model to retain the original information of the signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 1 is a flow chart of a method for training a signal denoising model in one embodiment;

[0043] Figure 2 is a schematic diagram of the architecture of a first network in one embodiment;

[0044] FIG3( a ) is a schematic diagram of the training process of the first network in one embodiment;

[0045] FIG3( b ) is a schematic diagram of the training process of the first network in another embodiment;

[0046] FIG4( a ) is a schematic diagram of the training process of the first network in another embodiment;

[0047] FIG4( b ) is a schematic diagram of the training process of the first network in yet another embodiment;

[0048] FIG4( c ) is a schematic diagram of the training process of the first network in another embodiment;

[0049] Figure 5 1 is a flow chart of a signal noise reduction method according to an embodiment;

[0050] Figure 6 Schematic diagram of a flow chart of a method for training a CT signal noise reduction model in one embodiment;

[0051] Figure 7 is a schematic diagram of the architecture of the second network in one embodiment;

[0052] Figure 8 1 is a schematic diagram of the training process of the second network in one embodiment;

[0053] Figure 9 Schematic diagram of a flow chart of a CT signal noise reduction method in one embodiment;

[0054] Figure 10 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0056] The present application provides a method for training a signal denoising model, which can be executed by a computer device. The method for training a signal denoising model may include: Figure 1 Steps shown.

[0057] Step S101: Acquire multi-scale training input signal data.

[0058] The training data used for signal noise reduction training can be referred to as training signal data. The training signal data includes input data and target data. The input data can be referred to as training input signal data, and the target data can be referred to as training target signal data.

[0059] In order to improve the signal denoising model's ability to retain signals, this embodiment performs training using multi-scale training input signal data, where each scale of training input signal data has corresponding scale of training target signal data.

[0060] Step S102: Acquire a first network.

[0061] The first network includes a scale-by-scale encoding path and a scale-by-scale decoding path.

[0062] The first network may include a scale-by-scale encoding path and a scale-by-scale decoding path. The encoding path performs scale-by-scale encoding from large scale to small scale, and the decoding path performs scale-by-scale decoding from small scale to large scale.

[0063] The encoding path includes encoding nodes of multiple scales, which are connected in series from large to small. The output of the large-scale encoding node is connected to the input of the small-scale encoding node, so that the encoding path can perform scale-by-scale encoding from large to small. The decoding path includes decoding nodes of multiple scales, which are connected in series from small to large. The output of the small-scale decoding node is connected to the input of the large-scale decoding node, so that the decoding path can perform scale-by-scale decoding from small to large.

[0064] A connection may exist between encoding nodes and decoding nodes of the same scale. Specifically, the output of the encoding node may be connected to the input of the decoding node. An intermediate unit may be provided between encoding nodes and decoding nodes of the same scale. In some cases, the encoding node with the smallest scale in the encoding path may be connected to the decoding node with the smallest scale in the decoding path via an intermediate unit.

[0065] Encoding can be implemented through downsampling, and decoding can be implemented through upsampling; accordingly, the encoding node can be called a downsampling node, and the decoding node can be called an upsampling node.

[0066] Assume that the multiple scales include a first scale, a second scale, and a third scale, where the first scale is larger than the second scale, and the second scale is larger than the third scale. For example, the first scale may be N×N, the second scale may be N / 2×N / 2, and the third scale may be N / 4×N / 4.

[0067] Reference Figure 2 The encoding path of the first network includes encoding nodes of the first scale, encoding nodes of the second scale, and encoding nodes of the third scale. The decoding path of the first network includes decoding nodes of the first scale, decoding nodes of the second scale, and decoding nodes of the third scale.

[0068] The outputs of the encoding nodes at the first scale are connected to the inputs of the decoding nodes at the first scale. The outputs of the encoding nodes at the second scale are connected to the inputs of the decoding nodes at the second scale. The outputs of the encoding nodes at the third scale are connected to the inputs of the intermediate units, and the outputs of the intermediate units are connected to the inputs of the decoding nodes at the third scale.

[0069] The output end of the coding node of the first scale is connected to the input end of the coding node of the second scale, and the output end of the coding node of the second scale is connected to the input end of the coding node of the third scale.

[0070] The output end of the decoding node of the third scale is connected to the input end of the decoding node of the second scale, and the output end of the decoding node of the second scale is connected to the input end of the decoding node of the first scale.

[0071] Step S103: Input the training input signal data of each scale into the encoding node of the corresponding scale in the encoding path, and / or input the training input signal data of each scale into the decoding node of the corresponding scale in the decoding path, so as to train the first network and obtain a signal denoising model.

[0072] Assume that the multi-scale includes a first scale, a second scale, and a third scale, the first scale is larger than the second scale, and the second scale is larger than the third scale.

[0073] The training input signal data of the first scale may be input into the encoding node of the first scale in the encoding path, and / or the training input signal data of the first scale may be input into the decoding node of the first scale in the decoding path.

[0074] The training input signal data of the second scale may be input into the encoding nodes of the second scale in the encoding path, and / or the training input signal data of the second scale may be input into the decoding nodes of the second scale in the decoding path.

[0075] The training input signal data of the third scale may be input into the encoding nodes of the third scale in the encoding path, and / or the training input signal data of the third scale may be input into the decoding nodes of the third scale in the decoding path.

[0076] Referring to Figure 3(a), a training diagram illustrating inputting training input signal data at each scale into coding nodes at the corresponding scale is shown. Training input signal data at a first scale is input into coding nodes at the first scale in the coding path, training input signal data at a second scale is input into coding nodes at the second scale in the coding path, and training input signal data at a third scale is input into coding nodes at the third scale in the coding path.

[0077] The encoding nodes at the first scale encode the training input signal data at the first scale to obtain an encoding result (the encoding result is referred to as a first encoding result). The encoding nodes at the first scale input the first encoding result to the decoding nodes at the first scale and the encoding nodes at the second scale, respectively.

[0078] The encoding nodes at the second scale encode the second-scale training input signal data and the first encoding result to obtain an encoding result (referred to as a second encoding result). The encoding nodes at the second scale input the second encoding result to the decoding nodes at the second scale and the encoding nodes at the third scale, respectively.

[0079] The encoding nodes at the third scale encode the training input signal data at the third scale and the second encoding result to obtain an encoding result (referred to as a third encoding result). The encoding nodes at the third scale input the third encoding result to the decoding nodes and intermediate units at the third scale, respectively.

[0080] After processing the third encoding result, the intermediate unit inputs the processing result to the decoding node of the third scale.

[0081] The decoding node at the third scale decodes the processing result and the third encoding result to obtain a decoding result (the decoding result is referred to as a third decoding result). The decoding node at the third scale inputs the third decoding result to the decoding node at the second scale.

[0082] The decoding node at the second scale decodes the third decoding result and the second encoding result to obtain a decoding result (the decoding result is referred to as a second decoding result). The decoding node at the second scale inputs the second decoding result to the decoding node at the first scale.

[0083] The decoding node of the first scale decodes the second decoding result and the first encoding result to obtain a decoding result (this decoding result is referred to as a first decoding result).

[0084] According to the first decoding result, a first-scale output result is obtained. According to the first-scale output result and the first-scale training target signal data, a loss value corresponding to the first scale can be obtained; specifically, the first-scale output result is theoretically consistent with the first-scale training target signal data; according to the loss function, the difference between the first-scale output result and the first-scale training target signal data can be calculated to obtain the loss value corresponding to the first scale; the loss function may include but is not limited to the L1 loss function, the L2 loss function, and the SSIM loss function; the full name of SSIM is Structural Similarity Index Measure in English, and the SSIM loss function can be used to calculate image similarity. According to the loss value corresponding to the first scale, the first network is back-propagated to obtain a gradient value. According to the gradient value, the first network is gradient descent to update the parameters of the first network and complete one iteration. After several iterations according to actual needs, a signal denoising model is obtained.

[0085] Referring to Figure 3(b), unlike Figure 3(a), the second-scale training input signal data is input to the second-scale decoding nodes. The specific training process is similar to that of Figure 3(a) and will not be repeated here.

[0086] In the training method of the above-mentioned signal denoising model, multi-scale training input signal data is used when training the signal denoising model. The training input signal data of each scale can be input into the encoding node of the corresponding scale in the encoding path, and / or, the training input signal data of each scale can be input into the decoding node of the corresponding scale in the decoding path. During the network processing process, the original information of the signal can be better retained, thereby improving the ability of the signal denoising model to retain the original information of the signal.

[0087] In one embodiment, inputting the training input signal data of each scale into the coding node of the corresponding scale in the coding path includes: processing the training input signal data of each scale through an input head and then inputting the training input signal data of each scale into the coding node of the corresponding scale in the coding path.

[0088] After the training input signal data of a certain scale is processed by the input head, the scale remains consistent with the scale of the corresponding encoding node.

[0089] Exemplarily, before the training input signal data at the second scale is input into the coding node at the second scale in the coding path, the training input signal data at the second scale may be input into an input head for processing; the scale of the training input signal data at the second scale after being processed by the input head remains the same as the input scale of the coding node at the second scale; and the training input signal data at the second scale processed by the input head is input into the coding node at the second scale in the coding path.

[0090] In one embodiment, inputting the training input signal data of each scale into the decoding node of the corresponding scale in the decoding path includes: processing the training input signal data of each scale through an input head and then inputting the training input signal data of each scale into the decoding node of the corresponding scale in the decoding path.

[0091] After the training input signal data of a certain scale is processed by the input head, the scale is consistent with the scale of the corresponding decoding node.

[0092] Exemplarily, before inputting the training input signal data at the second scale into the decoding node at the second scale in the decoding path, the training input signal data at the second scale may be input to an input head for processing; the scale of the training input signal data at the second scale after being processed by the input head remains the same as the input scale of the decoding node at the second scale; and the training input signal data at the second scale processed by the input head is input to the decoding node at the second scale in the decoding path.

[0093] In one embodiment, training the first network to obtain a signal noise reduction model includes:

[0094] Obtain the output result of each scale of the first network; obtain the loss value corresponding to each scale based on the output result of each scale and the training target signal data of the corresponding scale; perform statistics on the loss values corresponding to multiple scales to obtain a comprehensive loss value; and perform gradient descent on the first network based on the comprehensive loss value to obtain a signal denoising model.

[0095] Assume that the multi-scale includes a first scale, a second scale, and a third scale, the first scale is larger than the second scale, and the second scale is larger than the third scale.

[0096] In this example, the training input signal data of each scale is input to the encoding node of the corresponding scale as an example for explanation.

[0097] Referring to Figure 4(a), training input signal data at a first scale is input into a coding node at a first scale in the coding path. Training input signal data at a second scale is input into a coding node at a second scale in the coding path. Training input signal data at a third scale is input into a coding node at a third scale in the coding path.

[0098] The encoding nodes at the first scale encode the training input signal data at the first scale to obtain an encoding result (the encoding result is referred to as a first encoding result). The encoding nodes at the first scale input the first encoding result to the decoding nodes at the first scale and the encoding nodes at the second scale, respectively.

[0099] The encoding nodes at the second scale encode the second-scale training input signal data and the first encoding result to obtain an encoding result (referred to as a second encoding result). The encoding nodes at the second scale input the second encoding result to the decoding nodes at the second scale and the encoding nodes at the third scale, respectively.

[0100] The encoding nodes at the third scale encode the training input signal data at the third scale and the second encoding result to obtain an encoding result (referred to as a third encoding result). The encoding nodes at the third scale input the third encoding result to the decoding nodes and intermediate units at the third scale, respectively.

[0101] After processing the third encoding result, the intermediate unit inputs the processing result to the decoding node of the third scale.

[0102] The decoding node at the third scale decodes the processing result and the third encoding result to obtain a decoding result (referred to as a third decoding result). The decoding node at the third scale inputs the third decoding result to the decoding node at the second scale. Based on the third decoding result, a third-scale output result is obtained.

[0103] The decoding node at the second scale decodes the third decoding result and the second encoding result to obtain a decoding result (referred to as the second decoding result). The decoding node at the second scale inputs the second decoding result to the decoding node at the first scale. Based on the second decoding result, a second-scale output result is obtained.

[0104] The decoding node of the first scale decodes the second decoding result and the first encoding result to obtain a decoding result (this decoding result is referred to as a first decoding result). Based on the first decoding result, a first scale output result is obtained.

[0105] Based on the output results at the first scale and the training target signal data at the first scale, we can obtain the loss value loss-1 corresponding to the first scale. Based on the output results at the second scale and the training target signal data at the second scale, we can obtain the loss value loss-2 corresponding to the second scale. Based on the output results at the third scale and the training target signal data at the third scale, we can obtain the loss value loss-3 corresponding to the third scale.

[0106] The loss values loss-1, loss-2, and loss-3 are statistically analyzed to obtain a comprehensive loss value. Backpropagation is performed on the first network based on the comprehensive loss value. A gradient value is obtained, and gradient descent is performed on the first network based on the gradient value, completing one iteration. After several iterations, a signal denoising model is obtained.

[0107] In one embodiment, the signal data includes any one of CT raw data and CT images.

[0108] The full name of CT in English is Computed Tomography, and in Chinese it is computer tomography.

[0109] In one embodiment, when the signal data includes a CT image, obtaining multi-scale training input signal data includes: performing multi-scale filtered back-projection reconstruction on the CT raw data to obtain a multi-scale training input CT image.

[0110] Taking the first, second, and third scales as an example, after obtaining the raw CT data, we perform filtered back projection reconstruction on the raw CT data at the first scale to obtain a training input CT image at the first scale. We then perform filtered back projection reconstruction on the raw CT data at the second scale to obtain a training input CT image at the second scale. We then perform filtered back projection reconstruction on the raw CT data at the third scale to obtain a training input CT image at the third scale.

[0111] In this embodiment, multi-scale filtered back projection reconstruction is performed on the same CT raw data, which can better retain the original information in the multi-scale training input CT image. Subsequently, the multi-scale training input CT image is input into the encoding node or decoding node of the corresponding scale, which can improve the network's ability to retain the original information.

[0112] To better understand the above method, the following describes an application example of the training method for the signal denoising model of the present application. In this application example, the multi-scale includes a first scale, a second scale, and a third scale, where the first scale is larger than the second scale, and the second scale is larger than the third scale. In this example, the signal data is a CT image, the training input signal data is referred to as the training input CT image, and the training target signal data is referred to as the training target CT image.

[0113] Multi-scale filtered back-projection reconstruction can be performed on the same CT raw data to obtain a first-scale training input CT image, a second-scale training input CT image, and a third-scale training input CT image.

[0114] The following example illustrates the input of training input CT images of various scales into encoding nodes of corresponding scales.

[0115] Referring to Figure 4(b), the training input CT image at the first scale passes through the input head and is then input into the coding node at the first scale in the coding path. The training input CT image at the second scale passes through the input head and is then input into the coding node at the second scale in the coding path. The training input CT image at the third scale passes through the input head and is then input into the coding node at the third scale in the coding path.

[0116] The encoding node at the first scale encodes the training input CT image at the first scale to obtain an encoding result (the encoding result is referred to as a first encoding result). The encoding node at the first scale inputs the first encoding result to the decoding node at the first scale and the encoding node at the second scale, respectively.

[0117] The second-scale encoding node encodes the second-scale training input CT image and the first encoding result to obtain an encoding result (referred to as a second encoding result). The second-scale encoding node inputs the second encoding result to the second-scale decoding node and the third-scale encoding node, respectively.

[0118] The encoding node at the third scale encodes the training input CT image at the third scale and the second encoding result to obtain an encoding result (the encoding result is referred to as the third encoding result). The encoding node at the third scale inputs the third encoding result to the decoding node and the intermediate unit at the third scale.

[0119] After processing the third encoding result, the intermediate unit inputs the processing result to the decoding node of the third scale.

[0120] The decoding node at the third scale decodes the processing result and the third encoding result to obtain a decoding result (this decoding result is referred to as a third decoding result). The decoding node at the third scale inputs the third decoding result to the decoding node at the second scale, and then inputs the third decoding result to output head 3 for processing.

[0121] The decoding node at the second scale decodes the third decoding result and the second encoding result to obtain a decoding result (this decoding result is referred to as the second decoding result). The decoding node at the second scale inputs the second decoding result to the decoding node at the first scale, and then inputs the second decoding result to output head 2 for processing.

[0122] The decoding node of the first scale decodes the second decoding result and the first encoding result to obtain a decoding result (this decoding result is referred to as a first decoding result). The decoding node of the first scale inputs the first decoding result to the output head 1 for processing.

[0123] The first-scale output result is obtained based on the output of output head 1. The second-scale output result is obtained based on the output of output head 2. The third-scale output result is obtained based on the output of output head 3.

[0124] Based on the output results at the first scale and the training target CT image at the first scale, we can obtain the loss value loss-1 corresponding to the first scale. Based on the output results at the second scale and the training target CT image at the second scale, we can obtain the loss value loss-2 corresponding to the second scale. Based on the output results at the third scale and the training target CT image at the third scale, we can obtain the loss value loss-3 corresponding to the third scale.

[0125] The loss values loss-1, loss-2, and loss-3 are statistically analyzed, and the statistical result obtained can be called a comprehensive loss value. According to the comprehensive loss value, the first network is gradient-decreased to obtain a signal noise reduction model.

[0126] Referring to Figure 4(c), the network architecture of the first network used in Figure 4(c) differs from that of the first network used in Figure 4(b). Compared to the first network in Figure 4(b), in the first network in Figure 4(c), two intermediate units are connected in series between the first-scale encoding node and the first-scale decoding node, and the outputs of these two intermediate units are connected to the input of output head 1. An intermediate unit is connected in series between the second-scale encoding node and the second-scale decoding node, and the output of this intermediate unit is connected to the input of output head 2.

[0127] The first network involved in the embodiment of the present application may be, but is not limited to, any one of a UNet network (U-type network), a UNet++ network, a DenseNet network (Densely Connected Convolutional Networks) and a transformer network (transformer network).

[0128] The connection between the training signal input data and the encoding nodes may be, but is not limited to, splicing in the channel dimension or splicing in the spatial dimension. The connection between the training signal input data and the decoding nodes may be, but is not limited to, splicing in the channel dimension or splicing in the spatial dimension. The connection between the encoding nodes and the decoding nodes may be, but is not limited to, splicing in the channel dimension or splicing in the spatial dimension.

[0129] The CT raw data of the embodiment of the present application may be, but is not limited to, collected by photon counting CT, electron integration CT, or dual-source CT.

[0130] The embodiments of the present application are applicable to CT data domain noise reduction, image domain noise reduction or dual-domain joint noise reduction.

[0131] The network of the embodiment of the present application is designed for noise reduction tasks and can also be applied to other tasks, such as classification tasks and segmentation tasks.

[0132] The present application provides a signal noise reduction method, which can be executed by a computer device. The signal noise reduction method may include Figure 5 Steps shown.

[0133] Step S501: Acquire a signal to be denoised.

[0134] The signal to be de-noised may include but is not limited to CT raw data to be de-noised or CT images to be de-noised.

[0135] Step S502: obtaining a denoised signal according to the signal to be denoised and a signal denoising model.

[0136] The signal denoising model is obtained by training according to the signal denoising model training method introduced in the above embodiment.

[0137] Taking the CT image to be denoised as an example, after obtaining the CT image to be denoised, the CT image to be denoised can be input into the signal denoising model, and the denoised CT image can be obtained according to the output result of the signal denoising model.

[0138] In the above-mentioned signal denoising method, the signal denoising model is trained according to the training method of the signal denoising model of the above-mentioned embodiment. Therefore, the signal denoising model has a strong ability to retain the original information of the signal, so that the denoised signal retains more effective original information, which is helpful for subsequent signal analysis and processing.

[0139] In one embodiment, when the signal denoising model is trained according to the training method of the signal denoising model involving statistical multi-scale loss values described in the above embodiment, a denoised signal is obtained according to the signal to be denoised and the signal denoising model, including:

[0140] A first target scale is obtained based on the scale corresponding to the signal to be denoised. If the first target scale is not the largest scale among the multiple scales involved in the signal denoising model, encoding nodes and decoding nodes with scales larger than the first target scale are pruned from the signal denoising model, and the signal to be denoised is input into the pruned signal denoising model to obtain a denoised signal.

[0141] The training process shown in Figure 4 (a) belongs to the training method involving statistical multi-scale loss values. Take Figure 4 (a) as an example. After completing the training according to the training process of Figure 4 (a), a signal denoising model can be obtained. The network architecture of the signal denoising model is as follows Figure 2 shown.

[0142] If the scale of the signal to be denoised is the second scale (that is, the first target scale is the second scale), and the largest scale among the multiple scales involved in the signal denoising model is the first scale, the encoding nodes and decoding nodes of the first scale in the signal denoising model can be pruned. The input scale of the pruned signal denoising model remains consistent with the scale of the signal to be denoised (that is, the first target scale). The signal to be denoised can be input into the encoding nodes of the second scale in the pruned signal denoising model to obtain the denoised signal.

[0143] If the scale of the signal to be denoised is the first scale (that is, the first target scale is the first scale), and the largest scale among the multiple scales involved in the signal denoising model is the first scale, then the signal to be denoised can be directly input into the encoding node of the first scale in the signal denoising model without having to trim the signal denoising model to obtain the denoised signal.

[0144] The training process shown in Figure 4(c) belongs to a training method involving statistical multi-scale loss values. Taking Figure 4(c) as an example, after completing the training according to the training process in Figure 4(c), a signal denoising model can be obtained.

[0145] If the scale of the signal to be denoised is the second scale (that is, the first target scale is the second scale), and the largest scale among the multiple scales involved in the signal denoising model is the first scale, then the encoding nodes and decoding nodes of the first scale in the signal denoising model can be pruned. Furthermore, the two intermediate units connected in series between the encoding nodes and decoding nodes of the first scale can be pruned, and the connections between these two intermediate units and other nodes and between these two intermediate units and other intermediate units can be pruned. After the pruning is completed, the signal to be denoised is input into the encoding nodes of the second scale in the pruned signal denoising model to obtain the denoised signal.

[0146] In this embodiment, the signal denoising model trained according to the training method of the signal denoising model involving statistical multi-scale loss values introduced in the above embodiment can be applied to the signal denoising inference stage at different scales through cropping, without the need for multiple repeated training, thereby improving the signal denoising efficiency.

[0147] The present application provides a training method for a CT signal noise reduction model, which can be executed by a computer device. The training method for a CT signal noise reduction model may include: Figure 6 Steps shown.

[0148] Step S601 : obtaining at least one scale of training input CT raw data and multi-scale training input CT images based on the detected CT raw data.

[0149] If the training input CT raw data is of a single scale, the detected CT raw data can be directly used as the single-scale CT raw data. If the training input CT raw data is of multiple scales, the detected CT raw data can be upsampled or downsampled to obtain multi-scale training input CT raw data.

[0150] Multi-scale filtered back-projection reconstruction is performed on the detected CT raw data to obtain multi-scale training input CT images, such as a first-scale training input CT image, a second-scale training input CT image, and a third-scale training input CT image.

[0151] Step S602: Acquire the second network.

[0152] The second network includes a scale-by-scale raw data encoding path, a scale-by-scale image encoding path, a scale-by-scale image decoding path, and a scale-by-scale raw data decoding path; the end of the raw data encoding path and the starting end of the image encoding path are connected through a back projection unit, and the end of the image decoding path and the starting end of the raw data decoding path are connected through a forward projection unit.

[0153] The raw data encoding path performs scale-by-scale raw data encoding from large scale to small scale, the image encoding path performs scale-by-scale image encoding from large scale to small scale, the raw data decoding path performs scale-by-scale raw data decoding from small scale to large scale, and the image decoding path performs scale-by-scale image decoding from small scale to large scale.

[0154] The raw data coding path includes multi-scale raw data coding nodes, which are connected in series from large to small. The output end of the large-scale raw data coding node is connected to the input end of the small-scale raw data coding node, so that the raw data coding path can perform scale-by-scale raw data coding from large scale to small scale.

[0155] The image coding path includes multi-scale image coding nodes, which are connected in series from large to small. The output end of the large-scale image coding node is connected to the input end of the small-scale image coding node, so that the image coding path can perform scale-by-scale image coding from large scale to small scale.

[0156] The image decoding path includes multi-scale image decoding nodes, which are connected in series from small to large. The output end of the small-scale image decoding node is connected to the input end of the large-scale image decoding node, so that the image decoding path performs scale-by-scale image decoding from small scale to large scale.

[0157] The raw data decoding path includes multi-scale raw data decoding nodes, which are connected in series from small to large. The output end of the small-scale raw data decoding node is connected to the input end of the large-scale raw data decoding node, so that the raw data decoding path performs scale-by-scale raw data decoding from small scale to large scale.

[0158] Each scale of raw data encoding nodes has a corresponding scale of raw data decoding nodes. A connection can exist between raw data encoding nodes and raw data decoding nodes of the same scale, specifically, the output of the raw data encoding node can be connected to the input of the raw data decoding node.

[0159] Each scale of the image encoding node has a corresponding scale of the image decoding node. The image encoding node and the image decoding node of the same scale can be connected, specifically, the output of the image encoding node can be connected to the input of the image decoding node.

[0160] In some cases, the image encoding node with the smallest scale in the image encoding path and the image decoding node with the smallest scale in the image decoding path may be connected via an intermediate unit.

[0161] Encoding can be implemented through downsampling, and decoding can be implemented through upsampling; accordingly, the encoding node can be called a downsampling node, and the decoding node can be called an upsampling node.

[0162] The output end of the minimum-scale raw data encoding node and the maximum-scale image encoding node of the image encoding path are connected through a back-projection unit, and the minimum-scale raw data decoding node and the maximum-scale image decoding node are connected through a forward projection unit.

[0163] Reference Figure 7 The raw data encoding path of the second network includes raw data encoding nodes at scale a and raw data encoding nodes at scale b. The raw data decoding path of the second network includes raw data decoding nodes at scale a and raw data decoding nodes at scale b. Scale a is larger than scale b.

[0164] The output of the raw data encoding node at scale a is connected to the input of the raw data decoding node at scale a. The output of the raw data encoding node at scale b is connected to the input of the raw data decoding node at scale b.

[0165] The image encoding path of the second network includes image encoding nodes at a first scale, image encoding nodes at a second scale, and image encoding nodes at a third scale. The image decoding path of the first network includes image decoding nodes at a first scale, image decoding nodes at a second scale, and image decoding nodes at a third scale. The first scale is larger than the second scale, and the second scale is larger than the third scale.

[0166] The output of the image encoding node at the first scale is connected to the input of the image decoding node at the first scale. The output of the image encoding node at the second scale is connected to the input of the image decoding node at the second scale. The output of the image encoding node at the third scale is connected to the input of the intermediate unit, and the output of the intermediate unit is connected to the input of the image decoding node at the third scale.

[0167] The output end of the image coding node of the first scale is connected to the input end of the image coding node of the second scale, and the output end of the image coding node of the second scale is connected to the input end of the image coding node of the third scale.

[0168] The output end of the image decoding node of the third scale is connected to the input end of the image decoding node of the second scale, and the output end of the image decoding node of the second scale is connected to the input end of the image decoding node of the first scale.

[0169] A back projection unit is connected between the raw data encoding node of scale b and the image encoding node of the first scale, and a forward projection unit is connected between the raw data decoding node of scale b and the image decoding node of the first scale.

[0170] The back-projection unit mainly converts information in the CT raw data domain into information in the CT image domain. It can be constructed through a deep learning network to achieve filtered back-projection reconstruction. The forward projection unit mainly converts information in the CT image domain into information in the CT raw data domain.

[0171] Step S603: Input the training input CT raw data of at least one scale into the raw data encoding node of the corresponding scale in the raw data encoding path, and / or input the training input CT raw data of at least one scale into the raw data decoding node of the corresponding scale in the raw data decoding path.

[0172] Taking single-scale training input CT raw data as an example, if the single-scale training input CT raw data is scale a training input CT raw data, the scale a training input CT raw data can be input into the scale a raw data encoding node, and / or the scale a training input CT raw data can be input into the scale a raw data decoding node.

[0173] Taking two scales of training input CT raw data as an example, the training input CT raw data at scale a can be input into the raw data encoding node at scale a and / or the raw data decoding node at scale a. The training input CT raw data at scale b can be input into the raw data encoding node at scale b and / or the raw data decoding node at scale b.

[0174] Step S604: input the training input CT image of each scale into the image encoding node of the corresponding scale in the image encoding path, and / or input the training input CT image of each scale into the image decoding node of the corresponding scale in the image decoding path.

[0175] Take the training input CT images at the first scale, second scale, and third scale as an example.

[0176] The training input CT image of the first scale may be input into an image encoding node of the first scale in the image encoding path, and / or the training input CT image of the first scale may be input into an image decoding node of the first scale in the image decoding path.

[0177] The training input CT image of the second scale may be input into an image encoding node of the second scale in the image encoding path, and / or the training input CT image of the second scale may be input into an image decoding node of the second scale in the image decoding path.

[0178] The training input CT image of the third scale may be input into an image encoding node of the third scale in the image encoding path, and / or the training input CT image of the third scale may be input into an image decoding node of the third scale in the image decoding path.

[0179] Step S605: After inputting at least one scale of training input CT raw data and multi-scale training input CT images, the second network is trained to obtain a CT signal denoising model.

[0180] After inputting at least one scale of training input CT raw data and multi-scale training input CT images into the second network, the second network is trained to obtain a CT signal denoising model.

[0181] In the training method of the above-mentioned CT signal denoising model, multi-scale training input CT signal data is used when training the CT signal denoising model. The training input CT signal data of each scale can be input into the encoding node of the corresponding scale in the encoding path, and / or, the training input CT signal data of each scale can be input into the decoding node of the corresponding scale in the decoding path. During the network processing process, the original information of the CT signal can be better retained, thereby improving the ability of the CT signal denoising model to retain the original information of the CT signal; and, the above-mentioned training method belongs to a dual-domain denoising training process of the CT raw data domain and the CT image domain, which can further improve the ability of the CT signal denoising model to retain the original information of the CT signal.

[0182] In one embodiment, training the second network to obtain a CT signal noise reduction model includes:

[0183] Obtain raw data output results of at least one scale and multi-scale image output results of the second network; obtain a raw data loss value corresponding to at least one scale based on the raw data output results of at least one scale and the training target CT raw data of the corresponding scale; obtain an image loss value corresponding to each scale based on the image output results of each scale and the training target CT image of the corresponding scale; perform statistics on the raw data loss values corresponding to at least one scale and the multi-scale image loss values to obtain a dual-domain joint loss value; perform gradient descent on the second network based on the dual-domain joint loss value to obtain a CT signal denoising model.

[0184] In the following example, the training input CT raw data is of a single scale and the single scale is a scale.

[0185] Reference Figure 8 After the training input CT raw data at scale a passes through the input header, it is input into the raw data encoding node at scale a. After the training input CT images at scale 1 pass through the input header, they are input into the image encoding node at scale 1 in the image encoding path. After the training input CT images at scale 2 pass through the input header, they are input into the image encoding node at scale 2 in the image encoding path. After the training input CT images at scale 3 pass through the input header, they are input into the image encoding node at scale 3 in the image encoding path.

[0186] The output of the output head connected to the raw data decoding node at scale a can be used to obtain the raw data output at scale a. The output of the output head connected to the image decoding node at scale one can be used to obtain the image output at scale one. The output of the output head connected to the image decoding node at scale two can be used to obtain the image output at scale two. The output of the output head connected to the image decoding node at scale three can be used to obtain the image output at scale three.

[0187] Based on the raw data output results at scale a and the training target CT raw data at scale a, the raw data loss value corresponding to scale a is obtained. Based on the image output results at the first scale and the training target CT image at the first scale, the image loss value corresponding to the first scale is obtained. Based on the image output results at the second scale and the training target CT image at the second scale, the image loss value corresponding to the second scale is obtained. Based on the image output results at the third scale and the training target CT image at the third scale, the image loss value corresponding to the third scale is obtained.

[0188] The raw data loss corresponding to scale a, the image loss corresponding to scale 1, the image loss corresponding to scale 2, and the image loss corresponding to scale 3 are statistically analyzed to obtain a dual-domain joint loss value. Based on this dual-domain joint loss value, backpropagation is performed on the second network to obtain a gradient value. Based on this gradient value, the second network is gradient-decreased, completing one iteration. After several iterations, a CT signal denoising model is obtained.

[0189] The present application provides a CT signal noise reduction method, which can be executed by a computer device. The CT signal noise reduction method may include Figure 9 Steps shown.

[0190] Step S901: Acquire a CT signal to be de-noised.

[0191] The CT signal to be de-noised may include but is not limited to CT raw data to be de-noised or a CT image to be de-noised.

[0192] Step S902: Obtain a CT signal after noise reduction based on the CT signal to be noise reduced and a CT signal noise reduction model.

[0193] The CT signal denoising model is obtained by training according to the training method for the CT signal denoising model introduced in the above embodiment.

[0194] Taking the CT signal to be denoised as a CT image to be denoised as an example, after obtaining the CT image to be denoised, the CT image to be denoised can be input into the CT signal denoising model, and the denoised CT image can be obtained according to the output result of the CT signal denoising model.

[0195] In the above-mentioned CT signal denoising method, the CT signal denoising model is trained according to the training method of the CT signal denoising model in the above-mentioned embodiment. Therefore, the CT signal denoising model has a strong ability to retain the original information of the CT signal, so that the denoised CT signal retains more effective original information, which is helpful for subsequent CT signal analysis and processing.

[0196] In one embodiment, when the CT signal denoising model is trained according to the training method for the CT signal denoising model involving statistical dual-domain joint loss values described in the above embodiment, obtaining a denoised CT signal based on the CT signal to be denoised and the CT signal denoising model includes:

[0197] A second target scale is obtained based on the scale corresponding to the CT signal to be denoised. If the CT signal to be denoised is raw CT data to be denoised, and the second target scale is a non-maximum scale in the CT raw data domain scale involved in the CT signal denoising model, raw data encoding nodes and raw data decoding nodes in the CT signal denoising model whose CT raw data domain scale is larger than the second target scale are cropped, and the CT raw data to be denoised is input into the cropped CT signal denoising model to obtain denoised CT raw data. If the CT signal to be denoised is a CT image to be denoised, and the second target scale is a non-maximum scale in the CT image domain scale involved in the CT signal denoising model, image encoding nodes and image decoding nodes in the CT signal denoising model whose CT image domain scale is larger than the second target scale are cropped, and the CT image to be denoised is input into the cropped CT signal denoising model to obtain a denoised CT image.

[0198] Figure 8 The training process shown belongs to the training method of the CT signal denoising model involving the statistical dual-domain joint loss value. Figure 8 For example. Figure 8 After the training process is completed, the CT signal denoising model can be obtained. The network architecture of the CT signal denoising model is as follows: Figure 7 shown; reference Figure 7 The raw data processing node of the CT signal denoising model is at the upper layer of the image processing node. The raw data processing node includes a raw data encoding node and a raw data decoding node. The image processing node includes an image encoding node and an image decoding node.

[0199] If the signal to be denoised is CT raw data at scale b (i.e., the second target scale is scale b), and the maximum scale in the CT raw data domain used by the CT signal denoising model is scale a, then the raw data encoding nodes and raw data decoding nodes at scale a in the CT signal denoising model can be pruned. After the pruned CT raw data is input into the raw data encoding nodes at scale b in the pruned CT signal denoising model, the denoised CT raw data is obtained.

[0200] If the signal to be denoised is CT raw data to be denoised, and the scale is scale a (that is, the second target scale is scale a), the maximum scale in the CT raw data domain involved in the CT signal denoising model is scale a. In this case, there is no need to crop the CT signal denoising model, and the CT raw data to be denoised can be directly input into the raw data encoding node of scale a in the CT signal denoising model to obtain the denoised CT raw data.

[0201] If the signal to be denoised is a CT image and the scale is the second scale (i.e., the second target scale is the second scale), and the largest scale in the CT image domain involved in the CT signal denoising model is the first scale, then the image encoding nodes and image decoding nodes at the first scale in the CT signal denoising model can be pruned. Because the raw data processing nodes are above the image processing nodes, the raw data encoding nodes and raw data decoding nodes at each scale can also be pruned. After pruned, the CT image to be denoised is input into the image encoding nodes at the second scale in the pruned CT signal denoising model to obtain the denoised CT image.

[0202] If the signal to be denoised is a CT image and its scale is the first scale (i.e., the second target scale is the first scale), and the largest scale in the CT image domain involved in the CT signal denoising model is the first scale, then the image encoding and decoding nodes in the CT signal denoising model do not need to be pruned. Instead, the raw data encoding and decoding nodes at each scale can be pruned. After pruned, the CT image to be denoised is input into the image encoding nodes at the first scale in the CT signal denoising model to obtain the denoised CT image.

[0203] It can be understood that in a scenario where the image processing node is located above the raw data processing node, if the CT signal to be denoised is the CT raw data to be denoised, and the second target scale is a non-maximum scale in the CT raw data domain scale involved in the CT signal denoising model, in addition to cropping the raw data encoding nodes and raw data decoding nodes in the CT signal denoising model whose CT raw data domain scale is greater than the second target scale, image encoding nodes and image decoding nodes of various scales can also be cropped; the CT raw data to be denoised is input into the cropped CT signal denoising model to obtain the denoised CT raw data.

[0204] In this embodiment, the CT signal denoising model trained according to the training method involving the statistical dual-domain joint loss value introduced in the above embodiment can be applied to the CT signal denoising inference stage at different scales through cropping, without the need for multiple repeated training, thereby improving the denoising efficiency of the CT signal.

[0205] The second network involved in the above embodiments of the present application can be, but is not limited to, any one of a UNet network, a UNet++ network, a DenseNet network, and a transformer network.

[0206] Based on the same inventive concept, the embodiments of the present application further provide a device for implementing each of the methods involved above. The implementation solutions provided by the devices corresponding to each method are similar to the implementation solutions described in the corresponding method. Therefore, the specific limitations in the device embodiments corresponding to each method provided below can be found in the limitations of the corresponding method above and will not be repeated here.

[0207] The training device for the signal denoising model provided in this application includes:

[0208] A multi-scale signal acquisition module is used to obtain multi-scale training input signal data;

[0209] A first network acquisition module, configured to acquire a first network; the first network includes a scale-by-scale encoding path and a scale-by-scale decoding path;

[0210] The first network training module is used to input the training input signal data of each scale into the encoding node of the corresponding scale in the encoding path, and / or input the training input signal data of each scale into the decoding node of the corresponding scale in the decoding path, so as to train the first network and obtain a signal denoising model.

[0211] In one embodiment, the first network training module is further configured to: process the training input signal data of each scale through an input head and then input the data into a coding node of a corresponding scale in the coding path.

[0212] In one embodiment, the first network training module is further configured to: process the training input signal data of each scale through an input head and then input the data into a decoding node of a corresponding scale in the decoding path.

[0213] In one embodiment, the first network training module is further used to: obtain the output results of each scale of the first network; obtain the loss value corresponding to each scale based on the output results of each scale and the training target signal data of the corresponding scale; perform statistics on the loss values corresponding to multiple scales to obtain a comprehensive loss value; and perform gradient descent on the first network based on the comprehensive loss value to obtain a signal denoising model.

[0214] In one embodiment, the signal data includes any one of CT raw data and CT images.

[0215] In one embodiment, when the signal data includes a CT image, the multi-scale signal acquisition module is further configured to:

[0216] Multi-scale filtered back-projection reconstruction is performed on the CT raw data to obtain multi-scale training input CT images.

[0217] The signal noise reduction device provided in this application includes:

[0218] A signal acquisition module for noise reduction, used to obtain the signal for noise reduction;

[0219] The signal denoising module is used to obtain a denoised signal based on the signal to be denoised and a signal denoising model; the signal denoising model is trained according to the signal denoising model training method introduced in the above embodiment.

[0220] In one embodiment, when the signal denoising model is trained according to the training method for the signal denoising model involving statistical multi-scale loss values described in the above embodiment, the signal denoising module is further used to: obtain a first target scale based on the scale corresponding to the signal to be denoised; if the first target scale is a non-maximum scale among the multiple scales involved in the signal denoising model, then trim the encoding nodes and decoding nodes in the signal denoising model whose scales are larger than the first target scale, and input the signal to be denoised into the trimmed signal denoising model to obtain a denoised signal.

[0221] The training device for the CT signal denoising model provided in this application includes:

[0222] A CT signal acquisition module is used to obtain at least one scale of training input CT raw data and multi-scale training input CT images based on the CT raw data obtained by detection;

[0223] a second network module, configured to obtain a second network; the second network comprising a scale-by-scale raw data encoding path, a scale-by-scale image encoding path, a scale-by-scale image decoding path, and a scale-by-scale raw data decoding path; an end of the raw data encoding path and a starting end of the image encoding path are connected via a backprojection unit, and an end of the image decoding path and a starting end of the raw data decoding path are connected via a forward projection unit;

[0224] a raw data input module, configured to input the training input CT raw data of at least one scale into the raw data encoding node of the corresponding scale in the raw data encoding path, and / or input the training input CT raw data of at least one scale into the raw data decoding node of the corresponding scale in the raw data decoding path;

[0225] an image input module, configured to input the training input CT image of each scale into the image encoding node of the corresponding scale in the image encoding path, and / or input the training input CT image of each scale into the image decoding node of the corresponding scale in the image decoding path;

[0226] The second network training module is used to train the second network after inputting training input CT raw data of at least one scale and training input CT images of multiple scales to obtain a CT signal denoising model.

[0227] In one embodiment, the second network training module is further used to: obtain the raw data output results of at least one scale and the multi-scale image output results of the second network; obtain the raw data loss value corresponding to at least one scale based on the raw data output results of at least one scale and the training target CT raw data of the corresponding scale; obtain the image loss value corresponding to each scale based on the image output results of each scale and the training target CT image of the corresponding scale; perform statistics on the raw data loss values corresponding to at least one scale and the multi-scale image loss values to obtain a dual-domain joint loss value; and perform gradient descent on the second network based on the dual-domain joint loss value to obtain a CT signal denoising model.

[0228] The CT signal noise reduction device provided in this application includes:

[0229] A CT signal acquisition module for noise reduction, used to acquire the CT signal for noise reduction;

[0230] The CT signal denoising module is used to obtain a denoised CT signal based on the CT signal to be denoised and a CT signal denoising model; the CT signal denoising model is trained according to the CT signal denoising model training method introduced in the above embodiment.

[0231] In one embodiment, when a CT signal denoising model is trained according to the training method for a CT signal denoising model involving a statistical dual-domain joint loss value described in the above embodiment, the CT signal denoising module is further configured to: obtain a second target scale based on the scale corresponding to the CT signal to be denoised; if the CT signal to be denoised is raw CT data to be denoised, and the second target scale is a non-maximum scale among the CT raw data domain scales involved in the CT signal denoising model, crop the raw data encoding nodes and raw data decoding nodes in the CT signal denoising model whose CT raw data domain scales are larger than the second target scale, and input the raw CT data to be denoised into the cropped CT signal denoising model to obtain denoised raw CT data; if the CT signal to be denoised is a CT image to be denoised, and the second target scale is a non-maximum scale among the CT image domain scales involved in the CT signal denoising model, crop the image encoding nodes and image decoding nodes in the CT signal denoising model whose CT image domain scales are larger than the second target scale, and input the CT image to be denoised into the cropped CT signal denoising model to obtain a denoised CT image.

[0232] Each module in each of the above-mentioned devices may be implemented in whole or in part through software, hardware, or a combination thereof. Each of the above-mentioned modules may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in the computer device in the form of software, so that the processor can call and execute the corresponding operations of each of the above modules.

[0233] The present application provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-mentioned various method embodiments. Figure 10 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store data involved in the above method. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a method introduced in the above embodiment.

[0234] Those skilled in the art will understand that Figure 10 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0235] The present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps in the above-mentioned various method embodiments are implemented.

[0236] The present application provides a computer program product having a computer program stored thereon, wherein the computer program is used by a processor to execute the steps in the above-mentioned various method embodiments.

[0237] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0238] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), quantum computing-based data processing logic devices, artificial intelligence (AI) processors, and the like.

[0239] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0240] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for training a signal denoising model, characterized in that: The method comprises: Obtain multi-scale training input signal data; Acquire a first network; the first network includes a scale-by-scale encoding path and a scale-by-scale decoding path; Inputting the training input signal data of each scale into the encoding node of the corresponding scale in the encoding path, and / or inputting the training input signal data of each scale into the decoding node of the corresponding scale in the decoding path, so as to train the first network and obtain a signal denoising model.

2. The method according to claim 1, characterized in that Inputting the training input signal data of each scale into the encoding node of the corresponding scale in the encoding path includes: After the training input signal data of each scale is processed by the input head, it is input into the encoding node of the corresponding scale in the encoding path.

3. The method according to claim 1, characterized in that Inputting the training input signal data of each scale into the decoding node of the corresponding scale in the decoding path includes: The training input signal data of each scale is processed by the input head and then input into the decoding node of the corresponding scale in the decoding path.

4. The method according to claim 1, wherein The first network is trained to obtain a signal noise reduction model, including: Obtaining each scale output result of the first network; According to the output results of each scale and the training target signal data of the corresponding scale, the loss value corresponding to each scale is obtained; Count the loss values corresponding to multiple scales to obtain the comprehensive loss value; Performing gradient descent on the first network according to the comprehensive loss value to obtain a signal noise reduction model.

5. The method according to any one of claims 1 to 4, characterized in that The signal data includes either CT raw data or CT images.

6. The method according to any one of claims 1 to 4, characterized in that When the signal data includes CT images, multi-scale training input signal data is obtained, including: Multi-scale filtered back-projection reconstruction is performed on the CT raw data to obtain multi-scale training input CT images.

7. A signal noise reduction method, characterized in that: The method comprises: Obtaining a signal to be denoised; A denoised signal is obtained according to the signal to be denoised and the signal denoising model; the signal denoising model is obtained by training according to the signal denoising model training method according to any one of claims 1 to 6.

8. The method according to claim 7, characterized in that When the signal denoising model is trained according to the signal denoising model training method according to claim 4, obtaining a denoised signal according to the signal to be denoised and the signal denoising model includes: Obtaining a first target scale according to a scale corresponding to the signal to be denoised; If the first target scale is a non-maximum scale among the multiple scales involved in the signal denoising model, encoding nodes and decoding nodes with scales larger than the first target scale are pruned from the signal denoising model, and the signal to be denoised is input into the pruned signal denoising model to obtain a denoised signal.

9. A training method for a CT signal noise reduction model, characterized in that: The method comprises: Obtaining at least one scale of training input CT raw data and multi-scale training input CT images based on the detected CT raw data; Obtaining a second network; the second network includes a scale-by-scale raw data encoding path, a scale-by-scale image encoding path, a scale-by-scale image decoding path, and a scale-by-scale raw data decoding path; an end of the raw data encoding path and a starting end of the image encoding path are connected via a backprojection unit, and an end of the image decoding path and a starting end of the raw data decoding path are connected via a forward projection unit; Inputting the training input CT raw data of at least one scale into the raw data encoding node of the corresponding scale in the raw data encoding path, and / or inputting the training input CT raw data of at least one scale into the raw data decoding node of the corresponding scale in the raw data decoding path; Inputting the training input CT image of each scale into the image encoding node of the corresponding scale in the image encoding path, and / or inputting the training input CT image of each scale into the image decoding node of the corresponding scale in the image decoding path; After inputting at least one scale of training input CT raw data and multi-scale training input CT images, the second network is trained to obtain a CT signal denoising model.

10. The method according to claim 9, characterized in that The second network is trained to obtain a CT signal denoising model, including: Obtaining raw data output results of at least one scale and multi-scale image output results of the second network; Obtaining a raw data loss value corresponding to at least one scale according to the raw data output result of at least one scale and the training target CT raw data of the corresponding scale; According to the image output results of each scale and the training target CT image of the corresponding scale, the image loss value corresponding to each scale is obtained; The raw data loss value corresponding to at least one scale and the multi-scale image loss value are statistically analyzed to obtain the dual-domain joint loss value; According to the dual-domain joint loss value, gradient descent is performed on the second network to obtain a CT signal denoising model.

11. A CT signal noise reduction method, characterized in that: The method comprises: Acquire the CT signal to be de-noised; A CT signal after noise reduction is obtained according to the CT signal to be noise reduced and the CT signal noise reduction model; the CT signal noise reduction model is trained according to the training method of the CT signal noise reduction model according to claim 9 or 10.

12. The method according to claim 11, characterized in that When the CT signal denoising model is trained according to the training method for the CT signal denoising model according to claim 10, obtaining the denoised CT signal according to the CT signal to be denoised and the CT signal denoising model includes: Obtaining a second target scale according to a scale corresponding to the CT signal to be denoised; If the CT signal to be denoised is CT raw data to be denoised, and the second target scale is a non-maximum scale among the CT raw data domain scales involved in the CT signal denoising model, pruning raw data encoding nodes and raw data decoding nodes in the CT signal denoising model whose CT raw data domain scales are larger than the second target scale, and inputting the CT raw data to be denoised into the denoised CT signal denoising model after pruning to obtain denoised CT raw data. If the CT signal to be denoised is a CT image to be denoised, and the second target scale is a non-maximum scale among the CT image domain scales involved in the CT signal denoising model, image encoding nodes and image decoding nodes in the CT signal denoising model whose CT image domain scales are larger than the second target scale are cropped, and the CT image to be denoised is input into the cropped CT signal denoising model to obtain a denoised CT image.

Images