Training method of bias field prediction model, method for correcting image bias field and related product

Abstract

The invention discloses a training method of a bias field prediction model, a method for correcting an image bias field and a related product. The training method comprises the following steps: acquiring an original image sample and a corresponding bias field sample; constructing a first training data set based on the original image sample and the bias field sample; and performing first training on the bias field prediction model by using the first training data set. According to the scheme of the embodiment of the invention, the method can achieve the quick and accurate prediction of the image bias field, and facilitates the accurate correction of the bias field effect of the image.

Description

Technical Field

[0001] The present disclosure generally relates to the field of image processing technology. More specifically, the present disclosure relates to a method for training a bias field prediction model, a method for correcting an image bias field, and related products. Background Art

[0002] Magnetic resonance imaging (MRI) is an imaging technology widely used in medical diagnosis and research. It uses magnetic fields and radio frequency pulses to make hydrogen nuclei in tissues resonate, generate magnetic resonance signals, and then use gradient magnetic fields for spatial encoding to finally obtain images of tissues. However, due to the influence of hardware equipment and environmental factors, MRI images often have uneven intensity during the acquisition process. This phenomenon is called bias field effect, which will seriously affect the quality of the image and subsequent diagnosis and analysis.

[0003] In view of this, there is an urgent need to provide a solution for correcting the image bias field so as to improve the image quality and facilitate subsequent image applications. Summary of the invention

[0004] In order to at least solve one or more of the technical problems mentioned above, the present disclosure proposes a training method for a bias field prediction model, a method and device for correcting an image bias field, and a computer-readable storage medium solution in multiple aspects.

[0005] In a first aspect, the present disclosure provides a method for training a bias field prediction model for predicting image bias field information, comprising: obtaining original image samples and corresponding bias field samples; constructing a first training data set based on the original image samples and the bias field samples; and performing a first training on the bias field prediction model using the first training data set.

[0006] In some embodiments, obtaining original image samples and corresponding bias field samples includes: obtaining the original image samples and corrected image samples of the original image samples; and dividing the original image samples by corresponding elements in the corrected image samples to obtain corresponding bias field samples.

[0007] In some other embodiments, constructing a first training data set based on the original image samples and the bias field samples includes: forming image pairs of the original image samples and the corresponding bias field samples as the first training data set.

[0008] In some further embodiments, constructing a first training data set based on the original image samples and the bias field samples includes: taking the inverse of each element in the bias field samples to obtain corresponding bias field inverse samples; and combining the original image samples and the corresponding bias field inverse samples into image pairs as the first training data set.

[0009] In some embodiments, the training method further comprises: downsampling samples in the first training data set so as to use the downsampled first training data set to perform a first training on the bias field prediction model.

[0010] In some other embodiments, the training method further includes: using a diffusion model to amplify the first training data set to obtain a second training data set; and using the second training data set to perform a second training on the bias field prediction model after the first training.

[0011] In some further embodiments, the bias field prediction model includes an encoder and a decoder, and using a diffusion model to amplify the first training data set to obtain a second training data set includes: using the latent vector output by the encoder of the bias field prediction model after the first training to perform a third training on the diffusion model; using the diffusion model after the third training to add noise to the latent vector to generate a new latent vector; inputting the new latent vector into the decoder to output new bias field information; and amplifying the first training data set based on a new image pair consisting of the new bias field information and corresponding original image samples to obtain the second training data set.

[0012] In some embodiments, the training method further includes: downsampling samples in the second training data set so as to use the downsampled second training data set to perform the second training on the bias field prediction model.

[0013] In other embodiments, the bias field prediction model includes at least one of a VAE model, a generative adversarial network, and variants thereof.

[0014] In yet other embodiments, the original image samples include MRI image samples.

[0015] In a second aspect, the present disclosure provides a method for correcting an image bias field, comprising: inputting an image to be corrected into a bias field prediction model trained according to any training method described in the first aspect of the present disclosure to output bias field information of the image to be corrected; and obtaining a corrected image based on the image to be corrected and the corresponding bias field information.

[0016] In some embodiments, if the bias field information is a bias field image, then based on the image to be corrected and the corresponding bias field information, obtaining the corrected image includes: dividing the image to be corrected by corresponding elements in the bias field image to obtain the corrected image; or if the bias field information is a bias field reciprocal image, then based on the image to be corrected and the corresponding bias field information, obtaining the corrected image includes: multiplying the image to be corrected by corresponding elements in the bias field reciprocal image to obtain the corrected image.

[0017] In some other embodiments, the method further includes: downsampling the image to be corrected; inputting the downsampled image to be corrected into the bias field prediction model to output bias field information of the downsampled size; and upsampling the bias field information of the downsampled size to the same size as the image to be corrected, so as to obtain a corrected image of the same size as the image to be corrected after calculation with the image to be corrected.

[0018] In some further embodiments, the image to be corrected includes an MRI image.

[0019] In a third aspect, the present disclosure provides a device for correcting an image bias field, comprising: a processor for executing program instructions; and a memory storing the program instructions, wherein when the program instructions are loaded and executed by the processor, the processor executes the training method described in any one of the first aspects of the present disclosure or executes the method described in any one of the second aspects of the present disclosure.

[0020] In a fourth aspect, the present disclosure provides a computer-readable storage medium having computer-readable instructions stored thereon, which, when executed by one or more processors, implement the training method as described in any one of the first aspects of the present disclosure or the method as described in any one of the second aspects of the present disclosure.

[0021] It can be seen from the training method of the bias field prediction model and the scheme for correcting the image bias field provided above that the disclosed embodiment can achieve fast and accurate image bias field estimation by using a first training data set composed of original image samples and corresponding bias field samples to train the bias field prediction model, thereby facilitating accurate correction of the bias field effect of the image, so as to improve the quality of the image and facilitate subsequent analysis based on the image (such as image classification, image segmentation, etc.) and medical diagnosis, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] By reading the detailed description below with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0023] Figure 1 An exemplary flow chart of a method for training a bias field prediction model according to some embodiments of the present disclosure is shown;

[0024] Figure 2 A schematic structural block diagram of a VAE model is shown;

[0025] Figure 3 An exemplary flow chart of a training method for performing first training using bias field reciprocal samples according to some embodiments of the present disclosure is shown;

[0026] Figure 4 An exemplary flow chart showing a training method including a second training according to some embodiments of the present disclosure;

[0027] Figure 5 A flow chart of a method for correcting an image bias field according to some embodiments of the present disclosure is shown;

[0028] Figure 6 A schematic block diagram of a system for correcting an image bias field according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present disclosure.

[0030] It should be understood that the terms "include" and "comprising" used in the specification and claims of the present disclosure indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0031] It should also be understood that the terms used in this disclosure are only for the purpose of describing specific embodiments and are not intended to limit the disclosure. As used in this disclosure and claims, the singular forms of "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should also be further understood that the term "and / or" used in this disclosure and claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations.

[0032] As used in this specification and claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0033] The inventors have found that in order to solve the bias field problem of MRI images, methods based on physical / mathematical models and statistics are currently commonly used. Methods based on physical / mathematical models usually simulate the uneven distribution of intensity by establishing a mathematical model, and then restore the original image by inversion. Methods based on statistics usually estimate a correction parameter by counting the pixel values ​​in the image, and then use this parameter to correct the image.

[0034] However, these methods often require complex calculation processes, which are time-consuming and not conducive to the clinical application of medical images. In addition, these methods often require prior knowledge or assumptions, and the accuracy of these prior knowledge or assumptions may affect the correction effect. For example, the method based on mathematical models to simulate intensity distribution requires understanding of the specific acquisition parameters of magnetic resonance imaging, such as scanning sequence, magnetic field strength, etc. For example, the N4 bias field correction method requires the assumption that the bias field is a smooth low-frequency signal.

[0035] Based on this, the present disclosure hopes to provide a technical solution that can quickly and accurately correct the image bias field, thereby improving the quality of the image and facilitating the subsequent further application of the image. The specific implementation of the present disclosure is described in detail below in conjunction with the accompanying drawings.

[0036] Figure 1 An exemplary flow chart of a method for training a bias field prediction model according to some embodiments of the present disclosure is shown. Figure 1As shown in , a training method 100 for a bias field prediction model for predicting image bias field information may include: in step S110, an original image sample and a corresponding bias field sample may be obtained. In some embodiments, the original image sample may include a medical image sample. In other embodiments, the original image sample may include an MRI image sample. According to such a setting, the trained bias field prediction model can be used for the prediction of bias field information of MRI images.

[0037] The corresponding bias field samples mentioned above refer to the bias field samples corresponding to the original image samples. In other words, each bias field sample can be estimated from a corresponding original image sample. In other embodiments, one original image sample can obtain multiple bias field samples accordingly. In some embodiments, the original image samples and the corresponding bias field samples can be obtained from existing historical measurement data or historical training data sets. In other embodiments, the original image samples and the corresponding bias field samples can be obtained from the correction results of traditional technologies.

[0038] In some other embodiments, obtaining the original image sample and the corresponding bias field sample may include: obtaining the original image sample and the corrected image sample of the original image sample; and dividing the original image sample with the corresponding element in the corrected image sample to obtain the corresponding bias field sample. Here, the corresponding element may be a pixel at a corresponding position in the two image samples. By dividing the pixel value at the corresponding position in the original image sample with the pixel value in the corrected image sample, the division result obtained is the bias field sample corresponding to the original image sample.

[0039] In a specific implementation, first, multiple original image samples may be collected; then, one or more of the conventional methods such as the N4 bias field correction method, the N3 bias field correction method, and other methods based on mathematical models and statistics may be used to perform bias field correction on the multiple original image samples to obtain corrected image samples corresponding to the multiple original image samples; further, the corresponding elements in each pair of original image samples and corrected image samples may be divided to obtain the bias field samples corresponding to each original image sample. In other embodiments, conventional correction methods such as the N4 bias field correction method and the N3 bias field correction method may be used to process the original image samples to directly output the corresponding corrected image samples and bias field samples, thereby directly obtaining the original image samples and the corresponding bias field sample pairs.

[0040] Next, in step S120, a first training data set may be constructed based on the original image samples and the bias field samples. In some embodiments, step S120 may include: forming an image pair with the original image samples and the corresponding bias field samples as the first training data set. The first training data set may include multiple image pairs, where the original image samples and the corresponding bias field samples in each image pair may correspond one to one.

[0041] Further, in step S130, the bias field prediction model may be first trained using the first training data set. The original image samples in the first training data set may be used as input data of the bias field prediction model, and the bias field samples in the first training data set may be used as training targets of the bias field prediction model to perform first training on the bias field prediction model. In other embodiments, the bias field prediction model may include, for example, at least one of a variational autoencoder (VAE) model, a generative adversarial network (GAN) and its variants (e.g., a conditional generative adversarial network CGAN, a deep convolutional generative adversarial network DCGAN, a semi-supervised generative adversarial network SGAN, etc.).

[0042] In a preferred embodiment, the bias field prediction model may be a VAE model. Compared with other types of models, the VAE model has a better training effect because it has a model structure of an encoder and a decoder. Figure 2 Give a description. Figure 2 Figure 2 shows a schematic structural block diagram of the VAE model. Figure 2 As shown in , the VAE model may include an encoder 201 and a decoder 202. In this embodiment, the encoder 201 of the VAE model may encode the input original image sample into a mean parameter and a variance parameter in a latent space, and sample a latent vector from the latent state distribution (or latent representation) of the mean parameter and the variance parameter. The latent vector is used as the input of the decoder 202, and the bias field information output by the decoder 202 can be obtained. In some embodiments, the output of the decoder 202 may be the bias field image itself predicted by the bias field prediction model. In other embodiments, the VAE model may include two encoders, one of which may be used to encode to obtain the mean parameter, and the other may be used to encode to obtain the variance parameter.

[0043] During the training of the bias field prediction model, the training can be performed by minimizing the reconstruction error and / or the KL divergence of the latent representation. By maximizing the similarity between the training target and the decoded image, while minimizing the difference between the latent representation and the standard normal distribution, the bias field prediction model can learn the latent structure of the original image sample, thereby better understanding the bias field characteristics and for correcting the bias field.

[0044] Return below Figure 1 Continue to describe. In some embodiments, the training method 100 may further include: downsampling the samples in the first training data set so as to use the downsampled first training data set to perform a first training on the bias field prediction model. In some embodiments, the samples in the first training data set may include original image samples and bias field samples. In other embodiments, the sizes of each dimension of the downsampled samples may be the same or different. For example, in some embodiments, the samples in the first training data set may be downsampled to a size of, for example, 96*96*96. Of course, it is not limited to this, and other required downsampling sizes (i.e., the size of the downsampled samples) may also be set as needed. The sizes of each dimension in the downsampling size may also be set differently. The bias field prediction model is first trained using the downsampled first training data set, and the size of the output (i.e., bias field information) of the bias field prediction model after the first training is the same as the downsampled size of the samples in the first training data set. According to such a setting, it is beneficial to reduce the amount of training data, thereby reducing the size of the video memory occupied during the training process, thereby reducing the requirements for hardware (graphics card memory), and improving the training speed and training efficiency.

[0045] Combination of the above Figure 1 The training method of the bias field prediction model for predicting image bias field information according to the disclosed embodiment is described exemplarily. It is to be understood that the above description is exemplarily and not restrictive. For example, in step S120, the construction of the first training data set based on the original image samples and the bias field samples may not be limited to directly using the original image samples and the bias field samples as the first training data set, but may also process the bias field samples to obtain other bias field information, so that other bias field information related to the bias field samples may be used as the first training data set. For ease of understanding, the following will be combined with Figure 3 Provide detailed explanation.

[0046] Figure 3 FIG. 4 is an exemplary flow chart of a training method for performing first training using bias field inverse samples according to some embodiments of the present disclosure. Figure 3 As shown in , the training method 300 may include: in step S310, an original image sample and a corresponding bias field sample may be obtained. Step S310 may be combined with the above or Figure 1 The described step S110 is the same or similar and will not be repeated here.

[0047] Next, in step S320, the inverse of each element in the bias field sample can be taken to obtain a corresponding bias field inverse sample. Each element in the bias field sample can be a pixel value of each pixel. The bias field inverse sample obtained by taking the inverse operation corresponds to the original image sample corresponding to the bias field sample before taking the inverse. Then, in step S330, the original image sample and the corresponding bias field inverse sample can be combined into an image pair as a first training data set. The first training data set can include multiple such image pairs.

[0048] It is understandable that step S320 and step S330 may be a combination of the above. Figure 1 One implementation of step S120 described above is still to construct a first training data set based on the original image samples and the bias field samples. After obtaining the first training data set, step S340 will be performed below. In step S340, the first training data set can be used to perform a first training on the bias field prediction model. The bias field prediction model after the first training outputs a predicted bias field reciprocal image.

[0049] Combination of the above Figure 3 An exemplary description is given of the training method for performing the first training using the inverse samples of the bias field according to the embodiment of the present disclosure. It should be noted that the training method 300 is a preferred implementation of the present disclosure. By using the inverse samples of the bias field for training, the bias field prediction model in the training process is easier to converge, which is beneficial to reducing the difficulty of training and speeding up the training speed. At the same time, better training effects can be obtained, and the prediction results of the bias field prediction model after training are more accurate.

[0050] It can also be understood that the above description is exemplary rather than restrictive. For example, in some embodiments, a first training data set including an image pair of original image samples and corresponding bias field inverse samples can be downsampled, and the downsampled first training data set is used to perform a first training on the bias field prediction model. The bias field prediction model after the first training outputs a bias field inverse image of the downsampled size. By combining the bias field inverse training data with the downsampling operation, the model training speed can be further accelerated and the training efficiency can be improved.

[0051] For example, the training method of some embodiments of the present disclosure may not be limited to using only the aforementioned first training data set, but may also use other training data sets for training. The training method of some embodiments of the present disclosure may not be limited to including only the first training, but may also include other training processes to further optimize the bias field prediction model. Figure 4 Provide further explanation.

[0052] Figure 4An exemplary flow chart of a training method including a second training according to some embodiments of the present disclosure is shown. Figure 4 As shown in , the training method 400 may include: in step S410, an original image sample and a corresponding bias field sample may be obtained. In step S420, a first training data set may be constructed based on the original image sample and the bias field sample. In step S430, the first training data set may be used to perform a first training on the bias field prediction model. Steps S410 to S430 have been described in conjunction with the above. Figure 1 - Figure 3 The detailed description is given above and will not be repeated here. Step S440 and step S450 will be described below.

[0053] In step S440, the first training data set may be amplified using a diffusion model to obtain a second training data set. Diffusion Models is a generative model that simulates a Markov chain of data generation, generating data by gradually introducing noise and removing the noise in an inverse process. In some embodiments, the diffusion model may include one or more of Gaussian Diffusion Models, Non-Markov Diffusion Models, Fractional Brownian Motion (fBm) Models, Generative Diffusion Models, Conditional Diffusion Models, Iterative Diffusion Models, Adaptive Diffusion Models, etc.

[0054] The amplification of the first training data set described above may be to add new training data on the basis of the first training data set to form a second training data set. In some embodiments, the second training data set may be obtained by inputting samples in the first training data set into a diffusion model to generate new sample data.

[0055] In some other embodiments, the bias field prediction model includes an encoder and a decoder, and step S440 may include: using the latent vector output by the encoder of the bias field prediction model after the first training to perform a third training on the diffusion model; using the diffusion model after the third training to add noise to the latent vector to generate a new latent vector; inputting the new latent vector into the decoder to output new bias field information; and amplifying the first training data set based on a new image pair consisting of the new bias field information and the corresponding original image samples to obtain a second training data set.

[0056] Take the bias field prediction model as a VAE model as an example, Figure 2 As shown in , after the first training of the VAE model, the original image samples in the first training data set can be used to input into the encoder 201 of the first trained VAE model, and the latent vector output by the encoder 201 after the first training is used as the training data for the third training of a diffusion model. Specifically, the latent vector can be input into the diffusion model, and the third training of the diffusion model is achieved by adding noise and denoising the latent vector. Then, the noise introduced by the third trained diffusion model is sampled, and the latent vector output by the encoder 201 after the first training is denoised to generate a new latent vector after adding noise.

[0057] Afterwards, the generated new latent vector can be input to the decoder 202 to obtain the new bias field information output by the decoder 202. In some embodiments, when the first training data set includes image pairs of original image samples and bias field samples, the bias field prediction model after the first training outputs the bias field image itself, and the new bias field information here is the new bias field sample. In other embodiments, when the first training data set includes image pairs of original image samples and bias field inverse samples, the bias field prediction model after the first training outputs the bias field inverse image, and the new bias field information here is the new bias field inverse sample.

[0058] Furthermore, a new image pair consisting of the new bias field information and the corresponding original image sample can be added to the first training data set, and together with the image pair consisting of the original image sample and the corresponding bias field information (bias field sample or bias field inverse sample) in the original first training data set, they can be used as training data in the second training data set. The original image sample corresponding to the new bias field information in the new image pair is the original image sample corresponding to the latent vector before noise addition.

[0059] Return below Figure 4 Continue to describe. Figure 4As further shown in FIG. 4 , in step S450, the second training data set may be used to perform a second training on the bias field prediction model after the first training. In the second training, the original image samples of each image pair / new image pair in the second training data set may be used as the input of the bias field prediction model after the first training, and the bias field information / new bias field information corresponding to the original image sample in the second training data set may be used as the target output of the bias field prediction model to perform a second training on the bias field prediction model. The second training may be an optimization and fine-tuning of the bias field prediction model after the first training.

[0060] By combining the bias field prediction model with the diffusion model as mentioned above, the characteristics and advantages of the two different models are creatively utilized. The diffusion model is used to augment the latent vectors generated by the VAE model, which can generate more diverse and realistic latent vectors, thereby helping to improve the generalization and representation capabilities of the bias field prediction model, enabling the bias field prediction model to better capture the potential structure of the data and further improve the effect of bias field correction.

[0061] Combination of the above Figure 4 The training method including the second training according to the embodiment of the present disclosure is described exemplarily. It can be understood that the above description is exemplary and not restrictive. For example, in some embodiments, the training method 400 may further include: downsampling the samples in the second training data set so as to use the downsampled second training data set to perform the second training on the bias field prediction model. According to such a setting, it is helpful to reduce the amount of data for the second training and the occupation of the video memory, thereby improving the training speed and training efficiency.

[0062] After obtaining the trained bias field prediction model, the present disclosure also provides a method for correcting the image bias field. Figure 5 An exemplary description is given.

[0063] Figure 5 FIG. 4 is a flow chart showing a method for correcting an image bias field according to some embodiments of the present disclosure. Figure 5 As shown in the method 500, the method 500 may include: in step S510, the image to be corrected may be input into the image to be corrected according to the present disclosure in the foregoing Figure 1 - Figure 4In the bias field prediction model trained by any of the training methods, the bias field information of the image to be corrected is output. In some embodiments, the image to be corrected may be a medical image. In other embodiments, the image to be corrected may include an MRI image. In other embodiments, the bias field information may include the bias field image itself, or the bias field reciprocal image. It is understandable that the bias field information output by the bias field prediction model is related to the content of the image pairs in the first training data set and / or the second training data set used above, and will not be repeated here.

[0064] Then, the process may proceed to step S520, and a corrected image may be obtained based on the image to be corrected and the corresponding bias field information. In some embodiments, when the bias field information is a bias field image, obtaining the corrected image based on the image to be corrected and the corresponding bias field information may include: dividing the image to be corrected by corresponding elements in the bias field image to obtain the corrected image; or when the bias field information is a bias field reciprocal image, obtaining the corrected image based on the image to be corrected and the corresponding bias field information may include: multiplying the image to be corrected by corresponding elements in the bias field reciprocal image to obtain the corrected image.

[0065] In other embodiments, when downsampled training data is used in the training process of the bias field prediction model, method 500 may further include: downsampling the image to be corrected; inputting the downsampled image to be corrected into the bias field prediction model to output bias field information of the downsampled size; and upsampling the bias field information of the downsampled size to the same size as the image to be corrected, so as to obtain a corrected image of the same size as the image to be corrected after calculation with the image to be corrected.

[0066] For example, the image to be corrected may be downsampled to the same downsampled size as the first training data set and / or the second training data set (for example, downsampled to a size of 96*96*96), and then the image to be corrected of the downsampled size is input into the trained bias field prediction model, and the bias field prediction model outputs the bias field information of the downsampled size (for example, a size of 96*96*96). Afterwards, the bias field information of the downsampled size may be upsampled to the same size as the image to be corrected (that is, the original size of the image to be corrected) through interpolation (for example, one or more of bilinear interpolation, trilinear interpolation, nearest neighbor interpolation, etc.), so as to obtain the corrected image after performing the above-mentioned calculation of corresponding element division or multiplication with the image to be corrected.

[0067] In summary, the training method of the disclosed embodiment predicts the bias field information of the image by using a bias field prediction model based on deep learning. The bias field prediction model can automatically learn and extract the bias field information in the training data, and is not limited by prior experience or assumptions. Therefore, even in the case of uneven magnetic field distribution or complex tissue structure, the bias field prediction can be accurately performed. Therefore, when performing subsequent bias field correction and other operations based on the prediction results, more accurate correction results can be obtained, which is conducive to improving the quality of the image. Compared with existing mathematical models and statistical methods, the scheme of the disclosed embodiment also has the advantages of simpler calculation process, shorter time consumption, more accurate correction results, etc., and is more suitable for image analysis, clinical applications, etc.

[0068] Furthermore, in some embodiments, the training data is amplified by combining the diffusion model, especially by amplifying the latent vector, and the bias field prediction model is trained twice, so that the performance of the bias field prediction model after the secondary training is significantly optimized, especially the generalization ability of the bias field prediction model is improved, so that it can be applied to bias field correction of images from different sources, or different types of images from the same source (for example, different types of MRI images), thereby having a broader application prospect.

[0069] In other preferred embodiments, the bias field prediction model is trained by using the bias field reciprocal samples, so that the training process of the bias field prediction model is easier to converge, that is, the training process is easier, the training speed is faster, and it is easier to obtain better training results. This makes it possible to obtain a model suitable for the required scenario more quickly when the bias field prediction model is used for training data from different sources or different types, so that it has a broader application prospect and the ability to adapt to the needs of rapid information changes.

[0070] The present disclosure also provides a device for correcting an image bias field, comprising: a processor for executing program instructions; and a memory storing the program instructions, wherein when the program instructions are loaded and executed by the processor, the processor executes the method described in the foregoing combined method according to the present disclosure. Figure 1 - Figure 4 Any of the training methods or implementations disclosed herein in combination with Figure 5 The following will combine Figure 6 The system shown is described.

[0071] Figure 6The schematic block diagram of the system for correcting the image bias field according to the embodiment of the present disclosure is shown. The system 600 may include a device 601 according to the embodiment of the present disclosure and its peripheral devices and external network, wherein the device 601 is used to train the bias field prediction model or to perform bias field correction on the image to be corrected, so as to achieve the above-mentioned combination Figure 1 - Figure 5 The technical solution of any of the embodiments of the present disclosure.

[0072] like Figure 6 As shown in , the device 601 may include a CPU 6011, which may be a general-purpose CPU, a dedicated CPU, or other information processing and program execution unit. Further, the device 601 may also include a large-capacity memory 6012 and a read-only memory ROM 6013, wherein the large-capacity memory 6012 may be configured to store various types of data, including training data sets, bias field information, images to be corrected and correction results, etc., as well as various programs required for running the neural network, and the ROM 6013 may be configured to store the power-on self-test of the device 601, the initialization of each functional module in the system, the basic input / output driver of the system, and the data required for booting the operating system.

[0073] Further, the device 601 also includes other hardware platforms or components, such as the TPU 6014, GPU 6015, FPGA 6016 and MLU 6017 shown. It is understood that although a variety of hardware platforms or components are shown in the device 601, this is only exemplary and not restrictive, and those skilled in the art can add or remove corresponding hardware according to actual needs. For example, the device 601 may include only a CPU as a well-known hardware platform and another hardware platform as a test hardware platform disclosed herein.

[0074] The device 601 of the present disclosure also includes a communication interface 6018, so that it can be connected to a local area network / wireless local area network (LAN / WLAN) 605 through the communication interface 6018, and then connected to a local server 606 or to the Internet ("Internet") 607 through the LAN / WLAN. Alternatively or additionally, the device 601 of the present disclosure can also be directly connected to the Internet or a cellular network through the communication interface 6018 based on wireless communication technology, such as wireless communication technology based on the third generation ("3G"), the fourth generation ("4G") or the fifth generation ("5G") . In some application scenarios, the device 601 of the present disclosure can also access a server 608 of an external network and a possible database 609 as needed to obtain various known original image samples and their corresponding bias field samples, etc., and can remotely store various parameters or intermediate data.

[0075] The peripheral devices of the device 601 may include a display device 602, an input device 603, and a data transmission interface 604. In one embodiment, the display device 602 may include, for example, one or more speakers and / or one or more visual displays, which are configured to perform voice prompts and / or image video display on the computing process or correction results of the disclosed device. The input device 603 may include, for example, a keyboard, a mouse, a microphone, a gesture capture camera, or other input buttons or controls, which are configured to receive input or user instructions of the original image sample or the image to be corrected. The data transmission interface 604 may include, for example, a serial interface, a parallel interface or a universal serial bus interface ("USB"), a small computer system interface ("SCSI"), a serial ATA, a firewire ("FireWire"), a PCI Express, and a high-definition multimedia interface ("HDMI"), etc., which are configured for data transmission and interaction with other devices or systems. According to the scheme disclosed herein, the data transmission interface 604 can receive a training data set for training, or an image to be corrected, etc., and transmit various types of data and results to the device 601.

[0076] The CPU 6011, mass storage 6012, read-only memory ROM 6013, TPU 6014, GPU 6015, FPGA 6016, MLU 6017 and communication interface 6018 of the device 601 of the present disclosure can be interconnected through a bus 6019, and data interaction with peripheral devices can be achieved through the bus. In one embodiment, through the bus 6019, the CPU 6011 can control other hardware components in the device 601 and its peripheral devices.

[0077] In operation, the processor CPU 6011 of the device 601 disclosed herein can receive a training data set or an image to be corrected for training through an input device 603 or a data transmission interface 604, and retrieve computer program instructions or codes (such as codes involving a neural network) stored in a memory 6012 to train the received training data or correct the image to be corrected, so as to obtain the parameters of the trained bias field prediction model or the bias field information results. After the CPU 6011 determines the bias field information or the correction result by executing the program instructions, the bias field information or the correction result can be displayed on the display device 602 or output by voice prompts. In addition, the device 601 can also upload the bias field information or the correction result to a network, such as a remote database 609, through a communication interface 6018.

[0078] It should also be understood that any module, unit, component, server, computer, terminal or device that executes instructions of the disclosed examples may include or otherwise access computer-readable media, such as storage media, computer storage media or data storage devices (removable and / or non-removable) such as disks, optical disks or tapes. Computer storage media may include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, program modules or other data.

[0079] Based on the above, the present disclosure further provides a computer-readable storage medium having computer-readable instructions stored thereon, which, when executed by one or more processors, implements the above-mentioned Figure 1 - Figure 4 The training method described in any embodiment of the present invention or as described above in combination with Figure 5 The method for correcting image bias field is described.

[0080] The computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium, such as RRAM (Resistive Random Access Memory), DRAM (Dynamic Random Access Memory), SRAM (Static Random-Access Memory), EDRAM (Enhanced Dynamic Random Access Memory), HBM (High-Bandwidth Memory), HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by the application, module, or both. Any such computer storage medium may be part of the device or accessible or connectable to the device. Any application or module described in the present invention may be implemented using computer-readable / executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

[0081] Although multiple embodiments of the present disclosure have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art may think of many changes, modifications, and alternatives without departing from the thought and spirit of the present disclosure. It should be understood that in the process of practicing the present disclosure, various alternatives to the embodiments of the present disclosure described herein may be adopted. The attached claims are intended to define the scope of protection of the present disclosure, and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for training a bias field prediction model for predicting image bias field information, comprising: Obtaining original image samples and corresponding bias field samples; Constructing a first training data set based on the original image samples and the bias field samples; as well as The bias field prediction model is first trained using the first training data set.

2. The training method according to claim 1, wherein obtaining original image samples and corresponding bias field samples comprises: Acquire the original image sample and a corrected image sample of the original image sample; The original image samples are divided by corresponding elements in the corrected image samples to obtain corresponding bias field samples.

3. The training method according to claim 1 or 2, wherein constructing a first training data set based on the original image samples and the bias field samples comprises: The original image samples and the corresponding bias field samples are combined into image pairs to serve as the first training data set.

4. The training method according to claim 1 or 2, wherein constructing a first training data set based on the original image samples and the bias field samples comprises: Taking the reciprocal of each element in the bias field sample to obtain a corresponding bias field reciprocal sample; The original image samples and the corresponding bias field inverse samples are combined into image pairs to serve as the first training data set.

5. The training method according to any one of claims 1 to 4, further comprising: The samples in the first training data set are downsampled so as to use the downsampled first training data set to perform a first training on the bias field prediction model.

6. The training method according to any one of claims 1 to 5, further comprising: Amplifying the first training data set using a diffusion model to obtain a second training data set; as well as The second training data set is used to perform a second training on the bias field prediction model after the first training.

7. The training method according to claim 6, wherein the bias field prediction model comprises an encoder and a decoder, and the first training data set is amplified using a diffusion model to obtain a second training data set comprising: Performing a third training on the diffusion model using the latent vector output by the encoder of the bias field prediction model after the first training; Noising the latent vector using a third trained diffusion model to generate a new latent vector; Inputting the new latent vector into the decoder to output new bias field information; as well as Based on a new image pair consisting of the new bias field information and corresponding original image samples, the first training data set is augmented to obtain the second training data set.

8. The training method according to claim 6 or 7, further comprising: The samples in the second training data set are downsampled so as to use the downsampled second training data set to perform the second training on the bias field prediction model.

9. The training method according to any one of claims 1 to 8, wherein The bias field prediction model includes at least one of a VAE model, a generative adversarial network and variants thereof.

10. The training method according to any one of claims 1-9, wherein the original image samples include MRI image samples.

11. A method for correcting an image bias field, comprising: Inputting the image to be corrected into the bias field prediction model trained by the training method according to any one of claims 1 to 10 to output the bias field information of the image to be corrected; as well as A corrected image is obtained based on the image to be corrected and corresponding bias field information.

12. The method according to claim 11, wherein The bias field information is a bias field image, and obtaining a corrected image based on the image to be corrected and the corresponding bias field information includes: Dividing the image to be corrected by corresponding elements in the bias field image to obtain the corrected image; or If the bias field information is a bias field reciprocal image, obtaining the corrected image based on the image to be corrected and the corresponding bias field information includes: multiplying the image to be corrected with corresponding elements in the bias field reciprocal image to obtain the corrected image.

13. The method according to claim 11 or 12, further comprising: Downsampling the image to be corrected; Inputting the downsampled image to be corrected into the bias field prediction model to output bias field information of the downsampled size; as well as The bias field information of the downsampled size is upsampled to the same size as the image to be corrected, so that after calculation with the image to be corrected, a corrected image of the same size as the image to be corrected is obtained.

14. The method according to any one of claims 11 to 13, wherein the image to be corrected comprises an MRI image.

15. A device for correcting an image bias field, comprising: a processor for executing program instructions; as well as A memory storing the program instructions, which, when loaded and executed by the processor, enables the processor to execute the training method according to any one of claims 1-10 or the method according to any one of claims 11-14.

16. A computer-readable storage medium having computer-readable instructions stored thereon, which, when executed by one or more processors, implement the training method as described in any one of claims 1-10 or the method as described in any one of claims 11-14.

Images