WO2024032098A1 - 伪影去除模型的训练方法、装置、设备、介质及程序产品 - Google Patents

伪影去除模型的训练方法、装置、设备、介质及程序产品 Download PDF

Info

Publication number
WO2024032098A1
WO2024032098A1 PCT/CN2023/096836 CN2023096836W WO2024032098A1 WO 2024032098 A1 WO2024032098 A1 WO 2024032098A1 CN 2023096836 W CN2023096836 W CN 2023096836W WO 2024032098 A1 WO2024032098 A1 WO 2024032098A1
Authority
WO
WIPO (PCT)
Prior art keywords
artifact
model
removal
sample
image
Prior art date
Application number
PCT/CN2023/096836
Other languages
English (en)
French (fr)
Inventor
王红
郑冶枫
Original Assignee
腾讯科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技（深圳）有限公司 filed Critical 腾讯科技（深圳）有限公司
Publication of WO2024032098A1 publication Critical patent/WO2024032098A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/73—Deblurring; Sharpening
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/50—Image enhancement or restoration using two or more images, e.g. averaging or subtraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]

Definitions

  • This application relates to the field of machine learning, and in particular to a training method, device, equipment, media and program products for an artifact removal model.
  • CT computed tomography
  • a dual-domain network (DuDoNet) is used to remove artifacts in CT images.
  • the dual-domain network is composed of two modules.
  • a CT value processing window is preset to perform chord diagrams containing artifacts respectively.
  • Image domain processing, as well as image domain processing of CT images containing artifacts, output the repaired sinusoidal image and enhanced CT image, and finally use the back-projection layer to output the CT image with artifacts removed.
  • the method of removing CT images containing artifacts through dual-domain networks results in lower restoration realism, lower image accuracy, and poorer image quality of the CT images after artifact removal.
  • Embodiments of the present application provide a training method, device, equipment, media and program products for an artifact removal model, which can improve the accuracy of the output results of the artifact removal model.
  • the technical solutions are as follows:
  • a training method for an artifact removal model is provided, the method is executed by a computer device, and the method includes:
  • the reference image is an image generated after scanning a sample detection object that does not contain an implant.
  • the artifact image is a reference image that contains an artifact.
  • the artifact is the shadow produced by the implant during the scanning process;
  • the artifact image is input into multiple sample removal models to obtain artifact removal results corresponding to the artifact images respectively output by the multiple sample removal models.
  • Different sample removal models correspond to different preset window ranges.
  • the sample removal model is used to remove artifacts in the artifact image based on the corresponding preset window range;
  • multiple sample removal models are trained to obtain an artifact removal model composed of multiple artifact removal sub-models, and the artifact removal sub-model is used to based on the corresponding preset Artifact removal is performed on the target image within the window range.
  • a training device for an artifact removal model includes:
  • An acquisition module used to acquire a reference image and an artifact image that match the image content.
  • the reference image is an image generated after scanning a sample detection object that does not contain an implant.
  • the artifact image is a reference image that contains artifacts.
  • the pseudo The shadow is the shadow produced by the implant during the scanning process;
  • An input module configured to input the artifact image into multiple sample removal models, and obtain artifact removal results corresponding to the artifact images respectively output by the multiple sample removal models.
  • Different sample removal models correspond to different presets. Window range, the sample removal model is used to remove artifacts in the artifact image based on the corresponding preset window range;
  • a determination module configured to determine prediction loss values corresponding to multiple sample removal models based on the pixel difference between the artifact removal result and the reference image
  • the input module is used to input the predicted loss values corresponding to multiple sample removal models into the sample weight model, and output the weight parameters corresponding to the multiple predicted loss values.
  • the weight parameters are used to evaluate the sample removal model. Parameter update and weight adjustment;
  • a training module configured to train multiple sample removal models based on the predicted loss value and the weight parameter to obtain an artifact removal model composed of multiple artifact removal sub-models, where the artifact removal sub-model is used to Artifact removal is performed on the target image based on the corresponding preset window range.
  • a computer device includes a processor and a memory.
  • the memory stores at least one instruction, at least a program, a code set or an instruction set.
  • the at least one instruction, the at least A program, the code set or the instruction set is loaded and executed by the processor to implement the training method of the artifact removal model as described in any of the above embodiments of the present application.
  • a computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set.
  • the at least one instruction, the at least one program, the code is loaded and executed by the processor to implement the training method of the artifact removal model as described in any of the above embodiments of the present application.
  • a computer program product or computer program including computer instructions stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the training method of the artifact removal model described in any of the above embodiments.
  • Multiple sample removal models are trained using reference images and artifact images that match the image content.
  • the artifact images are input into multiple sample removal models and then multiple artifact removal results are output respectively, and multiple artifact removal results are determined.
  • the predicted loss value between the shadow removal result and the reference image is input into the sample weight model. Finally, the weight parameters corresponding to each predicted loss value are obtained.
  • Multiple sample removal models are trained based on the predicted loss value and weight parameters.
  • An artifact removal model containing multiple artifact removal sub-models is obtained, that is, by using weight parameters and prediction loss values to train multiple sample removal models corresponding to different preset window ranges, so that the artifact removal model obtained by the final training
  • the model can output artifact removal images corresponding to different window ranges, meet the artifact removal requirements of different images, and improve the artifact removal accuracy of the artifact removal results.
  • Figure 1 is a schematic diagram of a training method for an artifact removal model provided by an exemplary embodiment of the present application
  • Figure 2 is a schematic diagram of the implementation environment provided by an exemplary embodiment of the present application.
  • Figure 3 is a flow chart of a training method for an artifact removal model provided by an exemplary embodiment of the present application
  • Figure 4 is a flow chart of a training method for an artifact removal model provided by another exemplary embodiment of the present application.
  • Figure 5 is a schematic diagram of the DICD-Net model provided by another exemplary embodiment of the present application.
  • Figure 6 is a schematic diagram of a network structure provided by an exemplary embodiment of the present application.
  • Figure 7 is a schematic diagram of the network structure of the sample weight model provided by an exemplary embodiment of the present application.
  • Figure 8 is a schematic diagram of multiple sample removal models provided by an exemplary embodiment of the present application.
  • Figure 9 is a schematic diagram of a training method for an artifact removal model provided by another exemplary embodiment of the present application.
  • Figure 10 is a schematic diagram of the application process of the artifact removal model provided by an exemplary embodiment of the present application.
  • Figure 11 is a schematic diagram of the training method of the artifact removal model provided by an exemplary embodiment of the present application.
  • Figure 12 is a schematic diagram of the artifact removal model processing process provided by an exemplary embodiment of the present application.
  • Figure 13 is a structural block diagram of a training device for an artifact removal model provided by an exemplary embodiment of the present application.
  • Figure 14 is a structural block diagram of a training device for an artifact removal model provided by another illustrative embodiment of the present application.
  • Figure 15 is a schematic structural diagram of a server provided by an exemplary embodiment of the present application.
  • Window technology is a display technology used to observe normal tissues or lesions of different densities in computed tomography (CT) examinations, including window width (Window Width) and window level (Window Level). Due to various Various tissue structures or lesions have different CT values. Therefore, when you want to display the details of a specified tissue structure on a CT image, you need to select a window width and window level suitable for observing the specified tissue structure to form a specified window range to obtain the specific information for the specified tissue structure. Specify the optimal display mode of the tissue structure and generate a grayscale image corresponding to the CT value in the specified window range.
  • Figure 1 shows a schematic diagram of a training method for an artifact removal model provided by an exemplary embodiment of the present application.
  • a training image set 100 is obtained, where in the training image set 100 It includes a reference image 101 and an artifact image 102 whose image content matches.
  • the reference image 101 and the artifact image 102 belong to a sample image pair.
  • the reference image 101 and the artifact image 102 are both computed tomography scans of the abdomen.
  • the CT image obtained after CT while the artifact image 102 is an image containing artifacts (an abdominal CT image contaminated by artifacts), and the reference image 101 does not contain artifacts (an abdominal CT image not contaminated by artifacts).
  • the artifact image 102 is input into multiple sample removal models 110, and artifact removal results 111 of the artifact image 102 are respectively output.
  • Each sample removal model in the multiple sample removal models 110 corresponds to a different preset window range, so the artifact
  • the shadow removal result 111 is implemented as an artifact removal image corresponding to different preset window ranges (for example: image 1111 is a CT image in the [-1000, 2000] HU window range, image 1112 is a CT image in [-320, 480] HU CT image under the window range, image 1113 is a CT image under the [-160, 240] HU window range).
  • the prediction loss values 112 respectively corresponding to the plurality of sample removal models 110 are determined according to the pixel difference between the artifact removal result 111 and the reference image 101 .
  • the weight parameter 121 is used to adjust the weight of the parameter update of the sample removal model 110. According to the predicted loss value 112 and the weight parameter 121 Multiple sample removal models 110 are trained to finally obtain an artifact removal model 130 composed of multiple artifact removal sub-models, where the artifact removal model 130 is used to remove artifacts from an input target image containing artifacts.
  • the implementation environment involves a terminal 210 and a server 220.
  • the terminal 210 and the server 220 are connected through a communication network 230.
  • the terminal 210 sends an artifact removal request to the server 220 , where the artifact removal request includes a target scan image.
  • the target scan image is implemented as a CT image contaminated by metal (that is, for During the CT scanning process of a designated part of the human body, the generated CT image is affected by the metal implanted in the designated part, resulting in metal artifacts).
  • the server 220 Artifact removal is performed on the metal artifacts, an artifact removal result is generated, and the artifact removal result is fed back to the terminal 210 .
  • the server 220 includes an artifact removal model 221.
  • the server 220 inputs the target scanned image into the artifact removal model 221 and outputs the artifact removal result.
  • the artifact removal result refers to the artifact area identified in the target scanned image.
  • the artifact removal model 221 is achieved by inputting the artifact images 222 used for training into multiple sample removal models 223, Multiple artifact removal results are output, and multiple prediction loss values 224 are determined based on the pixel difference between the artifact removal results and the reference image (an image that matches the image content of the artifact image 222 and does not contain artifacts), and the prediction loss is
  • the value 224 is input to the sample weight model 225, and the output is weight parameters 226 corresponding to multiple prediction loss values 224, which are obtained by training the sample removal model 223 based on the weight parameters 226 and the prediction loss values 224.
  • the above-mentioned terminal 210 can be a mobile phone, a tablet computer, a desktop computer, a portable notebook computer, a smart TV, a smart vehicle, and other terminal devices in various forms, which are not limited in the embodiments of the present application.
  • server 220 can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or it can provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, Cloud servers for basic cloud computing services such as network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms.
  • cloud services such as network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms.
  • CDN Content Delivery Network
  • cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or local area network to realize data calculation, storage, processing, and sharing.
  • the above-mentioned server 220 can also be implemented as a node in the blockchain system.
  • the information including but not limited to user equipment information, user personal information, etc.
  • data including but not limited to data used for analysis, stored data, displayed data, etc.
  • signals involved in this application All are authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.
  • the reference images and artifact images used for training and the verification images used for model verification involved in this application were obtained with full authorization.
  • Figure 3 shows a flow chart of the training method of the artifact removal model provided by an exemplary embodiment of this application.
  • This method is performed by a computer.
  • Device execution for example, this method can be executed by the terminal, or can be executed by the server, or can also be executed by the terminal and the server jointly.
  • the method is explained by being executed by the server.
  • the method include:
  • Step 310 Obtain the reference image and the artifact image whose image content matches.
  • the reference image is an image generated after scanning a sample detection object that does not contain implants
  • the artifact image is a reference image that contains artifacts.
  • the artifact is an image generated during the scanning process of a sample detection object that contains implants. Shadows caused by incoming objects. In other words, the above-mentioned artifacts are shadows produced by the implant during the scanning process.
  • the reference image refers to the medical image generated after scanning the sample detection object through a specified scanning technology.
  • the artifact image is implemented as a grayscale image.
  • the sample detection object is used to represent a specified tissue or organ (such as heart, abdomen, chest, lungs, etc.).
  • the reference image is a medical image obtained by scanning a sample detection object that does not contain an implant, that is, the reference image is a medical image that is not affected by the implant.
  • an implant refers to an object that contains metal parts and is implanted in the detection object, such as at least one of implant types such as dentures, pacemakers, and stents, which is not limited.
  • the specified scanning technology refers to CT scanning technology. Therefore, the images involved in the embodiments of this application are all CT images.
  • image content matching means that the content contained in the reference image and the content contained in the artifact image are the same.
  • the reference image and the artifact image are both CT images generated after a CT scan of the same abdomen.
  • the difference between the artifact image and the reference image is that the artifact image is a reference image containing artifacts. That is, the reference image and the artifact image are implemented as a sample image pair.
  • artifacts represent shadows (or dark bands) produced on the image by objects other than the sample detection object during the scanning process of the detection object.
  • Step 320 Input the artifact images into multiple sample removal models, and output artifact removal results corresponding to the artifact images respectively.
  • a computer device (such as a server) inputs artifact images into multiple sample removal models to obtain multiple samples to remove Artifact removal results corresponding to the artifact images output by the model.
  • sample removal models correspond to different preset window ranges, and the sample removal models are used to remove artifacts in the artifact image based on the corresponding preset window range.
  • the artifact removal result refers to removing the artifacts contained in the artifact image through the sample removal model, and outputting a scan image corresponding to the preset window range corresponding to the sample removal model, that is, the scan The image does not contain artifacts.
  • the contrast relationship of each area presented in the above-mentioned scan image is the same as the contrast relationship of each area presented in the artifact image, or is different, and this is not limited.
  • the preset window range is used to represent the contrast relationship between areas in the scanned image.
  • the scanned image includes area a and area b.
  • the brightness of area a in the scanned image is higher than Area b
  • the preset window range B the brightness of area a in the scanned image is lower than area b. That is to say, when the same scanned image corresponds to different preset window ranges, the contrast between the displayed areas is different, which facilitates targeted viewing of the designated area.
  • the preset window range corresponding to the sample removal model is a preset fixed window range, for example: the preset window range corresponding to the sample removal model A is [-1000, 2000] HU; or, the preset window range corresponding to the sample removal model A
  • the window range is an adjustable window range set according to actual needs and is not limited.
  • multiple sample removal models correspond to the same model structure; or, multiple sample removal models correspond to different model structures, which is not limited.
  • Step 330 Based on the pixel difference between the artifact removal result and the reference image, determine the prediction loss values corresponding to the multiple sample removal models.
  • the prediction loss value is used to represent the difference between each pixel between the artifact removal result and the reference image.
  • a loss function is preset, and the distance between the pixel value corresponding to the artifact removal result and the pixel value corresponding to the reference image is calculated through the loss function, and the calculated result is used as the corresponding prediction of multiple sample removal models. loss value.
  • Step 340 Input the prediction loss values corresponding to the multiple sample removal models into the sample weight model, and output the weight parameters corresponding to the multiple prediction loss values.
  • the weight parameter is used to adjust the weight of the parameter update of the sample removal model.
  • the prediction loss values corresponding to multiple sample removal models are input into the sample weight model respectively, and the scalar result is output as the weight parameter corresponding to a single prediction loss value.
  • the weight parameters corresponding to the multiple predicted loss values output are different; or, there are at least two weight parameters corresponding to the predicted loss values. Same, no limitation is placed on this.
  • the weight parameter is used to assign different weights to each predicted loss value in the process of training the sample removal model through the predicted loss value.
  • the multiple prediction loss values are simultaneously input into the sample weight model, and the weight parameters corresponding to the multiple prediction loss values are simultaneously output, that is, The weight parameters corresponding to multiple prediction loss values are obtained at the same time; or, after each prediction loss value corresponding to the sample removal model is obtained, it is input into the sample weight model, and the weight parameters corresponding to the sample removal model are output, that is, That is, the weight parameters corresponding to multiple prediction loss values are obtained sequentially, and this is not limited.
  • Step 350 Train the sample removal model based on the predicted loss value and weight parameters to obtain an artifact removal model composed of multiple artifact removal sub-models.
  • the artifact removal sub-model is used to remove artifacts from the target image based on the corresponding preset window range.
  • parameter adjustment is performed on the first model parameter of the sample removal model according to the predicted loss value and weight parameter, and the artifact removal sub-model is determined based on the adjusted parameters.
  • a single artifact removal sub-model is obtained after training a single sample removal model, and finally multiple artifact removal sub-models constitute an artifact removal model.
  • the training process of each sample removal model is performed simultaneously, or the training process of each sample removal model is performed sequentially, that is, , after training the first sample removal model, start training the second sample removal model, there is no limit to this.
  • the training method of the artifact removal model uses reference images and artifact images with matching image content to train multiple sample removal models, wherein the artifact images are input during the training process.
  • multiple artifact removal results are output respectively
  • the prediction loss value between the multiple artifact removal results and the reference image is determined
  • the prediction loss value is input into the sample weight model to finally obtain the weight corresponding to each prediction loss value.
  • Parameters multiple sample removal models are trained according to the predicted loss value and weight parameter, and finally an artifact removal model containing multiple artifact removal sub-models is obtained.
  • the weight parameter and the predicted loss value multiple corresponding different preset windows are obtained.
  • the sample removal model is trained with a range of samples, so that the finally trained artifact removal model can output artifact removal images corresponding to different window ranges, meet the artifact removal requirements of different images, and improve the artifact removal accuracy of the artifact removal results. .
  • the training process of the sample removal model is implemented as a multiple loop iterative training process.
  • Figure 4 shows the A flowchart of a training method for an artifact removal model provided in an exemplary embodiment is applied.
  • the method is executed by a computer device.
  • the method can be executed by a terminal, or by a server, or it can also be executed by both a terminal and a server.
  • the method is explained by being executed by the server, as shown in Figure 3, that is, step 350 includes step 351, step 352 and step 353, and step 340 also includes step 341.
  • the method includes the following steps:
  • Step 310 Obtain the reference image and the artifact image whose image content matches.
  • the reference image is an image generated after scanning a sample detection object that does not contain implants
  • the artifact image is a reference image that contains artifacts.
  • the artifact is an image generated during the scanning process of a sample detection object that contains implants. Shadows caused by incoming objects.
  • a metal artifact is taken as an example for illustration.
  • the artifact made of metal is implemented as an artifact with a strip structure.
  • the reference image is a CT image generated after performing a CT scan on the sample detection image
  • the artifact image is a CT image containing metal artifacts. That is, the current artifact image is a CT scan due to the presence of metal in the sample detection image.
  • the resulting image includes metallic artifacts.
  • the artifact image and the reference image are obtained directly from an authorized public data set; alternatively, the reference image is an image obtained directly from a public data set, and the artifact image is based on the reference image, combined with different metal
  • the corresponding metal mask information, the artificially synthesized reference image containing metal artifacts, is not limited to this.
  • the reference image and the artifact image are implemented as a sample image pair.
  • the reference image and the artifact image are scanned images corresponding to the same sample detection object.
  • Step 320 Input the artifact images into multiple sample removal models, and output artifact removal results corresponding to the artifact images respectively.
  • sample removal models correspond to different preset window ranges, and the sample removal models are used to remove artifacts in the artifact image based on the corresponding preset window range.
  • the preset window range of the sample removal model is a preset fixed window range.
  • the preset window range of sample removal model A is fixed to [-320, 480]HU.
  • different sample removal models correspond to different preset window ranges.
  • the sample removal model is used to remove artifacts from the artifact image, adjust the artifact image to a display mode corresponding to the preset window range according to the preset window range, and output the result: is the artifact removal result.
  • the artifact image is a CT image containing metal artifacts with a window range of [-1000,2000]HU. Enter the artifact image into the sample removal model (the default window range is [-320,480]HU ), first adjust the window range of the artifact image to [-320,480]HU and then input it into the sample removal model.
  • the sample removal model removes the metal artifacts in the artifact image and outputs the image as the artifact removal result.
  • sample removal model in the embodiment of the present application can be implemented as a deep interpretable convolutional dictionary network (Deep Interpretables Convolutional Dictionary Network, DICD-Net), a convolutional neural network (Convolutional Neural Network, CNN), U- Net network and other neural network models are not limited to this.
  • DICD-Net Deep Interpretables Convolutional Dictionary Network
  • CNN convolutional Neural Network
  • U- Net network and other neural network models are not limited to this.
  • DICD-Net500 includes N iterative processes. In any iterative process, the A single iteration process consists of network and X-Net are composed in sequence.
  • the artifact image 510 is input into DICD-Net for N iterative removal (N stages), and the artifact removal result 520 corresponding to the artifact image is output, where the artifact removal result 520 is implemented by N
  • the CT image obtained after removing artifacts from the network and X network. in, The network and X network complete the feature layer respectively. and update of artifact removal results 520.
  • FIG. 6 shows a schematic diagram of the network structure provided by an exemplary embodiment of the present application.
  • the structural diagram 600 corresponding to the network and the X network includes Network structure 610 and X network structure 620.
  • the network structure of network 610 can refer to the following formula 1:
  • Each residual block in the residual network 630 includes in turn: a convolution layer, a batch normalization layer (Batch
  • Y represents the input artifact image (and the artifact image contains metal artifacts);
  • X represents the artifact removal result (X (n-1) represents the artifact removal result obtained in the n-1th iteration stage);
  • I represents the mask information (Mask) corresponding to non-metallic artifacts in the artifact image; Indicates a recurring pattern of metal artifacts.
  • the display method of metal artifacts. is the feature layer, which represents the strip artifact structure of the metal artifact.
  • Feature Map feature map
  • ⁇ 1 is The update step size of the network.
  • the network structure of the X network 620 can refer to the following formula 2:
  • X (n) represents the output result of the X network during the nth iteration removal process
  • each residual block in the residual network 640 sequentially includes: a convolution layer, a batch normalization layer (Batch Normalization), a ReLU layer, a convolution layer, a batch normalization layer, and a cross-link layer.
  • Y represents the input artifact image (and the artifact image contains metal artifacts);
  • X represents the artifact removal result (X (n-1) represents the artifact removal result obtained in the n-1th iteration stage);
  • I represents the mask information (Mask) corresponding to non-metallic artifacts in the artifact image, Indicates a recurring pattern of metal artifacts.
  • the display method of metal artifacts. is the feature layer, representing the strip artifact structure of the metal artifact.
  • Feature Map feature map
  • eta 2 is the update step size of the X network.
  • the n-1th feature map obtained in the n-1th iteration removal stage And the artifact removal result (X (n-1) ) obtained in the n-1th iteration removal stage is input In the network, the output is the n-th feature map obtained in the n-th iteration removal stage.
  • the nth feature map And the artifact removal result (X (n-1) ) obtained in the n-1th iteration stage is input into the X network, and the artifact removal result (X (n) ) obtained in the nth iteration stage is output.
  • Step 330 Based on the pixel difference between the artifact removal result and the reference image, determine the prediction loss values corresponding to the multiple sample removal models.
  • a preset loss function is used to calculate the pixel value difference between the artifact removal result and the reference image through the preset loss function, and the calculated results are used as prediction loss values corresponding to multiple sample removal models.
  • the preset loss function is Among them, b represents the sample removal model corresponding to the b-th preset window range, and ⁇ represents the first model parameter of the sample removal model.
  • Step 341 Input the prediction loss obtained by the s-1th iterative training into the sample weight model obtained by the s-th iterative training, and output the weight parameters corresponding to the s-th iterative training.
  • the sample weight model also needs to be trained.
  • a verification reference image and a verification artifact image whose image content matches are obtained; the verification artifact image is input into multiple sample removal models, and verification removal results corresponding to the verification artifact image are output respectively; based on the verification removal result and Verify the pixel differences between reference images and determine the verification loss values corresponding to multiple sample removal models; based on the verification loss values, train the sample weight model.
  • the verification reference image is a CT image obtained by performing a CT scan on the verification detection object
  • the verification artifact image is a CT image containing metal artifacts, where the image content corresponding to the verification reference image and the verification artifact image is the same
  • the verification reference image and the verification artifact image are both CT images obtained after a CT scan of the same abdomen, where the verification artifact image includes metal artifacts and the verification reference image does not contain artifacts).
  • the verification reference image and the verification artifact image are an image pair in the verification sample.
  • the verification artifact image and the artifact image belong to different CT images.
  • the verification artifact image is input into multiple sample removal models. After the artifacts in the artifact image are removed through multiple sample removal models, an image corresponding to the preset window range of the sample removal model is generated as the verification removal model. result. Among them, each sample removal model corresponds to a verification removal result.
  • the loss function is preset, and the difference between the pixel points between multiple verification removal results and the verification reference image is calculated through the loss function, which is used as the verification loss value corresponding to the sample removal model.
  • multiple sample removal models respectively correspond to multiple verification loss values.
  • the preset loss function is implemented as The output result of this loss function is used to represent the verification loss value corresponding to the b-th sample removal model.
  • the second model parameters of the sample weight model are adjusted through multiple verification loss values.
  • the second model parameters of the sample weight model are gradient adjusted to obtain the samples corresponding to the sth iterative training. weight model.
  • the process of training the sample weight model includes training the sample weight model for N iterations (corresponding to the N iteration processes included in the above-mentioned DICD-Net network), that is, during the N iteration processes , the sample weight model and the sample removal model are updated iteratively and alternately.
  • the solution implemented by this application is to remove the verification loss values corresponding to the model through multiple samples, and train the sample weight model.
  • the process of training the sample weight model for a single verification loss value is used as an example to illustrate.
  • For the sample weight model For the method of gradient adjustment of the second model parameters, please refer to Formula 3 for details:
  • ⁇ (s) represents the sample weight model corresponding to the s-th iteration training
  • ⁇ is the preset second learning rate, which is used to represent the update step size for training the sample weight model
  • a mapping function about ⁇ is set to represent The mapping relationship between the first model parameter ( ⁇ ) corresponding to the sample removal model and the second model parameter ( ⁇ ) corresponding to the sample weight model. That is, based on the first model parameters obtained by s-1 iterative training, the corresponding mapping relationship between the first model parameter and the second model parameter in the s-1 iterative training process is determined; based on the mapping relationship, the corresponding mapping relationship between the first model parameter and the second model parameter is determined. The validation loss value obtained from s-1 iterations of training.
  • ⁇ represents the preset first learning rate
  • ⁇ represents the preset first learning rate
  • Step 351 Determine the weighted loss values corresponding to the multiple sample removal models based on the multiple predicted loss values and the weight parameters corresponding to the multiple loss values.
  • the loss function can be preset and the model parameters of multiple DICD-Nets can be adjusted separately to minimize the sum of the output loss values of the loss function, which is the artifact.
  • the final optimization goal of the removal model please refer to Formula 5:
  • DICD-Net B sample removal models
  • the weight parameters corresponding to the sample removal model are obtained through the sample weight model, therefore, in, Represents the network structure of the sample weight model, whose input is the predicted loss value
  • the output is the weight parameter W, and the corresponding second model parameter in the network structure is ⁇ .
  • MLP Multi-Layer Perception
  • Figure 7 shows a schematic diagram of the sample weight model network structure provided by an exemplary embodiment of the present application.
  • the MLP network is currently displayed.
  • the network includes an input layer 710, a hidden layer 720 and an output layer 730.
  • the input layer is the predicted loss value 711
  • the output layer 730 is the weight parameter 731 corresponding to the predicted loss value 711, where , the hidden layer 720 contains multiple neurons, and the number of neurons in the hidden layer 720 is limited according to actual needs.
  • the sth weighted loss value is determined based on the s-1th predicted loss value and the sth weight parameter.
  • the weighted loss value is used to represent the corresponding predicted loss value of the sample removal model after weight adjustment.
  • Formula 6 For details, please refer to Formula 6:
  • Step 352 Adjust the first model parameters of the multiple sample removal models respectively based on the weighted loss values corresponding to the multiple sample removal models to obtain multiple artifact removal sub-models.
  • the first model parameters of multiple sample removal models are gradient adjusted according to multiple weighted loss values, and the adjusted parameters are obtained as model parameters corresponding to the artifact removal model, and then the artifact removal model is obtained.
  • the specific training process is Please refer to Formula 7:
  • ⁇ (s-1) is used to represent the first model parameters obtained during the s-1th iterative training process (multiple first model parameters corresponding to the B sample removal models in a single iterative training), Used to represent the gradient adjustment of the s-1th first model parameter through the weighted loss value corresponding to the sth iterative training, that is, based on the weighted loss value corresponding to the sth iterative training, the sth sample removal model is The first model parameters obtained from -1 iterative training are gradient adjusted to obtain the first model parameters corresponding to the s-th iterative training, and the s+1-th cyclic adjustment is performed until the artifact removal model training is completed, s ⁇ 1 and s is an integer.
  • the corresponding training sequence is Formula 4 (parameterizing ⁇ to obtain the mapping function about ⁇ ), Formula 3 (parameterizing the sample weight model Gradient adjustment of the second model parameters) and Formula 7 (gradient adjustment of the first parameter models corresponding to multiple sample removal models).
  • ⁇ * ( ⁇ ) is used to represent the optimal solution corresponding to multiple first model parameters (that is, the first parameter that meets the adjustment effect conditions).
  • the optimal solution of multiple first model parameters can be calculated by calculating the weighted loss value. The corresponding minimum value is obtained.
  • ⁇ * is used to represent the optimal solution of the second model parameters.
  • the optimal solution can be obtained by calculating the minimum value corresponding to the predicted loss value, so that the second model parameters reach the optimal solution (or infinitely approximate the optimal solution) , the second model parameter is used as the second parameter obtained by the final training of the sample weight model.
  • Formula 8 and Formula 9 are only training goals set for the sample removal model and sample weight model.
  • the training process of the sample removal model and sample weight model is through Formula 4, Formula 3 and Formula 7 performs iterative loop training in the specified order.
  • pseudo shadow models corresponding to multiple sample removal models are generated, that is, a single artifact shadow model corresponds to a single sample removal model.
  • Step 353 Use multiple artifact removal sub-models as artifact removal models.
  • the artifact removal model includes multiple sample removal models that have completed training, that is, the artifact removal model includes multiple artifact removal sub-models.
  • the first model parameter obtained by the latest adjustment is determined as the first parameter; or in response to the adjustment effect of the first model parameter meeting the adjustment
  • the effect condition determines the first model parameter as the first parameter, and adjusts the effect condition to express the restriction requirement on the predicted loss value.
  • the model parameters in the artifact removal model are the first parameters.
  • the training target for multiple sample removal models is to obtain the first model parameters corresponding to the multiple sample removal models through gradient adjustment. The result is used as the final corresponding first parameter of the artifact removal model.
  • the times threshold is a preset specified number of times or can be set according to the training situation. For example, if the model is trained for 100 times to remove multiple samples, then 100 times is the times threshold. When the model is removed for multiple samples, When the number of iterative adjustments of the first model parameters reaches 100 times, the multiple first model parameters obtained from the 100th training are determined as the first parameters.
  • the adjustment effect condition means that after gradient adjustment is performed on multiple sample removal models, after the artifact image is input and the predicted loss value between the output artifact removal result and the reference image meets the adjustment effect condition,
  • the first model parameters respectively corresponding to the plurality of sample removal models are determined as the first parameters of the artifact removal network. That is, the first model parameters of this training meet the training objectives of the sample removal model.
  • the training method of the artifact removal model uses reference images and artifact images with matching image content to train multiple sample removal models, wherein the artifact images are input during the training process.
  • multiple artifact removal results are output respectively
  • the prediction loss value between the multiple artifact removal results and the reference image is determined
  • the prediction loss value is input into the sample weight model to finally obtain the weight corresponding to each prediction loss value.
  • Parameters multiple sample removal models are trained according to the predicted loss value and weight parameter, and finally an artifact removal model containing multiple artifact removal sub-models is obtained.
  • the weight parameter and the predicted loss value multiple corresponding different preset windows are obtained.
  • the sample removal model is trained with a range of samples, so that the finally trained artifact removal model can output artifact removal images corresponding to different window ranges, meet the artifact removal requirements of different images, and improve the artifact removal accuracy of the artifact removal results. .
  • the sample weight model is first trained, and then the prediction loss value is input into the sample weight model obtained according to the training and the weight parameter is output.
  • the first step of the model is removed from multiple samples.
  • the parameters of each model are gradient adjusted respectively, so that during the iterative training process, the sample weight model and the sample removal model are alternately trained iteratively, and the sample weight model can be trained to assist in training the sample removal model, improving the accuracy and accuracy of model training. Effect.
  • different preset window ranges also include a process of window conversion.
  • Step 310 Obtain the reference image and the artifact image whose image content matches.
  • the reference image is an image generated after scanning a sample detection object that does not contain implants
  • the artifact image is a reference image that contains artifacts.
  • the artifact is an image generated during the scanning process of a sample detection object that contains implants. Shadows caused by incoming objects.
  • the artifact image is a CT image synthesized from the authorized CT image obtained in advance in the public data set.
  • CT images that are not affected by metal in the public data set are first used as reference images, as well as different types of metal masks (Mask).
  • Mosk metal masks
  • the CT images and metal The mask is used for image synthesis, and CT images containing metal artifacts are synthesized as training data.
  • Each training data is randomly cropped into image blocks with a side length of 64 ⁇ 64, and then horizontal mirror flipping and vertical mirror flipping are randomly performed with a probability of 0.5 to finally generate different artifact images.
  • the reference image and the artifact image are implemented as a sample image pair.
  • multiple sample removal models respectively correspond to different preset window ranges and are used to mark the contrast relationship displayed in each area of the scanned image for different CT values.
  • three sample removal models are used as Example Note that, for the three sample removal models included in this embodiment, their corresponding preset window ranges are [-1000, 2000] HU, [-320, 480] HU and [-160, 240] HU respectively. , there is no restriction on this.
  • Figure 8 shows a schematic diagram of multiple sample removal models provided by an exemplary embodiment of the present application.
  • a schematic diagram of a multi-window sample removal model is currently displayed, wherein the multi-window
  • the sample removal model includes three sample removal models 810, sample removal models 820 and sample removal models 830 corresponding to three different preset window ranges.
  • the sample removal model is implemented as a DICD-Net structured model.
  • the artifact image 801 is input into the sample removal model 810, and the first prediction area 811 corresponding to the sample removal model 810 is output, where the first prediction area 811 is used to represent the first preset window range through the sample removal model 810, A CT image generated after removing metal artifacts from the artifact image 801.
  • the artifact image before the artifact image is input into the sample removal model, it also has a corresponding preset window range. Therefore, the preset window range for the artifact image and the preset window range of the input sample removal model are different.
  • the window range is specified, the artifact image needs to be window converted before inputting the sample removal model to generate an artifact image consistent with the preset window range corresponding to the sample removal model.
  • the artifact image 801 and the first prediction area 811 are input into the window conversion layer (Window Layer, marked W in Figure 8) for window conversion, and the first sample conversion result 802 corresponding to the artifact image 801 is obtained.
  • the first prediction conversion result 812 corresponding to the first prediction region 811 wherein the window range corresponding to the first sample conversion result 802 and the first prediction conversion result 812 is consistent with the preset window range of the sample removal model 820
  • the first prediction area 811 and the first prediction conversion result 812 are input to the channel fusion layer (Channel Concatenation, marked C in Figure 8), the first fusion result is input to the sample removal model 820, and the second prediction corresponding to the sample removal model 820 is output.
  • the region 821 inputs the artifact image 801, the first prediction region 811 and the second prediction region 821 into the window conversion layer for window conversion, and obtains the second sample conversion result 803 corresponding to the artifact image 801 and the second sample conversion result 803 corresponding to the first prediction region 811.
  • the prediction conversion result 813, and the third prediction conversion result 822 corresponding to the second prediction area 821 input the second sample conversion result 803, the second prediction conversion result 813 and the third prediction conversion result 822 into the channel fusion layer and then input the sample for removal.
  • the model 830 outputs the third prediction region 831 corresponding to the sample removal model 830.
  • determine the window range corresponding to the i-th sample removal model i is a positive integer; perform window conversion on the artifact image and the i-1th artifact removal result to obtain the artifact image and the i-1th artifact removal result.
  • the window conversion results corresponding to the shadow removal results are used as the model input of the i-th sample removal model.
  • X ori is used to represent the original image
  • X curr and Window range (L represents window level, H represents window height).
  • Step 320 Input the artifact images into multiple sample removal models, and output artifact removal results corresponding to the artifact images respectively.
  • sample removal models correspond to different preset window ranges, and the sample removal models are used to remove artifacts in the artifact image based on the corresponding preset window range.
  • three artifact removal results are obtained by outputting respectively three sample removal models with different preset window ranges.
  • preset loss function Calculate the pixel distances between the three artifact removal results and the reference image respectively, and output three prediction loss values, respectively and
  • Step 330 Based on the pixel difference between the artifact removal result and the reference image, determine the prediction loss values corresponding to the multiple sample removal models.
  • three prediction loss values are input into the sample weight model respectively, and three weight parameters are output.
  • the implementation of the sample weight model is:
  • Step 340 Input the prediction loss values corresponding to the multiple sample removal models into the sample weight model, and output the weight parameters corresponding to the multiple prediction loss values.
  • the weight parameter is used to adjust the weight of the parameter update of the sample removal model.
  • the training process of multiple sample removal models and sample weight models is carried out iteratively and alternately, that is, the first model parameters of the multiple sample removal models and the sample
  • the second model parameters of the weight model are updated alternately, and the first model parameters are adjusted through the updated second model parameters.
  • Figure 9 shows a schematic diagram of the training method of the artifact removal model provided by an exemplary embodiment of the present application.
  • the artifact image 910 is input to the sample removal model 920 (including three (the number of models is not shown in the figure), the output artifact removal results 931, artifact removal results 932 and artifact removal results 933 are obtained, in which the sample removal model 920 and the sample weight model 940 are trained alternately.
  • the first model parameter ( ⁇ (s-1) ) and the second model parameter ( ⁇ (s-1) ) are currently obtained after the s-1th iterative training.
  • the mapping function about ⁇ obtained after the s-1th iteration training Determine the verification loss value obtained after the s-1th iteration of training, and perform gradient adjustment on ⁇ (s-1) based on the verification loss value obtained after the s-1th iteration of training, to obtain the corresponding value after the s-th iteration of training.
  • Two model parameters ( ⁇ (s) ) determine the sample weight model corresponding to the s-th iterative training based on ⁇ (s) , and input the predicted loss value obtained after the s-1th iterative training into the sample corresponding to the s-th iterative training.
  • Weight model obtain the corresponding weight parameters after the s-th iterative training, and determine the s-th iterative training based on the predicted loss value obtained after the s-1th iterative training and the weight parameters corresponding to the s-th iterative training through Formula 7
  • the corresponding weighted loss value is used to perform gradient adjustment on ⁇ (s-1) to obtain the second model parameter ( ⁇ (s) ) corresponding to the s-th iterative training. This is used to perform loop iterative training until the sample removal model training is completed. .
  • both the first model parameter and the second model parameter are preset initial values.
  • a first learning decay rate is obtained.
  • the first learning decay rate is used to adjust the first learning rate in a decaying form according to the number of iterations.
  • the first learning rate is preset for multiple sample removal models.
  • the update step size of training in the process of training multiple sample removal models, perform gradient descent on the first learning rate based on the first learning attenuation rate to obtain the target learning rate corresponding to the artifact removal model.
  • the first learning rate is set to 2 ⁇ 10 -4 and the first learning decay rate is set every 30 epochs.
  • the first learning rate decays by 0.5.
  • the total number of training epochs is 200.
  • the batch size is 16, and the sample CT image block size is 64x64.
  • the second learning rate is set to 1 ⁇ 10 -5 for the sample weight model, and the number of neurons corresponding to the hidden layer is 100.
  • Figure 10 shows a schematic diagram of the artifact removal model processing process provided by an exemplary embodiment of the present application.
  • the current processing system includes a front-end A1010 (such as a CT scanner), Server 1020 and front-end B1030 (such as computer terminal or mobile phone terminal).
  • the front-end A1010 When the front-end A1010 performs a CT scan on the target detection object, if metal is implanted in the target detection object, the CT image generated by the current front-end A1010 will contain a metal artifact area caused by the metal.
  • the CT image is input into the artifact removal model in the server 1020 to perform artifact removal on the CT image.
  • the identified metal artifact areas are removed to generate CT images without metal artifacts in different preset window ranges. And feed it back to the front-end B1030 for doctors to make auxiliary diagnosis.
  • the training method of the artifact removal model uses reference images and artifact images with matching image content to train multiple sample removal models, wherein the artifact images are input during the training process.
  • multiple artifact removal results are output respectively
  • the prediction loss value between the multiple artifact removal results and the reference image is determined
  • the prediction loss value is input into the sample weight model to finally obtain the weight corresponding to each prediction loss value.
  • Parameters multiple sample removal models are trained according to the predicted loss value and weight parameter, and finally an artifact removal model containing multiple artifact removal sub-models is obtained.
  • the weight parameter and the predicted loss value multiple corresponding different preset windows are obtained.
  • the sample removal model is trained with a range of samples, so that the finally trained artifact removal model can output artifact removal images corresponding to different window ranges, meet the artifact removal requirements of different images, and improve the artifact removal accuracy of the artifact removal results. .
  • the method of training multiple sample removal models based on multiple window ranges enables multiple sample removal models with different window ranges to be trained simultaneously, and finally obtains a model that can output artifact removal results in different window ranges.
  • the artifact removal model can better improve the training effect of the model.
  • converting images in different window ranges through window conversion can enable better model learning between different window ranges and improve the accuracy and flexibility of model training.
  • Figure 11 shows a schematic diagram of the training method of the artifact removal model provided by an exemplary embodiment of the present application. As shown in Figure 11, the method includes the following:
  • Step 1110 start.
  • the training method of the artifact removal model provided by this application is executed by the server.
  • the server starts to execute the training process of the artifact removal model.
  • the server first determines whether the training phase is currently in the training phase or the testing phase. If it is in the training phase, step 1120 is executed. If it is in the testing phase, step 1150 is executed.
  • Step 1120 Obtain the artifact image.
  • the artifact image refers to a CT image containing metal artifacts.
  • the artifact image is a composite image containing metal artifacts obtained by obtaining a reference image in a public data set, combined with different metal mask information, and synthesized through a data simulation process. CT images of shadows as training data.
  • the reference image and the artifact image correspond to the same image content, that is, the reference image and the artifact image are implemented as a sample image pair.
  • Step 1130 Perform iterative loop training on the first model parameters of the sample removal model and the second model parameters of the sample weight model.
  • simultaneous gradient adjustment is performed on the first model parameters corresponding to the three sample removal models.
  • the artifact images are input into three sample removal models respectively, and three artifact removal results are obtained as output.
  • the preset loss function is used Determine the prediction loss values corresponding to the three sample removal models.
  • the output is the weight parameters W 1 , W 2 and W 3 corresponding to the three prediction loss values respectively.
  • the above-mentioned formula 4, formula 3 and formula 7 are followed in sequence to back propagate to the sample weight model and the three sample removal models, and successively target the second model of the sample weight model
  • the parameters and the first model parameters corresponding to the three sample removal models are gradient adjusted.
  • step 1140 is executed, otherwise step 1130 is continued.
  • Step 1140 save the training model.
  • the artifact removal model is determined using the plurality of first model parameters obtained in the last training as the first parameters, where the artifact removal model includes three training Good artifact removal submodel.
  • the server stores the trained artifact removal model.
  • Step 1150 Obtain a test artifact image.
  • test artifact image containing metal artifacts is obtained, where the test artifact image is implemented as a CT image used to test the effect of the trained artifact removal model.
  • Step 1160 Load the trained artifact removal model.
  • the server After the server obtains the test artifact image, it loads the stored artifact removal model that has been trained.
  • Step 1170 Obtain the artifact removal result generated by the artifact removal model through forward calculation.
  • the artifact removal result refers to generating a CT image within a preset window range after removing artifacts from the test artifact image.
  • Step 1180 Output CT images in different preset window ranges corresponding to the artifact removal results.
  • the three finally trained artifact removal sub-models are used as the artifact removal model. That is, after the current target image is input into the artifact removal model, three corresponding to different preset window ranges will be output at the same time. Artifact removal results.
  • Figure 12 shows a schematic diagram of the application process of the artifact removal model provided by an exemplary embodiment of the present application. As shown in Figure 12, when the target image 1210 (in this embodiment, an image for the same Abdominal CT images of three different display modes (image 1211, image 1212 and image 1213) of abdominal tissue in the same window range are input into the artifact removal model 1220, and artifacts in three different window ranges are simultaneously output.
  • Removal results (including three artifact removal results 111 corresponding to image 1211, three artifact removal results 122 corresponding to image 1212, and three artifact removal results 133 corresponding to image 1213).
  • the artifact removal results are used to assist doctors in diagnosing abdominal tissues.
  • the training method of the artifact removal model uses reference images and artifact images with matching image content to train multiple sample removal models, wherein the artifact images are input during the training process.
  • multiple artifact removal results are output respectively
  • the prediction loss value between the multiple artifact removal results and the reference image is determined
  • the prediction loss value is input into the sample weight model to finally obtain the weight corresponding to each prediction loss value.
  • Parameters multiple sample removal models are trained according to the predicted loss value and weight parameter, and finally an artifact removal model containing multiple artifact removal sub-models is obtained.
  • the weight parameter and the predicted loss value multiple corresponding different preset windows are obtained.
  • the sample removal model is trained with a range of samples, so that the finally trained artifact removal model can output artifact removal images corresponding to different window ranges, meet the artifact removal requirements of different images, and improve the artifact removal accuracy of the artifact removal results. .
  • the designed sample removal model (DICD-Net) has good interpretability, which allows users to have a good understanding of the functions of each module in the model;
  • the sample weight model is introduced between different window ranges, making the reconstruction and restoration learning process of different window ranges more flexible, which has better potential to fully improve the fidelity of the organizational structure;
  • Figure 13 is a structural block diagram of a training device for an artifact removal model provided by an exemplary embodiment of the present application. As shown in Figure 13, the device includes the following parts:
  • the acquisition module 1310 is used to acquire a reference image and an artifact image that match the image content.
  • the reference image is an image generated after scanning a sample detection object that does not contain an implant.
  • the artifact image is a reference image that contains artifacts. image, the artifact is a shadow produced by the implant during the scanning process of a sample detection object containing the implant; that is, the artifact is produced by the implant during the scanning process shadow;
  • the input module 1320 is used to input the artifact image into multiple sample removal models, and obtain artifact removal results corresponding to the artifact images output by the multiple sample removal models respectively.
  • Different sample removal models correspond to different predictions. Assuming a window range, the sample removal model is used to remove artifacts in the artifact image based on the corresponding preset window range;
  • Determining module 1330 configured to determine prediction loss values corresponding to multiple sample removal models based on the pixel difference between the artifact removal result and the reference image;
  • the input module 1320 is also used to input the predicted loss values corresponding to multiple sample removal models into the sample weight model, and output the weight parameters corresponding to the multiple predicted loss values.
  • the weight parameters are used to remove the sample.
  • the parameters of the model are updated and the weights are adjusted;
  • the training module 1340 is used to train multiple sample removal models based on the predicted loss value and the weight parameter to obtain an artifact removal model composed of multiple artifact removal sub-models.
  • the artifact removal sub-model is Artifact removal is performed on the target image based on the corresponding preset window range.
  • the training module 1340 includes:
  • the determination unit 1341 is configured to determine the weighted loss values corresponding to the multiple sample removal models based on the multiple predicted loss values and the weight parameters corresponding to the multiple loss values respectively;
  • the adjustment unit 1342 is configured to respectively adjust the first model parameters of the plurality of sample removal models based on the weighted loss values corresponding to the plurality of sample removal models to obtain multiple artifact removal sub-models;
  • the determining unit 1341 is also configured to use the plurality of artifact removal sub-models as the artifact removal model.
  • the determination unit 1341 is also configured to, in the s-th iterative training process, based on the predicted loss value obtained from the s-1th iterative training and the weight parameter obtained from the s-th iterative training. , determine the weighted loss value corresponding to the sth iteration training;
  • the adjustment unit 1342 is also configured to perform gradient adjustment on the first model parameters obtained by the s-1th iterative training of the sample removal model based on the weighted loss value corresponding to the sth iterative training, to obtain the sth
  • the corresponding first model parameters are trained iteratively for s+1 times, and the s+1th cyclic adjustment is performed until the artifact removal model training ends, s ⁇ 1 and s is an integer.
  • the input module 1320 is also used to input the prediction loss obtained by the s-1th iteration training into the sample weight model obtained by the sth iteration training, and output the corresponding training result of the sth iteration. weight parameter.
  • the acquisition module 1310 is also used to acquire a verification reference image and a verification artifact image whose image content matches;
  • the input module 1320 is also used to input the verification artifact image into multiple sample removal models, and output the verification removal results corresponding to the verification artifact image respectively;
  • the determination module 1330 is also configured to determine verification loss values corresponding to multiple sample removal models based on the pixel difference between the verification removal result and the verification reference image;
  • the training module 1340 is also used to train the sample weight model based on the verification loss value.
  • the training module 1340 is also used to, during the s iterative training process, calculate the second model of the sample weight model based on the verification loss value obtained from the s-1th iterative training.
  • the parameters are gradient adjusted to obtain the sample weight model corresponding to the sth iteration training.
  • the determination module 1330 is further configured to determine, based on the first model parameters obtained in the s-1 th iterative training, the difference between the first model parameter and the second model parameter in the s-1 th iteration.
  • the determination module 1330 is further configured to determine the most recently adjusted first model parameter as the first model parameter in response to the number of cyclic iterative adjustments of the first model parameter reaching a threshold. a parameter; or; in response to the adjustment effect of the first model parameter meeting the adjustment effect condition, determining the first model parameter as the first parameter, and the adjustment effect condition is used to represent the prediction loss value restriction requirements.
  • the obtaining module 1310 is also used to obtain a first learning decay rate, which is used to adjust the first learning rate in a decaying form according to the number of iterations.
  • the first learning rate is a preset update step size for training multiple sample removal models;
  • the device also includes:
  • the gradient descent module 1350 is configured to perform gradient descent on the first learning rate based on the first learning attenuation rate in the process of training multiple sample removal models to obtain the target learning corresponding to the artifact removal model. Rate.
  • the determination module 1330 is also used to determine the window range corresponding to the i-th sample removal model, where i is a positive integer;
  • the device also includes:
  • the conversion module 1360 is used to perform window conversion on the artifact image and the i-1th artifact removal result, and obtain the window conversion results corresponding to the artifact image and the i-1th artifact removal result respectively, as the i-1th artifact removal result.
  • the training device for artifact removal models uses reference images and artifact images with matching image content to train multiple sample removal models, where the artifact images are input during the training process.
  • multiple artifact removal results are output respectively
  • the prediction loss value between the multiple artifact removal results and the reference image is determined
  • the prediction loss value is input into the sample weight model to finally obtain the weight corresponding to each prediction loss value.
  • Parameters multiple sample removal models are trained according to the predicted loss value and weight parameter, and finally an artifact removal model containing multiple artifact removal sub-models is obtained.
  • the weight parameter and the predicted loss value multiple corresponding different preset windows are obtained.
  • the sample removal model is trained with a range of samples, so that the finally trained artifact removal model can output artifact removal images corresponding to different window ranges, meet the artifact removal requirements of different images, and improve the artifact removal accuracy of the artifact removal results. .
  • the training device for the artifact removal model provided by the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above function allocation can be completed by different functional modules as needed, that is, The internal structure of the device is divided into different functional modules to complete all or part of the functions described above.
  • the training device of the artifact removal model provided in the above embodiments and the training method embodiment of the artifact removal model belong to the same concept. The specific implementation process can be found in the method embodiments and will not be described again here.
  • Figure 15 shows a schematic structural diagram of a server provided by an exemplary embodiment of the present application. Specifically:
  • the server 1500 includes a central processing unit (Central Processing Unit, CPU) 1501, a system memory 1504 including a random access memory (Random Access Memory, RAM) 1502 and a read only memory (Read Only Memory, ROM) 1503, and a connection system memory 1504 and the system bus 1505 of the central processing unit 1501.
  • Server 1500 also includes a mass storage device 1506 for storing operating system 1513, applications 1514, and other program modules 1515.
  • Mass storage device 1506 is connected to central processing unit 1501 through a mass storage controller (not shown) connected to system bus 1505 .
  • Mass storage device 1506 and its associated computer-readable media provide non-volatile storage for server 1500 . That is, mass storage device 1506 may include computer-readable media (not shown) such as a hard disk or a Compact Disc Read Only Memory (CD-ROM) drive.
  • CD-ROM Compact Disc Read Only Memory
  • Computer-readable media may include computer storage media and communication media.
  • Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.
  • Computer storage media include RAM, ROM, Erasable Programmable Read Only Memory (EPROM), electrically erasable and programmable memory Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other solid-state storage technology, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, tape cassette, magnetic tape, disk storage or other Magnetic storage devices.
  • RAM random access memory
  • ROM Erasable Programmable Read Only Memory
  • EPROM Erasable Programmable Read Only Memory
  • EEPROM Electrically Erasable Programmable Read Only Memory
  • flash memory or other solid-state storage technology
  • CD-ROM Compact Disc
  • DVD Digital Versatile Disc
  • the server 1500 may also run on a remote computer connected to a network through a network such as the Internet. That is, the server 1500 can be connected to the network 1512 through the network interface unit 1511 connected to the system bus 1505, or the network interface unit 1511 can also be used to connect to other types of networks or remote computer systems (not shown).
  • the above-mentioned memory also includes one or more programs.
  • One or more programs are stored in the memory and configured to be executed by the CPU.
  • Embodiments of the present application also provide a computer device.
  • the computer device includes a processor and a memory.
  • the memory stores at least one instruction, at least a program, a code set or an instruction set. At least one instruction, at least a program, code.
  • the set or instruction set is loaded and executed by the processor to implement the training method of the artifact removal model provided by each of the above method embodiments.
  • Embodiments of the present application also provide a computer-readable storage medium, which stores at least one instruction, at least a program, a code set or an instruction set, at least one instruction, at least a program, a code set or a set of instructions.
  • the instruction set is loaded and executed by the processor to implement the training method of the artifact removal model provided by each of the above method embodiments.
  • Embodiments of the present application also provide a computer program product or computer program.
  • the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium.
  • the processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the training method of the artifact removal model described in any of the above embodiments.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Image Analysis (AREA)
  • Electrically Operated Instructional Devices (AREA)

Abstract

一种伪影去除模型的训练方法、装置、设备、介质及程序产品，涉及机器学习领域。该方法包括：获取图像内容匹配的参考图像和伪影图像(310)；将伪影图像输入多个样本去除模型，分别输出得到伪影图像对应的伪影去除结果(320)；确定多个样本去除模型分别对应的预测损失值(330)；将多个预测损失值输入样本权重模型，输出得到多个预测损失值分别对应的权重参数(340)；基于预测损失值和权重参数，对多个样本去除模型进行训练，得到多个伪影去除子模型构成的伪影去除模型伪影去除模型(350)。上述方案能够使得在对多个样本去除模型进行训练的过程中，根据权重参数进行灵活训练，提高伪影去除模型训练的准确度。

Description

伪影去除模型的训练方法、装置、设备、介质及程序产品
本申请要求于2022年08月09日提交的、申请号为202210951294.0、发明名称为“伪影去除模型的训练方法、装置、设备、介质及程序产品”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及机器学习领域，特别涉及一种伪影去除模型的训练方法、装置、设备、介质及程序产品。
背景技术
在计算机断层扫描(Computed Tomography，CT)扫描过程中，由于受不同因素影响，CT图像中会产生伪影，如：在对口腔进行CT扫描时，由于牙齿中植入了假牙，则生成的CT图像将受假牙影响从而产生条状阴影。
在相关技术中，采用双域网络(DuDoNet)去除CT图像中的伪影，其中，双域网络中由两个模块构成，通过预设一个CT值处理窗口分别对包含伪影的弦图进行弦图域的处理，以及对包含伪影的CT图像进行图像域的处理，输出得到修复的正弦图像和增强后的CT图像，最终利用反投影层输出得到去除伪影的CT图像。
然而在相关技术中，通过双域网络去除包含伪影的CT图像的方法，去除伪影后的CT图像的还原真实度较低，图像准确度较低，图像质量较差。
发明内容
本申请实施例提供了一种伪影去除模型的训练方法、装置、设备、介质及程序产品，能够提高伪影去除模型的输出结果的准确度。所述技术方案如下：
一方面，提供了一种伪影去除模型的训练方法，所述方法由计算机设备执行，所述方法包括：
获取图像内容匹配的参考图像和伪影图像，所述参考图像是对未包含植入物的样本检测对象扫描后生成的图像，所述伪影图像是包含伪影的参考图像，所述伪影是所述植入物在扫描过程中产生的阴影；
将所述伪影图像输入多个样本去除模型，获得所述多个样本去除模型分别输出的所述伪影图像对应的伪影去除结果，不同样本去除模型对应不同的预设窗口范围，所述样本去除模型用于以对应的预设窗口范围为基准，去除所述伪影图像中的伪影；
基于所述伪影去除结果和所述参考图像之间的像素点差异，确定多个样本去除模型分别对应的预测损失值；
将多个样本去除模型分别对应的预测损失值输入样本权重模型，输出得到多个预测损失值分别对应的权重参数，所述权重参数用于对所述样本去除模型的参数更新进行权重调整；
基于所述预测损失值和所述权重参数，对多个样本去除模型进行训练，得到多个伪影去除子模型构成的伪影去除模型，所述伪影去除子模型用于基于对应的预设窗口范围对目标图像进行伪影去除。
另一方面，提供了一种伪影去除模型的训练装置，所述装置包括：
获取模块，用于获取图像内容匹配的参考图像和伪影图像，所述参考图像是对未包含植入物的样本检测对象扫描后生成的图像，所述伪影图像是包含伪影的参考图像，所述伪 影是所述植入物在扫描过程中产生的阴影；
输入模块，用于将所述伪影图像输入多个样本去除模型，获得所述多个样本去除模型分别输出的所述伪影图像对应的伪影去除结果，不同样本去除模型对应不同的预设窗口范围，所述样本去除模型用于以对应的预设窗口范围为基准，去除所述伪影图像中的伪影；
确定模块，用于基于所述伪影去除结果和所述参考图像之间的像素点差异，确定多个样本去除模型分别对应的预测损失值；
所述输入模块，用于将多个样本去除模型分别对应的预测损失值输入样本权重模型，输出得到多个预测损失值分别对应的权重参数，所述权重参数用于对所述样本去除模型的参数更新进行权重调整；
训练模块，用于基于所述预测损失值和所述权重参数，对多个样本去除模型进行训练，得到多个伪影去除子模型构成的伪影去除模型，所述伪影去除子模型用于基于对应的预设窗口范围对目标图像进行伪影去除。
另一方面，提供了一种计算机设备，所述计算机设备包括处理器和存储器，所述存储器中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由所述处理器加载并执行以实现如上述本申请实施例中任一所述伪影去除模型的训练方法。
另一方面，提供了一种计算机可读存储介质，所述存储介质中存储有至少一条指令、至少一段程序、代码集或指令集，所述至少一条指令、所述至少一段程序、所述代码集或指令集由处理器加载并执行以实现如上述本申请实施例中任一所述的伪影去除模型的训练方法。
另一方面，提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述实施例中任一所述的伪影去除模型的训练方法。
本申请实施例提供的技术方案带来的有益效果至少包括：
采用图像内容匹配的参考图像和伪影图像对多个样本去除模型进行训练，其中，在训练过程中将伪影图像输入多个样本去除模型后分别输出多个伪影去除结果，确定多个伪影去除结果和参考图像之间的预测损失值，将预测损失值输入样本权重模型后最终得到各个预测损失值对应的权重参数，根据预测损失值和权重参数对多个样本去除模型进行训练，最终得到包含多个伪影去除子模型的伪影去除模型，也即，通过采用权重参数和预测损失值对多个对应不同预设窗口范围的样本去除模型进行训练，使得最终训练得到的伪影去除模型能够输出不同窗口范围对应的伪影去除图像，满足不同图像的伪影去除需求，提高了伪影去除结果的伪影去除准确度。
附图说明
图1是本申请一个示例性实施例提供的伪影去除模型的训练方法示意图；
图2是本申请一个示例性实施例提供的实施环境示意图；
图3是本申请一个示例性实施例提供的伪影去除模型的训练方法流程图；
图4是本申请另一个示例性实施例提供的伪影去除模型的训练方法流程图；
图5是本申请另一个示例性实施例提供的DICD-Net模型示意图；
图6是本申请一个示例性实施例提供的网络结构示意图；
图7是本申请一个示例性实施例提供的样本权重模型网络结构示意图；
图8是本申请一个示例性实施例提供的多个样本去除模型示意图；
图9是本申请另一个示例性实施例提供的伪影去除模型的训练方法示意图；
图10是本申请一个示例性实施例提供的伪影去除模型应用过程示意图；
图11是本申请一个示例性实施例提供的伪影去除模型的训练方法示意图；
图12是本申请一个示意性实施例提供的伪影去除模型处理过程示意图；
图13是本申请一个示意性实施例提供的伪影去除模型的训练装置结构框图；
图14是本申请另一个示意性实施例提供的伪影去除模型的训练装置结构框图；
图15是本申请一个示例性实施例提供的服务器的结构示意图。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚，下面将结合附图对本申请实施方式作进一步地详细描述。
首先，针对本申请实施例中涉及的名词进行简单介绍。
窗技术：窗技术是计算机断层扫描(Computed Tomography，CT)检查中用以观察不同密度的正常组织或者病变的一种显示技术，包括窗宽(Window Width)和窗位(Window Level)，由于各种组织结构或者病变具有不同的CT值，因此想要在CT图像上显示指定组织结构的细节时，需要选择适合观察该指定组织结构的窗宽和窗位，构成指定窗口范围，以获得针对该指定组织结构的最优显示模式，生成该指定窗口范围对应CT值的灰度图像。
示意性的，请参考图1，其示出了本申请一个示例性实施例提供的伪影去除模型的训练方法示意图，如图1所示，获取训练图像集100，其中，训练图像集100中包括图像内容匹配的参考图像101和伪影图像102，其中，参考图像101和伪影图像102属于一个样本图像对，参考图像101和伪影图像102都是对腹部进行计算机断层扫描(Computed Tomography，CT)后得到的CT图像，而伪影图像102是包含伪影的图像(受伪影污染的腹部CT图像)，参考图像101中不包含伪影(未受伪影污染的腹部CT图像)。
将伪影图像102输入多个样本去除模型110，分别输出得到伪影图像102的伪影去除结果111，其中，多个样本去除模型110中各样本去除模型对应不同的预设窗口范围，因此伪影去除结果111实现为对应不同预设窗口范围下的伪影去除图像(如：图像1111是在[-1000，2000]HU窗口范围下的CT图像，图像1112是在[-320，480]HU窗口范围下的CT图像，图像1113是在[-160，240]HU窗口范围下的CT图像)。根据伪影去除结果111和参考图像101之间的像素点差异确定多个样本去除模型110分别对应的预测损失值112。
将预测损失值112输入样本权重模型120，输出得到预测损失值112分别对应的权重参数121，权重参数121用于对样本去除模型110的参数更新进行权重调整，根据预测损失值112和权重参数121对多个样本去除模型110进行训练，最终得到多个伪影去除子模型构成的伪影去除模型130，其中，伪影去除模型130用于对输入的包含伪影的目标图像进行伪影去除。
其次，对本申请实施例中涉及的实施环境进行说明，示意性的，请参考图2，该实施环境中涉及终端210、服务器220，终端210和服务器220之间通过通信网络230连接。
在一些实施例中，终端210向服务器220发送伪影去除请求，其中，伪影去除请求中包括目标扫描图像，本实施例中，目标扫描图像实现为被金属污染的CT图像(也即，对人体指定部位进行CT扫描过程中，生成的CT图像受到指定部位中植入的金属影响，产生的金属伪影)，服务器220接收到来自终端发送的伪影去除请求后，对目标扫描图像中包含的金属伪影进行伪影去除，生成伪影去除结果，将伪影去除结果反馈至终端210。
其中，服务器220中包括伪影去除模型221，服务器220将目标扫描图像输入伪影去除模型221，输出得到伪影去除结果，伪影去除结果是指将目标扫描图像中识别得到的伪影区域进行去除后生成的CT增强图像。
其中，伪影去除模型221是通过将用于训练的伪影图像222输入多个样本去除模型223， 输出得到多个伪影去除结果，根据伪影去除结果和参考图像(与伪影图像222图像内容匹配且不存在伪影的图像)之间像素点差异确定多个预测损失值224，将预测损失值224输入样本权重模型225，输出得到与多个预测损失值224分别对应的权重参数226，根据权重参数226和预测损失值224对样本去除模型223进行训练后得到的。
上述终端210可以是手机、平板电脑、台式电脑、便携式笔记本电脑、智能电视、智能车载等多种形式的终端设备，本申请实施例对此不加以限定。
值得注意的是，上述服务器220可以是独立的物理服务器，也可以是多个物理服务器构成的服务器集群或者分布式系统，还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、内容分发网络(Content Delivery Network，CDN)、以及大数据和人工智能平台等基础云计算服务的云服务器。
其中，云技术(Cloud technology)是指在广域网或局域网内将硬件、软件、网络等系列资源统一起来，实现数据的计算、储存、处理和共享的一种托管技术。
在一些实施例中，上述服务器220还可以实现为区块链系统中的节点。
需要说明的是，本申请所涉及的信息(包括但不限于用户设备信息、用户个人信息等)、数据(包括但不限于用于分析的数据、存储的数据、展示的数据等)以及信号，均为经用户授权或者经过各方充分授权的，且相关数据的收集、使用和处理需要遵守相关国家和地区的相关法律法规和标准。例如，本申请中涉及到的用于训练的参考图像和伪影图像，以及用于模型验证的验证图像是在充分授权的情况下获取的。
示意性的，对本申请提供的伪影去除模型的训练方法进行说明，请参考图3，其示出了本申请一个示例性实施例提供的伪影去除模型的训练方法流程图，该方法由计算机设备执行，比如，该方法可以由终端执行，也可以由服务器执行，或者，也可以由终端和服务器共同执行，本实施例中以该方法由服务器执行进行说明，如图3所示，该方法包括：
步骤310，获取图像内容匹配的参考图像和伪影图像。
其中，参考图像是对未包含植入物的样本检测对象扫描后生成的图像，伪影图像是包含伪影的参考图像，伪影是在对包含植入物的样本检测对象的扫描过程中植入物产生的阴影。也即是说，上述的伪影是植入物在扫描过程中产生的阴影。
示意性的，参考图像是指通过指定扫描技术对样本检测对象进行扫描后生成的医学图像，一般性的，伪影图像实现为灰度图像。其中，样本检测对象用于表示某个指定组织器官(如：心脏、腹部、胸腔、肺部等)。
在一些实施例中，参考图像是针对未包含植入物的样本检测对象进行扫描后得到的医学图像，也即，参考图像是未受到植入物影响的医学图像。
在一些实施例中，植入物指包含金属部件的且植入检测对象中的物体，如：假牙、起搏器、支架等植入物类型中至少一种，对此不加以限定。
示意性的，指定扫描技术是指CT扫描技术，因此，本申请实施例中涉及的图像均为CT图像。
在一些实施例中，图像内容匹配是指参考图像中包含的内容和伪影图像中包含的内容相同，如：参考图像和伪影图像都是针对同一个腹部进行CT扫描后生成的CT图像。其中，伪影图像与参考图像之间的区别为伪影图像是包含伪影的参考图像。也即，参考图像和伪影图像实现为一个样本图像对。
示意性的，伪影表示检测对象在进行扫描过程中，除样本检测对象以外的对象在扫描过程中在图像上产生的阴影(或者暗带)。
步骤320，将伪影图像输入多个样本去除模型，分别输出得到伪影图像对应的伪影去除结果。
其中，计算机设备(比如服务器)将伪影图像输入多个样本去除模型，获得多个样本去 除模型分别输出的伪影图像对应的伪影去除结果。
其中，不同样本去除模型对应不同的预设窗口范围，样本去除模型用于以对应的预设窗口范围为基准，去除伪影图像中的伪影。
在一些实施例中，伪影去除结果是指通过样本去除模型，将伪影图像中包含的伪影进行去除，并输出与样本去除模型对应预设窗口范围对应的扫描图像，也即，该扫描图像中不包含伪影。
其中，上述扫描图像呈现的各区域对比度关系与伪影图像呈现的各区域的对比度关系相同，或者，不相同，对此不加以限定。
示意性的，预设窗口范围用于表示扫描图像中各区域之间的对比度关系，如：扫描图像中包括区域a和区域b，对应预设窗口范围A下，扫描图像中区域a亮度高于区域b，对应预设窗口范围B下，扫描图像中区域a亮度低于区域b。也即，同一张扫描图像对应不同的预设窗口范围时，其显示的各区域之间的对比度不同，便于对指定区域进行针对性查看。
可选地，样本去除模型对应的预设窗口范围是预先设置的固定窗口范围，如：样本去除模型A对应的预设窗口范围为[-1000，2000]HU；或者，样本去除模型对应的预设窗口范围是根据实际需要进行设定的可调整窗口范围，对此不加以限定。
可选地，多个样本去除模型对应相同的模型结构；或者，多个样本去除模型对应不同的模型结构，对此不加以限定。
步骤330，基于伪影去除结果和参考图像之间的像素点差异，确定多个样本去除模型分别对应的预测损失值。
示意性的，预测损失值用于表示伪影去除结果和参考图像之间各像素点之间的差值。
在一些实施例中，预设损失函数，通过损失函数计算伪影去除结果对应的像素值和参考图像对应的像素值之间的距离，将计算得到的结果作为多个样本去除模型分别对应的预测损失值。
步骤340，将多个样本去除模型分别对应的预测损失值输入样本权重模型，输出得到多个预测损失值分别对应的权重参数。
其中，权重参数用于对样本去除模型的参数更新进行权重调整。
在一些实施例中，将多个样本去除模型对应的预测损失值分别输入样本权重模型，输出得到标量结果，作为单个预测损失值对应的权重参数。
可选地，不同样本去除模型分别对应的预测损失值输入样本权重模型后，对应输出的多个预测损失值分别的权重参数各不相同；或者，存在至少两个预测损失值分别对应的权重参数相同，对此不加以限定。
示意性的，权重参数用于使得通过预测损失值对样本去除模型进行训练过程中，给各预测损失值分配不同的权重。
可选地，在获取将多个样本去除模型分别对应的预测损失值后，将多个预测损失值再同时输入样本权重模型，同时输出得到多个预测损失值分别对应的权重参数，也即，多个预测损失值分别对应的权重参数是同时获取得到的；或者，每获取一个样本去除模型对应的预测损失值后，即将其输入样本权重模型，输出得到该样本去除模型对应的权重参数，也即，多个预测损失值分别对应的权重参数是依次获取得到的，对此不加以限定。
步骤350，基于预测损失值和权重参数，对样本去除模型进行训练，得到多个伪影去除子模型构成的伪影去除模型。
其中，伪影去除子模型用于基于对应的预设窗口范围对目标图像进行伪影去除。
示意性的，根据预测损失值和权重参数，对样本去除模型的第一模型参数进行参数调整，根据调整后的参数确定伪影去除子模型。
在一些实施例中，单个样本去除模型训练完后得到单个伪影去除子模型，最终多个伪影去除子模型构成伪影去除模型。
可选地，根据预测损失值和权重参数对样本去除模型进行训练的过程中，每个样本去除模型的训练过程是同时进行的，或者，每个样本去除模型的训练过程是依次执行，也即，当训练完第一个样本去除模型后，再开始训练第二个样本去除模型，对此不加以限定。
综上所述，本申请实施例提供的伪影去除模型的训练方法，采用图像内容匹配的参考图像和伪影图像对多个样本去除模型进行训练，其中，在训练过程中将伪影图像输入多个样本去除模型后分别输出多个伪影去除结果，确定多个伪影去除结果和参考图像之间的预测损失值，将预测损失值输入样本权重模型后最终得到各个预测损失值对应的权重参数，根据预测损失值和权重参数对多个样本去除模型进行训练，最终得到包含多个伪影去除子模型的伪影去除模型，通过采用权重参数和预测损失值对多个对应不同预设窗口范围的样本去除模型进行训练，使得最终训练得到的伪影去除模型能够输出不同窗口范围对应的伪影去除图像，满足不同图像的伪影去除需求，提高了伪影去除结果的伪影去除准确度。
在一个可选的实施例中，以针对单个样本去除模型进行训练为例，对样本去除模型的训练过程实现为多次循环迭代训练过程，示意性的，请参考图4，其示出了本申请一个示例性实施例提供的伪影去除模型的训练方法流程图，该方法由计算机设备执行，比如，该方法可以由终端执行，也可以由服务器执行，或者，也可以由终端和服务器共同执行，本实施例中以该方法由服务器执行进行说明，如图3所示，也即，步骤350中包括步骤351、步骤352和步骤353，步骤340中还包括步骤341，该方法包括如下步骤：
步骤310，获取图像内容匹配的参考图像和伪影图像。
其中，参考图像是对未包含植入物的样本检测对象扫描后生成的图像，伪影图像是包含伪影的参考图像，伪影是在对包含植入物的样本检测对象的扫描过程中植入物产生的阴影。
本实施例中，以金属伪影为例进行举例说明，由金属构成的伪影实现为条状结构的伪影。
示意性的，参考图像是对样本检测图像进行CT扫描后生成的CT图像，伪影图像是包含金属伪影的CT图像，也即，当前伪影图像由于样本检测图像中存在金属而导致CT扫描后生成的图像中包括金属伪影。
可选地，伪影图像和参考图像是从授权的公开数据集中直接获取的；或者，参考图像是从公开数据集中直接获取的图像，伪影图像是在参考图像的基础上，结合不同的金属对应的金属掩膜信息，人工合成得到的包含金属伪影的参考图像，对此不加以限定。
本实施例中，参考图像和伪影图像实现为一个样本图像对。
在一些实施例中，参考图像和伪影图像针对同一个样本检测对象对应的扫描图像。
步骤320，将伪影图像输入多个样本去除模型，分别输出得到伪影图像对应的伪影去除结果。
其中，不同样本去除模型对应不同的预设窗口范围，样本去除模型用于以对应的预设窗口范围为基准，去除伪影图像中的伪影。
示意性的，样本去除模型的预设窗口范围是预先设置好的固定窗口范围，如：样本去除模型A预设窗口范围固定为[-320，480]HU。
示意性的，不同的样本去除模型对应不同的预设窗口范围。
在一些实施例中，样本去除模型用于对伪影图像中的伪影进行去除后，根据预设窗口范围，将伪影图像调整至与预设窗口范围对应的显示模式，输出得到的结果即为伪影去除结果，如：伪影图像是窗口范围为[-1000,2000]HU的包含金属伪影的CT图像，将伪影图像输入样本去除模型(预设窗口范围为[-320,480]HU)之前，先将伪影图像的窗口范围调整至[-320,480]HU后将其输入样本去除模型，样本去除模型将伪影图像中的金属伪影去除进行图像输出，作为伪影去除结果。
示意性的，不同预设窗口范围对应的伪影去除结果显示的图像内容保持一致，各伪影去除结果显示的区域对比度不同。
可选地，本申请实施例中的样本去除模型可实现为深度可解释性卷积字典网络(Deep Interpretables Convolutional Dictionary Network，DICD-Net)、卷积神经网络(Convolutional Neural Network,CNN)、U-net网络等神经网络模型，对此不加以限定。
下面针对样本去除模型实现为DICD Net为例进行介绍。
针对由金属造成的伪影，具有金属伪影特有的先验知识，也即，金属伪影具有非局部的条状结构，该先验知识对于样本去除模型的参数学习能够起到引导作用，示意性的，请参考图5，其示出了本申请一个示例性实施例提供的DICD-Net模型示意图，如图5所示，DICD-Net500中包括N次迭代过程，在任意迭代过程中，该单次迭代过程由网络和X网络(X-Net)按照先后顺序组成。
如图5所示，将伪影图像510输入DICD-Net中进行N次迭代去除(N个Stage)，输出得到伪影图像对应的伪影去除结果520，其中，伪影去除结果520实现为通过N个网络和X网络对伪影图像进行去除后得到的CT图像。其中，网络和X网络分别完成特征层和伪影去除结果520的更新。
下面针对N次迭代去除过程中单次Stage进行说明。
示意性的，如图6所示，其示出了本申请一个示例性实施例提供的网络结构示意图，如图6所示，当前显示单次Stage中，网络和X网络对应的结构示意图600，其中，包括网络结构610和X网络结构620。
示意性的，网络610的网络结构可参考如下公式一：
公式一：
其中，表示第n次迭代去除过程中，网络的输出结果，表示残差网络630，残差网络630中的每个残差块依次包括：卷积层，批标准化层(Batch
Normalization)，ReLU层，卷积层，Batch Normalization层，以及跨链接层。
其中，Y表示输入的伪影图像(且该伪影图像中包含金属伪影)；X表示伪影去除结果(X(n-1)表示第n-1次迭代阶段得到的伪影去除结果)；I表示伪影图像中非金属伪影对应的掩膜信息(Mask)；表示金属伪影重复出现的模式，本实施例中可理解为金属伪影的显示方式，是特征层，代表金属伪影的条状伪影结构，本实施例中可理解为金属伪影对应的特征图(Feature Map)；η1为网络的更新步长。
示意性的，X网络620的网络结构可参考如下公式二：
公式二：
其中，X(n)表示第n次迭代去除过程中，X网络的输出结果，表示残差网络640，残差网络640中的每个残差块依次包括：卷积层，批标准化层(Batch Normalization)，ReLU层，卷积层，批标准化层，以及跨链接层。
其中，Y表示输入的伪影图像(且该伪影图像中包含金属伪影)；X表示伪影去除结果(X(n-1)表示第n-1次迭代阶段得到的伪影去除结果)；I表示伪影图像中非金属伪影对应的掩膜信息(Mask)，表示金属伪影重复出现的模式，本实施例中可理解为金属伪影的显示方式，是特征层，代表金属伪影的条状伪影结构，本实施例中可理解为金属伪影对应的特征图(Feature Map)，η2为X网络的更新步长。
结合上述公式一和公式二，如图6所示，当第n-1次迭代去除阶段得到的第n-1个特征图和第n-1次迭代去除阶段得到的伪影去除结果(X(n-1))输入网络中，输出得到第n次迭代去除阶段得到的第n个特征图将第n个特征图和第n-1次迭代阶段得到的伪影去除结果(X(n-1))输入X网络中，输出得到第n次迭代阶段的伪影去除结果(X(n))。
步骤330，基于伪影去除结果和参考图像之间的像素点差异，确定多个样本去除模型分别对应的预测损失值。
示意性的，预设损失函数，将伪影去除结果和参考图像通过预设损失函数计算像素值差值，将计算得到的结果作为多个样本去除模型分别对应的预测损失值。
本实施例中，预设的损失函数为其中，b表示第b个预设窗口范围对应的样本去除模型，Θ表示样本去除模型的第一模型参数。
步骤341，将第s-1次迭代训练得到的预测损失输入第s次迭代训练得到的样本权重模型，输出得到第s次迭代训练对应的权重参数。
本实施例中，在针对样本去除模型进行训练的过程中，样本权重模型也需要进行训练。
下面，首先针对样本权重模型的训练过程进行说明。
在一些实施例中，获取图像内容匹配的验证参考图像和验证伪影图像；将验证伪影图像输入多个样本去除模型，分别输出得到验证伪影图像对应的验证去除结果；基于验证去除结果和验证参考图像之间的像素点差异，确定多个样本去除模型分别对应的验证损失值；基于验证损失值，对样本权重模型进行训练。
示意性的，验证参考图像是对验证检测对象进行CT扫描后得到的CT图像，验证伪影图像是包含金属伪影的CT图像，其中，验证参考图像和验证伪影图像对应的图像内容相同(如：验证参考图像和验证伪影图像均为对同一个腹部进行CT扫描后得到的CT图像，其中，验证伪影图像中包括金属伪影，验证参考图像中不存在伪影)。
本实施例中，验证参考图像和验证伪影图像为验证样本中的一个图像对。
本实施例中，验证伪影图像和伪影图像属于不同的CT图像。
示意性的，将验证伪影图像输入多个样本去除模型，通过多个样本去除模型将伪影图像中的伪影去除后，生成与样本去除模型的预设窗口范围对应的图像，作为验证去除结果。其中，每个样本去除模型对应一个验证去除结果。
示意性的，预设损失函数，通过损失函数计算多个验证去除结果和验证参考图像之间像素点的差值，作为样本去除模型对应的验证损失值。其中，多个样本去除模型分别对应多个验证损失值。
本实施例中，预设的损失函数实现为该损失函数的输出结果用于表示第b个样本去除模型对应的验证损失值。
示意性的，通过多个验证损失值对样本权重模型的第二模型参数进行调整。
在一些实施例中，在s次迭代训练过程中，基于第s-1次迭代训练得到的验证损失值，对样本权重模型的第二模型参数进行梯度调整，得到第s次迭代训练对应的样本权重模型。
本实施例中，针对样本权重模型进行训练的过程中包括对样本权重模型进行N次迭代训练(与上述DICD-Net网络中包含的N次迭代过程对应)，也即，在N次迭代过程中，样本权重模型和样本去除模型是迭代交替更新的。
本申请实现的方案为通过多个样本去除模型对应的验证损失值，均对样本权重模型进行训练，以针对单个验证损失值对样本权重模型进行训练的过程为例进行说明，针对样本权重模型的第二模型参数进行梯度调整的方式可具体参考公式三：
公式三：
其中，θ(s)表示第s次迭代训练对应的样本权重模型，β为预设的第二学习率，用于表示对样本权重模型进行训练的更新步长，表示第b个样本去除模型对应的 验证损失值。
其中，用于表示在第s-1次迭代过程中关于θ的映射函数，由于在当前s次迭代训练过程中，第二模型参数(θ)尚未进行更新，因此设置一个关于θ的映射函数用于表示样本去除模型对应的第一模型参数(Θ)和样本权重模型对应的第二模型参数(θ)之间的映射关系。也即，基于s-1次迭代训练得到的第一模型参数，确定第一模型参数和第二模型参数之间在第s-1次迭代训练过程中对应的映射关系；基于映射关系，确定第s-1次迭代训练得到的验证损失值。
示意性的，关于θ的映射函数具体可参考公式四：
公式四：
其中，用于表示在s-1次迭代训练过程中关于θ的映射函数，α表示预设的第一学习率，用于表示第s-1次迭代训练过程中的样本权重模型对应的网络结构，也即，包含θ的网络映射函数。
步骤351，基于多个预测损失值和多个损失值分别对应的权重参数，确定多个样本去除模型分别对应的加权损失值。
示意性的，针对多个DICD-Net的优化目标，可以通过预设损失函数，通过对多个DICD-Net的模型参数分别进行调整后，使得损失函数的输出损失值和最小，即为伪影去除模型最终的优化目标，因此，针对伪影去除模型中包含多个DICD-Net的损失函数具体可参考公式五：
公式五：
其中，为当模型中包含B个样本去除模型(DICD-Net)分别对应的预测损失值的损失总和，用于表示第b个DICD-Net对应伪影图像的预测损失值，Wb用于表示第b个预测损失值对应的权重参数。
由于本申请实施例中，样本去除模型对应的权重参数通过样本权重模型获得，因此，其中，表示样本权重模型的网络结构，其输入为预测损失值输出为权重参数W，该网络结构中对应的第二模型参数为θ。
本实施例中，将设置为包含一个隐藏层的多层感知器(Multi-Layer Perception，MLP)网络，示意性的，请参考图7，其示出了本申请一个示例性实施例提供的样本权重模型网络结构示意图，如图7所示，当前显示MLP网络，该网络中包括输入层710、隐藏层720和输出层730，输入层为预测损失值711，输出层730为预测损失值711对应的权重参数731，其中，隐藏层720中包含多个神经元，隐藏层720神经元个数根据实际需要进行限定。
在一些实施例中，在第s次迭代训练过程中，基于第s-1个预测损失值和第s个权重参数，确定第s个加权损失值。
示意性的，加权损失值用于表示经过权重调整后的样本去除模型的对应的预测损失值，具体可参考公式六：
公式六：
其中，用于表示预测损失值对应的权重参数，用于表示预测损失值，用于表示单次迭代训练过程中，B个样本去除模型分别对应的加权损失值的损失总和。
步骤352，基于多个样本去除模型分别对应的加权损失值对多个样本去除模型的第一模型参数分别进行调整，得到多个伪影去除子模型。
示意性的，根据多个加权损失值对多个样本去除模型的第一模型参数进行梯度调整，得到调整后参数作为伪影去除模型对应的模型参数，进而得到伪影去除模型，其具体训练过程可参考公式七：
公式七：
其中，Θ(s-1)用于表示第s-1次迭代训练过程中得到第一模型参数(针对单次迭代训练中B个样本去除模型分别对应的多个第一模型参数)，用于表示通过第s次迭代训练对应的加权损失值对第s-1个第一模型参数进行梯度调整，也即，基于第s次迭代训练对应的加权损失值，对样本去除模型的第s-1次迭代训练得到的第一模型参数进行梯度调整，得到第s次迭代训练对应的第一模型参数，并进行第s+1次循环调整，直至伪影去除模型训练结束，s≥1且s为整数。
值得注意的是，上述针对样本权重模型和样本去除模型进行训练的过程中，其对应的训练顺序为公式四(对θ进行参数化，得到关于θ的映射函数)、公式三(对样本权重模型的第二模型参数进行梯度调整)和公式七(对多个样本去除模型分别对应的第一参数模型进行梯度调整)。
本实施例中，针对样本去除模型的调整效果条件，给出了如公式八对应的训练目标：
公式八：
其中，Θ*(θ)用于表示多个第一模型参数对应的最优解(也即符合调整效果条件的第一参数)，多个第一模型参数的最优解可通过计算加权损失值对应的最小化值获取。
示意性的，针对样本权重模型的第二模型参数，我们也可以给出如公式九对应的训练目标：
公式九：
其中，θ*用于表示第二模型参数的最优解，该最优解可通过计算预测损失值对应的最小化值获取，将第二模型参数达到最优解(或者无限逼近最优解)，将该第二模型参数作为针对样本权重模型的最终训练得到的第二参数。
值得注意的是，公式八和公式九仅为针对样本去除模型和样本权重模型设定的训练目标，在实际训练过程中，对样本去除模型和样本权重模型的训练过程通过公式四、公式三和公式七按照指定顺序进行迭代循环训练。
示意性的，针对多个样本去除模型同时进行训练后，生成得到多个样本去除模型分别对应的伪影子模型，也即，单个伪影子模型和单个样本去除模型对应。
步骤353，将多个伪影去除子模型作为伪影去除模型。
示意性的，伪影去除模型中包括多个训练结束的样本去除模型，也即，伪影去除模型中包括多个伪影去除子模型。
在一些实施例中，响应于第一模型参数的循环迭代调整次数达到次数阈值，将最近一次调整得到的第一模型参数确定为第一参数；或者；响应于第一模型参数的调整效果符合调整效果条件，将第一模型参数确定为第一参数，调整效果条件用于表示对预测损失值的限制要求。
示意性的，伪影去除模型中的模型参数为第一参数，本实施例中针对多个样本去除模型的训练目标即为将多个样本去除模型分别对应的第一模型参数经过梯度调整后得到的结果作为伪影去除模型最终对应的第一参数。
可选地，次数阈值是预先设置好的指定次数或者根据训练情况进行自行设定的，如：针对多个样本去除模型训练100次，则100次即为次数阈值，当对多个样本去除模型的第一模型参数进行循环迭代调整的次数达到100次时，将第100次训练得到的多个第一模型参数确定为第一参数。
可选地，调整效果条件是指当对多个样本去除模型进行梯度调整后，伪影图像输入伪影图像后输出的伪影去除结果和参考图像之间的预测损失值符合调整效果条件后，将多个样本去除模型分别对应的第一模型参数确定为伪影去除网络的第一参数，也即，本次训练的第一模型参数符合对样本去除模型的训练目标。
综上所述，本申请实施例提供的伪影去除模型的训练方法，采用图像内容匹配的参考图像和伪影图像对多个样本去除模型进行训练，其中，在训练过程中将伪影图像输入多个样本去除模型后分别输出多个伪影去除结果，确定多个伪影去除结果和参考图像之间的预测损失值，将预测损失值输入样本权重模型后最终得到各个预测损失值对应的权重参数，根据预测损失值和权重参数对多个样本去除模型进行训练，最终得到包含多个伪影去除子模型的伪影去除模型，通过采用权重参数和预测损失值对多个对应不同预设窗口范围的样本去除模型进行训练，使得最终训练得到的伪影去除模型能够输出不同窗口范围对应的伪影去除图像，满足不同图像的伪影去除需求，提高了伪影去除结果的伪影去除准确度。
本实施例中，通过先对样本权重模型进行训练，再将预测损失值输入根据训练得到的样本权重模型中输出得到权重参数，根据权重参数和多个预测损失值对多个样本去除模型的第一模型参数分别进行梯度调整，从而实现在迭代训练过程中，样本权重模型和样本去除模型交替迭代训练，能够通过对样本权重模型的训练来辅助训练样本去除模型，提高模型训练的准确度和训练效果。
在一个可选的实施例中，针对不同预设窗口范围还包括窗口转换的过程，以图3为例的步骤对本申请的方案应用于辅助诊断领域为例进行说明：
步骤310，获取图像内容匹配的参考图像和伪影图像。
其中，参考图像是对未包含植入物的样本检测对象扫描后生成的图像，伪影图像是包含伪影的参考图像，伪影是在对包含植入物的样本检测对象的扫描过程中植入物产生的阴影。
伪影图像是针对预先获取的公开数据集中的授权CT图像进行图像合成后的CT图像。
本实施例中，针对伪影图像的获取过程，首先使用公开数据集中未被金属影响的CT图像作为参考图像，以及不同类型的金属掩膜(Mask)，按照数据仿真流程，将CT图像和金属掩膜进行图像合成，合成得到包含金属伪影的CT图像，作为训练数据。
将训练数据对应的CT值进行裁剪，得到窗口范围为[-1000,2000]HU的CT图像，然后转为衰减系数，将其归置到[0,1]范围，最终转换到[0,255]范围的CT图像。
对每个训练数据都随机进行图像裁剪，裁剪成边长为64×64大小的图像块，然后以0.5概率分别随机进行水平镜像翻转和垂直镜像翻转，用于最终生成不同的伪影图像。
本实施例中，参考图像和伪影图像实现为一个样本图像对。
示意性的，多个样本去除模型分别对应不同的预设窗口范围，用于标注扫描图像针对不同CT值的情况下，各区域显示的对比度关系，本实施例中，以三个样本去除模型为例进行 说明，也即，本实施例中包括的三个样本去除模型，其各自对应的预设窗口范围分别为[-1000，2000]HU、[-320，480]HU和[-160，240]HU，对此不加以限定。
示意性的，请参考图8，其示出了本申请一个示例性实施例提供的多个样本去除模型示意图，如图8所示，当前显示多窗口的样本去除模型示意图，其中，多窗口的样本去除模型中包括三个不同预设窗口范围分别对应的样本去除模型810、样本去除模型820和样本去除模型830，其中，样本去除模型实现为DICD-Net结构的模型，因此，样本去除模型810为对应第一预设窗口范围([-1000，2000]HU)的DICD-Net(b＝1)，样本去除模型820为对应第二预设窗口范围([-320，480]HU)的DICD-Net(b＝2)，样本去除模型830为对应第三预设窗口范围([-160，240]HU)的DICD-Net(b＝3)。
将伪影图像801输入样本去除模型810，输出得到样本去除模型810对应的第一预测区域811，其中，第一预测区域811用于表示通过样本去除模型810以第一预设窗口范围为基准，对伪影图像801中的金属伪影去除后生成的CT图像。
示意性的，在伪影图像在输入样本去除模型之前，本身也对应有预设窗口范围，因此，针对伪影图像的预设窗口范围和输入的样本去除模型的预设窗口范围之间属于不同的窗口范围时，需要在输入样本去除模型前将伪影图像进行窗口转换，生成与样本去除模型对应的预设窗口范围一致的伪影图像。
如图8所示，将伪影图像801和第一预测区域811输入窗口转换层(Window Layer，图8中标记为W)进行窗口转换，得到伪影图像801对应的第一样本转换结果802，以及第一预测区域811对应的第一预测转换结果812，其中，第一样本转换结果802和第一预测转换结果812对应的窗口范围与样本去除模型820的预设窗口范围保持一致，将第一预测区域811和第一预测转换结果812输入通道融合层(Channel Concatenation，图8中标记为C)，将第一融合结果输入样本去除模型820，输出得到样本去除模型820对应的第二预测区域821将伪影图像801、第一预测区域811和第二预测区域821输入窗口转换层进行窗口转换，得到伪影图像801对应的第二样本转换结果803、第一预测区域811对应的第二预测转换结果813，以及第二预测区域821对应的第三预测转换结果822，将第二样本转换结果803、第二预测转换结果813和第三预测转换结果822输入通道融合层后再输入样本去除模型830，输出得到样本去除模型830对应的第三预测区域831。
在一些实施例中，确定第i个样本去除模型对应窗口范围，i为正整数；将伪影图像和第i-1个伪影去除结果进行窗口转换，得到伪影图像和第i-1伪影去除结果分别对应的窗口转换结果，作为第i个样本去除模型的模型输入。
示意性的，窗口转换的过程可参考如下公式十至公式十二：
公式十：Xori＝Xcurr×(Hcurr-Lcurr)+Lcurr
公式十一：Xclip＝Clip(Xori；[Lnext，Hnext])
公式十二：
其中，Xori用于表示原始图像，Xcurr和Xnext分别表示窗口转换前后对应的扫描图像，[Lcurr，Hcurr]和[Lnext，Hnext]分别表示窗口转换前后分别对应的预设窗口范围(L表示窗位，H表示窗高)。
步骤320，将伪影图像输入多个样本去除模型，分别输出得到伪影图像对应的伪影去除结果。
其中，不同样本去除模型对应不同的预设窗口范围，样本去除模型用于以对应的预设窗口范围为基准，去除伪影图像中的伪影。
本实施例中，根据三个预设窗口范围不同的样本去除模型分别输出得到三个伪影去除结 果，通过预设损失函数分别计算三个伪影去除结果和参考图像之间的像素点距离，输出得到三个预测损失值，分别为和
步骤330，基于伪影去除结果和参考图像之间的像素点差异，确定多个样本去除模型分别对应的预测损失值。
本实施例中，将三个预测损失值分别输入样本权重模型，输出得到三个权重参数，其中，样本权重模型的实现为
步骤340，将多个样本去除模型分别对应的预测损失值输入样本权重模型，输出得到多个预测损失值分别对应的权重参数。
其中，权重参数用于对样本去除模型的参数更新进行权重调整。
本实施例中，根据三个预测损失值和分别对应的三个权重参数，对三个样本去除模型分别同时进行训练，得到三个伪影去除子模型，将这三个伪影去除子模型作为最终的伪影去除网络(Mar Network)。
本实施例中，在针对多个样本去除模型的训练过程中，多个样本去除模型和样本权重模型的训练过程是迭代交替进行的，也即，多个样本去除模型的第一模型参数和样本权重模型的第二模型参数是交替更新的，且通过更新后的第二模型参数对第一模型参数进行参数调整。
示意性的，请参考图9，其示出了本申请一个示例性实施例提供的伪影去除模型的训练方法示意图，如图9所示，将伪影图像910输入样本去除模型920(包含三个，图中未示出该模型数量)，输出得到伪影去除结果931、伪影去除结果932和伪影去除结果933，其中，样本去除模型920和样本权重模型940之间是交替训练的。
如图9所示，针对第s次迭代训练过程，当前获取第s-1次迭代训练后得到第一模型参数(Θ(s-1))和第二模型参数(θ(s-1))，首先根据上述公式四，通过第s-1次迭代训练后得到的第一模型参数确定第s-1次迭代训练后得到的关于θ的映射函数然后根据上述公式三，通过第s-1次迭代训练后得到的关于θ的映射函数确定第s-1次迭代训练后得到的验证损失值，根据第s-1次迭代训练后得到的验证损失值对θ(s-1)进行梯度调整，得到第s次迭代训练后对应的第二模型参数(θ(s))，根据θ(s)确定第s次迭代训练对应的样本权重模型，将第s-1次迭代训练后得到的预测损失值输入第s次迭代训练对应的样本权重模型，得到第s次迭代训练后对应的权重参数，通过公式七，根据第s-1次迭代训练后得到的预测损失值和第s次迭代训练对应的权重参数，确定第s次迭代训练对应的加权损失值，用于对Θ(s-1)进行梯度调整，得到第s次迭代训练对应第二模型参数(Θ(s))，以此进行循环迭代训练，直至样本去除模型训练结束。
值得注意的是，当进行第一次迭代训练时，第一模型参数和第二模型参数均为预先设置好的初始值。
在一些实施例中，获取第一学习衰减率，第一学习衰减率用于根据迭代次数以衰减形式对第一学习率进行调整，第一学习率是预先设定的对多个样本去除模型进行训练的更新步长；在对多个样本去除模型进行训练的过程中，基于第一学习衰减率对第一学习率进行梯度下降，得到伪影去除模型对应的目标学习率。
本实施例中，针对样本去除模型的训练，设置第一学习率为2×10-4，并设置第一学习衰减率为每30个epoch，第一学习率衰减0.5，总训练的epoch数量为200，在每次迭代训练过程中，batch大小为16，样本CT图像块大小为64x64，当第一学习率通过第一学习衰减率进行梯度下降后，若epoch数量达到预设数量(200)，则将当前最后一次梯度下降得到的第一学习率作为样本去除模型的目标学习率。
本实施例中，针对样本权重模型设置第二学习率为1×10-5，隐藏层对应的神经元数量为100。
示意性的，请参考图10，其示出了本申请一个示意性实施例提供的伪影去除模型处理过程示意图，如图10所示，当前处理系统包括前端A1010(如：CT扫描仪)、服务器1020和前端B1030(如：电脑终端或者手机终端)。
当通过前端A1010对目标检测对象进行CT扫描后，若目标检测对象中植入了金属，则当前前端A1010生成的CT图像中包含金属造成的金属伪影区域。
将该CT图像输入服务器1020中的伪影去除模型，对CT图像进行伪影去除，将识别得到的金属伪影区域进行移除后生成不同预设窗口范围下的无金属伪影的CT图像，并将其反馈至前端B1030，用于供医生进行辅助诊断。
综上所述，本申请实施例提供的伪影去除模型的训练方法，采用图像内容匹配的参考图像和伪影图像对多个样本去除模型进行训练，其中，在训练过程中将伪影图像输入多个样本去除模型后分别输出多个伪影去除结果，确定多个伪影去除结果和参考图像之间的预测损失值，将预测损失值输入样本权重模型后最终得到各个预测损失值对应的权重参数，根据预测损失值和权重参数对多个样本去除模型进行训练，最终得到包含多个伪影去除子模型的伪影去除模型，通过采用权重参数和预测损失值对多个对应不同预设窗口范围的样本去除模型进行训练，使得最终训练得到的伪影去除模型能够输出不同窗口范围对应的伪影去除图像，满足不同图像的伪影去除需求，提高了伪影去除结果的伪影去除准确度。
本实施例中，针对基于多个窗口范围的多个样本去除模型进行训练的方式，能够使多个不同窗口范围的样本去除模型进行同时训练，最终得到能够输出不同窗口范围的伪影去除结果的伪影去除模型，能够更好的提高模型的训练效果。
本实施例中，通过窗口转换将不同窗口范围的图像进行转换，能够使得不同窗口范围之间能够更好的进行模型学习，提高模型训练的准确度和灵活性。
请参考图11，其示出了本申请一个示例性实施例提供的伪影去除模型的训练方法示意图，如图11所示，该方法包括如下：
步骤1110，开始。
示意性的，本申请提供的伪影去除模型的训练方法由服务器执行，当前通过向服务器发送训练请求后，服务器开始执行对伪影去除模型的训练过程。
服务器首先判断当前进行训练阶段还是测试阶段，若为训练阶段，则执行步骤1120，若为测试阶段，则执行步骤1150。
步骤1120，获取伪影图像。
本实施例中，伪影图像是指包含金属伪影的CT图像，其中，伪影图像是通过获取公开数据集中的参考图像，结合不同金属掩膜信息，通过数据仿真流程合成得到的包含金属伪影的CT图像，作为训练数据。参考图像和伪影图像对应相同的图像内容，也即，参考图像和伪影图像实现为一个样本图像对。
步骤1130，对样本去除模型的第一模型参数和样本权重模型的第二模型参数进行迭代循环训练。
本实施例中，针对三个样本去除模型分别对应的第一模型参数进行同时梯度调整。
本实施例中，将伪影图像分别输入三个样本去除模型，输出得到三个伪影去除结果。
根据三个伪影去除结果和参考图像之间的像素点差异，通过预设的损失函数确定三个样本去除模型分别对应的预测损失值。
将三个预测损失值分别输入样本权重模型输出得到三个预测损失值分别对应的权重参数W1、W2和W3。
根据预测损失值和权重参数，在循环迭代训练过程中，依次按照上述公式四、公式三和公式七，反向传播到样本权重模型和三个样本去除模型，先后针对样本权重模型的第二模型参数和三个样本去除模型分别对应的第一模型参数进行梯度调整。
当针对第一模型参数进行循环迭代训练的次数达到次数阈值，执行步骤1140，反之继续执行步骤1130。
步骤1140，保存训练模型。
当针对第一模型参数的循环迭代训练次数达到次数阈值后，将最后一次训练得到的多个第一模型参数作为第一参数，确定伪影去除模型，其中，伪影去除模型中包括三个训练好的伪影去除子模型。
服务器将训练好的伪影去除模型进行存储。
步骤1150，获取测试伪影图像。
当服务器判定当前为测试阶段时，获取包含金属伪影的测试伪影图像，其中，测试伪影图像实现为用于对训练好的伪影去除模型进行效果测试的CT图像。
步骤1160，加载已经训练好的伪影去除模型。
服务器获取测试伪影图像后，将存储的已经训练好的伪影去除模型进行加载。
步骤1170，通过前向计算得到伪影去除模型生成的伪影去除结果。
将测试伪影图像输入伪影去除模型，通过针对测试伪影图像进行前向计算，得到测试伪影图像对应的伪影去除结果。其中，伪影去除结果是指对测试伪影图像去除伪影后生成预设窗口范围下的CT图像。
步骤1180，输出伪影去除结果对应的不同预设窗口范围的CT图像。
将通过伪影去除模型同时输出得到的针对三种不同预设窗口范围的伪影去除结果进行输出。
本实施例中，将最终训练好的三个伪影去除子模型，作为伪影去除模型，也即，当前将目标图像输入伪影去除模型后，将同时输出三个对应不同预设窗口范围的伪影去除结果。示意性的，请参考图12，其示出了本申请一个示例性实施例提供的伪影去除模型应用过程示意图，如图12所示，当目标图像1210(本实施例中，提供了针对同一腹部组织在同一窗口范围下，三种不同显示方式的腹部CT图像，分别为图像1211，图像1212和图像1213)，将其输入伪影去除模型1220后，同时输出三种不同窗口范围的伪影去除结果(包括图像1211对应的三种伪影去除结果111，图像1212对应的三种伪影去除结果122，图像1213对应的三种伪影去除结果133)。该伪影去除结果用于辅助医生对腹部组织进行诊断。
综上所述，本申请实施例提供的伪影去除模型的训练方法，采用图像内容匹配的参考图像和伪影图像对多个样本去除模型进行训练，其中，在训练过程中将伪影图像输入多个样本去除模型后分别输出多个伪影去除结果，确定多个伪影去除结果和参考图像之间的预测损失值，将预测损失值输入样本权重模型后最终得到各个预测损失值对应的权重参数，根据预测损失值和权重参数对多个样本去除模型进行训练，最终得到包含多个伪影去除子模型的伪影去除模型，通过采用权重参数和预测损失值对多个对应不同预设窗口范围的样本去除模型进行训练，使得最终训练得到的伪影去除模型能够输出不同窗口范围对应的伪影去除图像，满足不同图像的伪影去除需求，提高了伪影去除结果的伪影去除准确度。
本申请提供的有益效果包括如下几点：
1.所设计的样本去除模型(DICD-Net)具有良好的可解释性，这可以方便用户很好地了解模型中每个模块的功能；
2.在不同窗口范围之间，引入了样本权重模型，使得不同窗口范围的重建复原学习过程更加灵活，这对充分提升组织结构的保真度具有更好的潜力；
3.所重建复原的不同对比度CT图像(伪影去除结果)，有利于对不同组织和器官进行更细精细的观察，从而更好地方便后续的诊断。
图13是本申请一个示例性实施例提供的伪影去除模型的训练装置的结构框图，如图13所示，该装置包括如下部分：
获取模块1310，用于获取图像内容匹配的参考图像和伪影图像，所述参考图像是对未包含植入物的样本检测对象扫描后生成的图像，所述伪影图像是包含伪影的参考图像，所述伪影是在对包含所述植入物的样本检测对象的扫描过程中所述植入物产生的阴影；也就是说，伪影是所述植入物在扫描过程中产生的阴影；
输入模块1320，用于将所述伪影图像输入多个样本去除模型，获得所述多个样本去除模型分别输出的所述伪影图像对应的伪影去除结果，不同样本去除模型对应不同的预设窗口范围，所述样本去除模型用于以对应的预设窗口范围为基准，去除所述伪影图像中的伪影；
确定模块1330，用于基于所述伪影去除结果和所述参考图像之间的像素点差异，确定多个样本去除模型分别对应的预测损失值；
所述输入模块1320，还用于将多个样本去除模型分别对应的预测损失值输入样本权重模型，输出得到多个预测损失值分别对应的权重参数，所述权重参数用于对所述样本去除模型的参数更新进行权重调整；
训练模块1340，用于基于所述预测损失值和所述权重参数，对多个样本去除模型进行训练，得到多个伪影去除子模型构成的伪影去除模型，所述伪影去除子模型用于基于对应的预设窗口范围对目标图像进行伪影去除。
在一个可选的实施例中，如图14所示，所述训练模块1340，包括：
确定单元1341，用于基于多个预测损失值和多个损失值分别对应的权重参数，确定多个样本去除模型分别对应的加权损失值；
调整单元1342，用于基于多个样本去除模型分别对应的加权损失值对多个样本去除模型的第一模型参数分别进行调整，得到多个伪影去除子模型；
所述确定单元1341，还用于将所述多个伪影去除子模型作为所述伪影去除模型。
在一个可选的实施例中，所述确定单元1341，还用于在第s次迭代训练过程中，基于第s-1次迭代训练得到的预测损失值和第s次迭代训练得到的权重参数，确定第s次迭代训练对应的加权损失值；
所述调整单元1342，还用于基于所述第s次迭代训练对应的加权损失值，对所述样本去除模型的第s-1次迭代训练得到的第一模型参数进行梯度调整，得到第s次迭代训练对应的第一模型参数，并进行第s+1次循环调整，直至所述伪影去除模型训练结束，s≥1且s为整数。
在一个可选的实施例中，所述输入模块1320，还用于将第s-1次迭代训练得到的预测损失输入第s次迭代训练得到的样本权重模型，输出得到第s次迭代训练对应的权重参数。
在一个可选的实施例中，所述获取模块1310，还用于获取图像内容匹配的验证参考图像和验证伪影图像；
所述输入模块1320，还用于将所述验证伪影图像输入多个样本去除模型，分别输出得到所述验证伪影图像对应的验证去除结果；
所述确定模块1330，还用于基于所述验证去除结果和所述验证参考图像之间的像素点差异，确定多个样本去除模型分别对应的验证损失值；
所述训练模块1340，还用于基于所述验证损失值，对所述样本权重模型进行训练。
在一个可选的实施例中，所述训练模块1340，还用于在s次迭代训练过程中，基于第s-1次迭代训练得到的验证损失值，对所述样本权重模型的第二模型参数进行梯度调整，得到第s次迭代训练对应的样本权重模型。
在一个可选的实施例中，所述确定模块1330，还用于基于第s-1次迭代训练得到的第一模型参数，确定第一模型参数和第二模型参数之间在第s-1次迭代训练过程中对应的映射关 系；基于所述映射关系，确定第s-1次迭代训练得到的验证损失值。
在一个可选的实施例中，所述确定模块1330，还用于响应于所述第一模型参数的循环迭代调整次数达到次数阈值，将最近一次调整得到的第一模型参数确定为所述第一参数；或者；响应于所述第一模型参数的调整效果符合调整效果条件，将所述第一模型参数确定为所述第一参数，所述调整效果条件用于表示对所述预测损失值的限制要求。
在一个可选的实施例中，所述获取模块1310，还用于获取第一学习衰减率，所述第一学习衰减率用于根据迭代次数以衰减形式对第一学习率进行调整，所述第一学习率是预先设定的对多个样本去除模型进行训练的更新步长；
所述装置还包括：
梯度下降模块1350，用于在对多个样本去除模型进行训练的过程中，基于所述第一学习衰减率对所述第一学习率进行梯度下降，得到所述伪影去除模型对应的目标学习率。
在一个可选的实施例中，所述确定模块1330，还用于确定第i个样本去除模型对应窗口范围，i为正整数；
所述装置还包括：
转换模块1360，用于将所述伪影图像和第i-1个伪影去除结果进行窗口转换，得到所述伪影图像和第i-1伪影去除结果分别对应的窗口转换结果，作为第i个样本去除模型的模型输入。
综上所述，本申请实施例提供的伪影去除模型的训练装置，采用图像内容匹配的参考图像和伪影图像对多个样本去除模型进行训练，其中，在训练过程中将伪影图像输入多个样本去除模型后分别输出多个伪影去除结果，确定多个伪影去除结果和参考图像之间的预测损失值，将预测损失值输入样本权重模型后最终得到各个预测损失值对应的权重参数，根据预测损失值和权重参数对多个样本去除模型进行训练，最终得到包含多个伪影去除子模型的伪影去除模型，通过采用权重参数和预测损失值对多个对应不同预设窗口范围的样本去除模型进行训练，使得最终训练得到的伪影去除模型能够输出不同窗口范围对应的伪影去除图像，满足不同图像的伪影去除需求，提高了伪影去除结果的伪影去除准确度。
需要说明的是：上述实施例提供的伪影去除模型的训练装置，仅以上述各功能模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能模块完成，即将设备的内部结构划分成不同的功能模块，以完成以上描述的全部或者部分功能。另外，上述实施例提供的伪影去除模型的训练装置与伪影去除模型的训练方法实施例属于同一构思，其具体实现过程详见方法实施例，此处不再赘述。
图15示出了本申请一个示例性实施例提供的服务器的结构示意图。具体来讲：
服务器1500包括中央处理单元(Central Processing Unit，CPU)1501、包括随机存取存储器(Random Access Memory，RAM)1502和只读存储器(Read Only Memory，ROM)1503的系统存储器1504，以及连接系统存储器1504和中央处理单元1501的系统总线1505。服务器1500还包括用于存储操作系统1513、应用程序1514和其他程序模块1515的大容量存储设备1506。
大容量存储设备1506通过连接到系统总线1505的大容量存储控制器(未示出)连接到中央处理单元1501。大容量存储设备1506及其相关联的计算机可读介质为服务器1500提供非易失性存储。也就是说，大容量存储设备1506可以包括诸如硬盘或者紧凑型光盘只读存储器(Compact Disc Read Only Memory，CD-ROM)驱动器之类的计算机可读介质(未示出)。
不失一般性，计算机可读介质可以包括计算机存储介质和通信介质。计算机存储介质包括以用于存储诸如计算机可读指令、数据结构、程序模块或其他数据等信息的任何方法或技术实现的易失性和非易失性、可移动和不可移动介质。计算机存储介质包括RAM、ROM、可擦除可编程只读存储器(Erasable Programmable Read Only Memory，EPROM)、带电可擦可编 程只读存储器(Electrically Erasable Programmable Read Only Memory，EEPROM)、闪存或其他固态存储技术，CD-ROM、数字通用光盘(Digital Versatile Disc，DVD)或其他光学存储、磁带盒、磁带、磁盘存储或其他磁性存储设备。当然，本领域技术人员可知计算机存储介质不局限于上述几种。上述的系统存储器1504和大容量存储设备1506可以统称为存储器。
根据本申请的各种实施例，服务器1500还可以通过诸如因特网等网络连接到网络上的远程计算机运行。也即服务器1500可以通过连接在系统总线1505上的网络接口单元1511连接到网络1512，或者说，也可以使用网络接口单元1511来连接到其他类型的网络或远程计算机系统(未示出)。
上述存储器还包括一个或者一个以上的程序，一个或者一个以上程序存储于存储器中，被配置由CPU执行。
本申请的实施例还提供了一种计算机设备，该计算机设备包括处理器和存储器，该存储器中存储有至少一条指令、至少一段程序、代码集或指令集，至少一条指令、至少一段程序、代码集或指令集由处理器加载并执行以实现上述各方法实施例提供的伪影去除模型的训练方法。
本申请的实施例还提供了一种计算机可读存储介质，该计算机可读存储介质上存储有至少一条指令、至少一段程序、代码集或指令集，至少一条指令、至少一段程序、代码集或指令集由处理器加载并执行，以实现上述各方法实施例提供的伪影去除模型的训练方法。
本申请的实施例还提供了一种计算机程序产品或计算机程序，该计算机程序产品或计算机程序包括计算机指令，该计算机指令存储在计算机可读存储介质中。计算机设备的处理器从计算机可读存储介质读取该计算机指令，处理器执行该计算机指令，使得该计算机设备执行上述实施例中任一所述的伪影去除模型的训练方法。

Claims (20)

  1. 一种伪影去除模型的训练方法，所述方法由计算机设备执行，所述方法包括：
    获取图像内容匹配的参考图像和伪影图像，所述参考图像是对未包含植入物的样本检测对象扫描后生成的图像，所述伪影图像是包含伪影的参考图像，所述伪影是所述植入物在扫描过程中产生的阴影；
    将所述伪影图像输入多个样本去除模型，获得所述多个样本去除模型分别输出的所述伪影图像对应的伪影去除结果，不同样本去除模型对应不同的预设窗口范围，所述样本去除模型用于以对应的预设窗口范围为基准，去除所述伪影图像中的伪影；
    基于所述伪影去除结果和所述参考图像之间的像素点差异，确定多个样本去除模型分别对应的预测损失值；
    将多个样本去除模型分别对应的预测损失值输入样本权重模型，输出得到多个预测损失值分别对应的权重参数，所述权重参数用于对所述样本去除模型的参数更新进行权重调整；
    基于所述预测损失值和所述权重参数，对多个样本去除模型进行训练，得到多个伪影去除子模型构成的伪影去除模型，所述伪影去除子模型用于基于对应的预设窗口范围对目标图像进行伪影去除。
  2. 根据权利要求1所述的方法，所述基于所述预测损失值和所述权重参数，对多个样本去除模型进行训练，得到多个伪影去除子模型构成的伪影去除模型，包括：
    基于多个预测损失值和多个损失值分别对应的权重参数，确定多个样本去除模型分别对应的加权损失值；
    基于多个样本去除模型分别对应的加权损失值对多个样本去除模型的第一模型参数分别进行调整，得到多个伪影去除子模型；
    将所述多个伪影去除子模型作为所述伪影去除模型。
  3. 根据权利要求2所述的方法，所述基于所述预测损失值和所述权重参数，确定加权损失值，包括：
    在第s次迭代训练过程中，基于第s-1次迭代训练得到的预测损失值和第s次迭代训练得到的权重参数，确定第s次迭代训练对应的加权损失值；
    所述基于所述加权损失值对所述样本去除模型的第一模型参数进行调整，得到所述伪影去除模型，包括：
    基于所述第s次迭代训练对应的加权损失值，对所述样本去除模型的第s-1次迭代训练得到的第一模型参数进行梯度调整，得到第s次迭代训练对应的第一模型参数，并进行第s+1次循环调整，直至所述伪影去除模型训练结束，s≥1且s为整数。
  4. 根据权利要求3所述的方法，所述将所述预测损失值输入样本权重模型，输出得到权重参数，包括：
    将第s-1次迭代训练得到的预测损失输入第s次迭代训练得到的样本权重模型，输出得到第s次迭代训练对应的权重参数。
  5. 根据权利要求1至4任一所述的方法，所述将所述预测损失值输入样本权重模型，输出得到权重参数之前，还包括：
    获取图像内容匹配的验证参考图像和验证伪影图像；
    将所述验证伪影图像输入多个样本去除模型，分别输出得到所述验证伪影图像对应的验证去除结果；
    基于所述验证去除结果和所述验证参考图像之间的像素点差异，确定多个样本去除模型分别对应的验证损失值；
    基于所述验证损失值，对所述样本权重模型进行训练。
  6. 根据权利要求5所述的方法，所述基于所述验证损失值，对所述样本权重模型进行训练，包括：
    在s次迭代训练过程中，基于第s-1次迭代训练得到的验证损失值，对所述样本权重模型的第二模型参数进行梯度调整，得到第s次迭代训练对应的样本权重模型。
  7. 根据权利要求6所述的方法，所述基于第s-1次迭代训练得到的验证损失值，对所述样本权重模型的第二模型参数进行梯度调整，得到第s次迭代训练对应的样本权重模型之前，还包括：
    基于第s-1次迭代训练得到的第一模型参数，确定第一模型参数和第二模型参数之间在第s-1次迭代训练过程中对应的映射关系；
    基于所述映射关系，确定第s-1次迭代训练得到的验证损失值。
  8. 根据权利要求1至4任一所述的方法，所述方法还包括：
    响应于所述第一模型参数的循环迭代调整次数达到次数阈值，将最近一次调整得到的第一模型参数确定为所述第一参数；
    或者；
    响应于所述第一模型参数的调整效果符合调整效果条件，将所述第一模型参数确定为所述第一参数，所述调整效果条件用于表示对所述预测损失值的限制要求。
  9. 根据权利要求1至4任一所述的方法，所述方法还包括：
    获取第一学习衰减率，所述第一学习衰减率用于根据迭代次数以衰减形式对第一学习率进行调整，所述第一学习率是预先设定的对多个样本去除模型进行训练的更新步长；
    在对多个样本去除模型进行训练的过程中，基于所述第一学习衰减率对所述第一学习率进行梯度下降，得到所述伪影去除模型对应的目标学习率。
  10. 根据权利要求1至4任一所述的方法，所述方法还包括：
    确定第i个样本去除模型对应的窗口范围，i为正整数；
    将所述伪影图像和第i-1个伪影去除结果进行窗口转换，得到所述伪影图像和第i-1伪影去除结果分别对应的窗口转换结果，作为第i个样本去除模型的模型输入。
  11. 一种伪影去除模型的训练装置，所述装置包括：
    获取模块，用于获取图像内容匹配的参考图像和伪影图像，所述参考图像是对未包含植入物的样本检测对象扫描后生成的图像，所述伪影图像是包含伪影的参考图像，所述伪影是所述植入物在扫描过程中产生的阴影；
    输入模块，用于将所述伪影图像输入多个样本去除模型，获得所述多个样本去除模型分别输出的所述伪影图像对应的伪影去除结果，不同样本去除模型对应不同的预设窗口范围，所述样本去除模型用于以对应的预设窗口范围为基准，去除所述伪影图像中的伪影；
    确定模块，用于基于所述伪影去除结果和所述参考图像之间的像素点差异，确定多个样本去除模型分别对应的预测损失值；
    所述输入模块，用于将多个样本去除模型分别对应的预测损失值输入样本权重模型，输出得到多个预测损失值分别对应的权重参数，所述权重参数用于对所述样本去除模型的参数更新进行权重调整；
    训练模块，用于基于所述预测损失值和所述权重参数，对多个样本去除模型进行训练，得到多个伪影去除子模型构成的伪影去除模型，所述伪影去除子模型用于基于对应的预设窗口范围对目标图像进行伪影去除。
  12. 根据权利要求11所述的装置，所述训练模块，包括：
    确定单元，用于基于多个预测损失值和多个损失值分别对应的权重参数，确定多个样本去除模型分别对应的加权损失值；
    调整单元，用于基于多个样本去除模型分别对应的加权损失值对多个样本去除模型的第一模型参数分别进行调整，得到多个伪影去除子模型；
    所述确定单元，还用于将所述多个伪影去除子模型作为所述伪影去除模型。
  13. 根据权利要求12所述的装置，所述确定单元，还用于在第s次迭代训练过程中，基于第s-1次迭代训练得到的预测损失值和第s次迭代训练得到的权重参数，确定第s次迭代训练对应的加权损失值；
    所述调整单元，还用于基于所述第s次迭代训练对应的加权损失值，对所述样本去除模型的第s-1次迭代训练得到的第一模型参数进行梯度调整，得到第s次迭代训练对应的第一模型参数，并进行第s+1次循环调整，直至所述伪影去除模型训练结束，s≥1且s为整数。
  14. 根据权利要求13所述的装置，所述输入模块，还用于将第s-1次迭代训练得到的预测损失输入第s次迭代训练得到的样本权重模型，输出得到第s次迭代训练对应的权重参数。
  15. 根据权利要求11至14任一所述的装置，
    所述获取模块，还用于获取图像内容匹配的验证参考图像和验证伪影图像；
    所述输入模块，还用于将所述验证伪影图像输入多个样本去除模型，分别输出得到所述验证伪影图像对应的验证去除结果；
    所述确定模块，还用于基于所述验证去除结果和所述验证参考图像之间的像素点差异，确定多个样本去除模型分别对应的验证损失值；
    所述训练模块，还用于基于所述验证损失值，对所述样本权重模型进行训练。
  16. 根据权利要求15所述的装置，所述训练模块，还用于在s次迭代训练过程中，基于第s-1次迭代训练得到的验证损失值，对所述样本权重模型的第二模型参数进行梯度调整，得到第s次迭代训练对应的样本权重模型。
  17. 根据权利要求16所述的装置，所述确定模块，还用于基于第s-1次迭代训练得到的第一模型参数，确定第一模型参数和第二模型参数之间在第s-1次迭代训练过程中对应的映射关系；基于所述映射关系，确定第s-1次迭代训练得到的验证损失值。
  18. 一种计算机设备，所述计算机设备包括处理器和存储器，所述存储器中存储至少一段程序，所述至少一段程序由所述处理器加载并执行以实现如权利要求1至10任一所述的伪影去除模型的训练方法。
  19. 一种计算机可读存储介质，所述存储介质中存储有至少一段程序，所述至少一段程序由处理器加载并执行以实现如权利要求1至10任一所述的伪影去除模型的训练方法。
  20. 一种计算机程序产品，包括计算机指令，所述计算机指令被处理器执行时实现如权利要求1至10任一所述的伪影去除模型的训练方法。
PCT/CN2023/096836 2022-08-09 2023-05-29 伪影去除模型的训练方法、装置、设备、介质及程序产品 WO2024032098A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210951294.0A CN115330615A (zh) 2022-08-09 2022-08-09 伪影去除模型的训练方法、装置、设备、介质及程序产品
CN202210951294.0 2022-08-09

Publications (1)

Publication Number Publication Date
WO2024032098A1 true WO2024032098A1 (zh) 2024-02-15

Family

ID=83921833

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/096836 WO2024032098A1 (zh) 2022-08-09 2023-05-29 伪影去除模型的训练方法、装置、设备、介质及程序产品

Country Status (2)

Country Link
CN (1) CN115330615A (zh)
WO (1) WO2024032098A1 (zh)

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115330615A (zh) * 2022-08-09 2022-11-11 腾讯医疗健康(深圳)有限公司 伪影去除模型的训练方法、装置、设备、介质及程序产品
CN115690255B (zh) * 2023-01-04 2023-05-09 浙江双元科技股份有限公司 基于卷积神经网络的ct图像去伪影方法、装置及系统
CN116228916B (zh) * 2023-05-10 2023-07-11 中日友好医院(中日友好临床医学研究所) 一种图像去金属伪影方法、系统及设备
CN116993845B (zh) * 2023-06-09 2024-03-15 西安交通大学 一种基于集成深度网络DnCNN的CT图像去伪影方法
CN117649358B (zh) * 2024-01-30 2024-04-16 腾讯科技（深圳）有限公司 图像处理方法、装置、设备及存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3731144A1 (en) * 2019-04-25 2020-10-28 Koninklijke Philips N.V. Deep adversarial artifact removal
CN112446873A (zh) * 2020-12-11 2021-03-05 深圳高性能医疗器械国家研究院有限公司 一种图像伪影的去除方法
CN112862728A (zh) * 2021-03-22 2021-05-28 上海壁仞智能科技有限公司 伪影去除方法、装置、电子设备和存储介质
CN113689359A (zh) * 2021-09-23 2021-11-23 上海联影医疗科技股份有限公司 一种图像伪影去除模型及其训练方法和系统
CN114241070A (zh) * 2021-12-01 2022-03-25 北京长木谷医疗科技有限公司 Ct图像去金属伪影以及模型训练的方法和装置
CN115330615A (zh) * 2022-08-09 2022-11-11 腾讯医疗健康(深圳)有限公司 伪影去除模型的训练方法、装置、设备、介质及程序产品

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3731144A1 (en) * 2019-04-25 2020-10-28 Koninklijke Philips N.V. Deep adversarial artifact removal
CN112446873A (zh) * 2020-12-11 2021-03-05 深圳高性能医疗器械国家研究院有限公司 一种图像伪影的去除方法
CN112862728A (zh) * 2021-03-22 2021-05-28 上海壁仞智能科技有限公司 伪影去除方法、装置、电子设备和存储介质
CN113689359A (zh) * 2021-09-23 2021-11-23 上海联影医疗科技股份有限公司 一种图像伪影去除模型及其训练方法和系统
CN114241070A (zh) * 2021-12-01 2022-03-25 北京长木谷医疗科技有限公司 Ct图像去金属伪影以及模型训练的方法和装置
CN115330615A (zh) * 2022-08-09 2022-11-11 腾讯医疗健康(深圳)有限公司 伪影去除模型的训练方法、装置、设备、介质及程序产品

Also Published As

Publication number Publication date
CN115330615A (zh) 2022-11-11

Similar Documents

Publication Publication Date Title
WO2024032098A1 (zh) 伪影去除模型的训练方法、装置、设备、介质及程序产品
WO2022121160A1 (zh) 一种基于深度学习的增强ct图像质量和分辨率的方法
CN109727663A (zh) 生成用于图像处理的神经网络模型的系统和方法
CN108564553A (zh) 基于卷积神经网络的低剂量ct图像噪声抑制方法
CN112348936B (zh) 一种基于深度学习的低剂量锥束ct图像重建方法
CN108537794A (zh) 医学图像数据处理方法、装置和计算机可读存储介质
CN109949235A (zh) 一种基于深度卷积神经网络的胸部x光片去噪方法
CN107958471A (zh) 基于欠采样数据的ct成像方法、装置、ct设备及存储介质
CN112598649B (zh) 基于生成对抗网络的2d/3d脊椎ct非刚性配准方法
CN114881943B (zh) 基于人工智能的脑龄预测方法、装置、设备及存储介质
CN113744275A (zh) 一种基于特征变换的三维cbct牙齿图像的分割方法
CN113850796A (zh) 基于ct数据的肺部疾病识别方法及装置、介质和电子设备
CN116894783A (zh) 基于时变约束的对抗生成网络模型的金属伪影去除方法
JP7430232B2 (ja) グラフ学習に基づくスパース投影再構成方法
EP4009268A1 (en) Performing denoising on an image
JP2023546208A (ja) ノイズ除去のためのニューラルネットワークモデルをトレーニング及びチューニングする方法並びにシステム
CN112801994A (zh) 骨龄评估方法和系统
CN113538451B (zh) 一种深静脉血栓的磁共振图像分割方法及装置，电子设备及存储介质
CN115281709B (zh) 一种下颌第二磨牙c形根管检测装置及方法
CN114693831B (zh) 一种图像处理方法、装置、设备和介质
WO2022091869A1 (ja) 医用画像処理装置、医用画像処理方法及びプログラム
CN114638907B (zh) 基于深度学习的骨硬化伪影校正方法及装置
Mangileva et al. Removing artifacts from computed tomography images of heart using neural network with partial convolution layer
CN117173271A (zh) Ct图像处理方法、模型训练方法、设备、介质及产品
CN117036162B (zh) 轻量级胸部ct图像超分辨率的残差特征注意力融合方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23851326

Country of ref document: EP

Kind code of ref document: A1