Model training method and related device

Abstract

The invention discloses a model training method and a related device, and relates to the field of machine learning, and the method comprises the steps: obtaining a clean coil image of each sample coil of each sample cutting layer; and generating a bad characteristic coil image of the sample coil based on the bad characteristic component of the sample coil and the clean coil image. And generating a bad feature amplitude image of the sample cutting layer based on the bad feature coil image of each sample coil of the sample cutting layer. And acquiring a simulation coil image of each sample coil based on the bad characteristic amplitude image of the sample cutting layer and the bad characteristic coil image of each sample coil. And based on the training data, training a machine learning model, and if a training completion condition is reached, obtaining a bad feature removal model. The sample comprises a target complex component of a bad characteristic coil image of the sample coil and a target complex component of a simulation coil image, the label is a target complex component of a clean coil image of the sample coil, and the target complex component is one of a real part or an imaginary part.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a model training method and related devices. Background Art

[0002] Currently, machine learning models are trained to identify various undesirable features in MRI (Magnetic Resonance Imaging) images. These features are then removed from the images to improve MRI image quality. Therefore, improving the training effectiveness of machine learning models is an important technical approach to improving MRI image quality. Summary of the Invention

[0003] In view of the above problems, this application provides a model training method and related devices to achieve the purpose of improving the effect of the bad feature removal model. The specific solution is as follows:

[0004] A first aspect of the present application provides a model training method, comprising multiple iterative trainings: the iterative training comprises: acquiring a clean coil image of each sample coil of each sample slice;

[0005] For each of the sample coils, generating a bad characteristic coil image of the sample coil based on the bad characteristic component of the sample coil and a clean coil image;

[0006] generating a bad characteristic amplitude image of the sample slice based on the bad characteristic coil image of each sample coil of the sample slice;

[0007] acquiring a simulated coil image of each of the sample coils based on the defective characteristic amplitude image of the sample slice and the defective characteristic coil image of each of the sample coils;

[0008] Based on the training data, the preset machine learning model is trained. If the training completion conditions are met, a bad feature removal model is obtained;

[0009] The training data includes multiple samples and corresponding labels. The samples include a target complex component of a bad characteristic coil image of the sample coil and a target complex component of a simulated coil image. The label is a target complex component of a clean coil image of the sample coil. The target complex component is one of the real part and the imaginary part.

[0010] In a possible implementation, if the iterative training is the first iterative training, obtaining a clean coil image of each sample coil in each sample slice includes:

[0011] Scanning the sample object multiple times, and obtaining multiple original coil images of each sample coil in each sample slice based on the sample coil data of the sample object;

[0012] For each of the sample coils, calculating an arithmetic mean of all original coil images of the sample coil to obtain an average coil image of the sample coil;

[0013] reconstructing an amplitude image of the sample slice based on an average coil image of each of the sample coils corresponding to the sample slice;

[0014] A clean coil image of each of the sample coils in the sample slice is acquired based on the amplitude image of the sample slice and the sensitivity map and phase map of each of the sample coils.

[0015] In a possible implementation, acquiring a clean coil image of each sample coil of the sample slice based on the amplitude image of the sample slice and the sensitivity map and phase map of each sample coil includes:

[0016] performing smoothing processing on the amplitude image of the sample slice and the amplitude image of each sample coil respectively;

[0017] calculating a ratio of the smoothed amplitude image of the sample coil to the amplitude image of the sample slice to obtain a sensitivity map of the sample coil;

[0018] Acquiring an original phase image of the sample coil using a preset angle acquisition function;

[0019] Smoothing the original phase image of the sample coil to obtain a smoothed phase image of the sample coil;

[0020] Normalizing the smoothed phase image of the sample coil to obtain the phase image of the target coil;

[0021] Based on the amplitude image of the sample slice and the sensitivity map and phase map of each sample coil, a clean coil image of each sample coil in the sample slice is obtained, where the real part of the clean coil image of the sample coil is the product of the amplitude image of the sample slice, the sensitivity map of the sample coil, and the cosine value of the phase map, and the imaginary part is the product of the amplitude image of the sample slice, the sensitivity map of the sample coil, and the sine value of the phase map.

[0022] In a possible implementation, acquiring a simulated coil image of each of the sample coils based on the bad characteristic amplitude image of the sample slice and the bad characteristic coil image of each of the sample coils includes:

[0023] Calculating a sensitivity map and a phase map based on the smoothed bad characteristic amplitude image of the sample slice and the bad characteristic coil image of each of the sample coils;

[0024] For each sample slice, a simulated coil image of each sample coil is acquired based on the defective characteristic amplitude image of the sample slice and the sensitivity map and phase map of each sample coil. The real part of the simulated coil image of the sample coil is the product of the defective characteristic amplitude image of the sample slice, the sensitivity map of the sample coil, and the cosine value of the phase map. The imaginary part is the product of the defective characteristic amplitude image of the sample slice, the sensitivity map of the sample coil, and the sine value of the phase map.

[0025] In a possible implementation, if the iterative training is any target iterative training except the first iterative training, generating, for each sample coil, a bad characteristic coil image of the sample coil based on the bad characteristic component of the sample coil and a clean coil image includes:

[0026] Determine whether the training completion conditions are met;

[0027] If not, constructing a bad feature component of each sample coil of each sample slice as a new bad feature component;

[0028] For each sample coil, the target complex components of the original undesirable characteristic coil image of the sample coil are updated based on the target complex components of the original undesirable characteristic components of the sample coil and the target complex components of the new undesirable characteristic components to obtain the target complex components of the new undesirable characteristic coil image of the sample coil.

[0029] In a possible implementation, generating the bad characteristic amplitude image of the sample slice based on the bad characteristic coil image of each sample coil of the sample slice includes:

[0030] For each of the sample slices, determining a local update set of the sample slice; wherein the local update set includes a plurality of target sample coils;

[0031] Based on the new bad characteristic coil image of each target sample coil in the local update set, the original bad characteristic amplitude image of the sample slice is updated to obtain a new bad characteristic amplitude image of the sample slice.

[0032] A second aspect of the present application provides a model training device, characterized in that it includes multiple training units for performing iterative training; the iterative training units include:

[0033] a clean image acquisition unit, configured to acquire a clean coil image of each sample coil of each sample slice;

[0034] a defective coil image acquiring unit, configured to generate, for each of the sample coils, a defective characteristic coil image of the sample coil based on the defective characteristic component of the sample coil and a clean coil image;

[0035] a bad amplitude image acquiring unit, configured to generate a bad characteristic amplitude image of the sample slice based on the bad characteristic coil images of each of the sample coils of the sample slice;

[0036] a simulated coil image acquisition unit, configured to acquire a simulated coil image of each of the sample coils based on the bad characteristic amplitude image of the sample slice and the bad characteristic coil image of each of the sample coils;

[0037] The iteration unit is used to train the preset machine learning model based on the training data. If the training completion conditions are met, a bad feature removal model is obtained;

[0038] The training data includes multiple samples and corresponding labels. The samples include a target complex component of a bad characteristic coil image of the sample coil and a target complex component of a simulated coil image. The label is a target complex component of a clean coil image of the sample coil. The target complex component is one of the real part and the imaginary part.

[0039] The third aspect of the present application provides a computer program product, including computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements the model training method of the above-mentioned first aspect or any implementation of the first aspect.

[0040] A fourth aspect of the present application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0041] The memory is used to store computer programs;

[0042] The processor is used to execute the computer program so that the electronic device can implement the model training method of the above-mentioned first aspect or any implementation method of the first aspect.

[0043] In a fifth aspect, the present application provides a computer storage medium, which carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can use the model training method of the above-mentioned first aspect or any implementation of the first aspect.

[0044] By means of the above technical solution, the model training and related devices provided by the present application include obtaining clean coil images of each sample coil of each sample slice. For each sample coil, a bad feature coil image of the sample coil is generated based on the bad feature component and the clean coil image of the sample coil. A bad feature amplitude image of the sample slice is generated based on the bad feature coil images of each sample coil of the sample slice. Based on the bad feature amplitude image of the sample slice and the bad feature coil images of each sample coil, a simulated coil image of each sample coil is obtained. Based on the training data, a machine learning model is trained. If the training completion condition is met, a bad feature removal model is obtained. The training data includes multiple samples and corresponding labels. The samples include the target complex component of the bad feature coil image of the sample coil and the target complex component of the simulated coil image. The label is the target complex component of the clean coil image of the sample coil. The target complex component is one of the real part or the imaginary part. It can be seen that in each iterative training of this method, the target complex components of the bad feature coil image of the sample coil and the target complex components of the simulated coil image are used as sample images, and the target complex components of the clean coil image of the sample coil are used as labels, wherein the simulated coil image is simulated based on the bad feature amplitude image of the sample slice and the bad feature coil image of each sample coil. Therefore, the machine learning model can simultaneously utilize the low proportion of bad features in the amplitude image and the simple distribution characteristics of the bad features in the coil image to construct training data to execute each iteration, thereby improving the model performance of the bad feature removal model to achieve the accuracy of removing bad features. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and other features, advantages, and aspects of the various embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that the originals and elements are not necessarily drawn to scale.

[0046] Figure 1 A schematic diagram of the system architecture provided for this application;

[0047] Figure 2 A flowchart of a model training method provided in an embodiment of the present application;

[0048] Figure 3 A flowchart for the specific implementation of a model training method provided in an embodiment of the present application;

[0049] Figure 4 A flowchart of another model training method provided in an embodiment of the present application;

[0050] Figure 5 A schematic diagram of the structure of a model training device provided in an embodiment of the present application;

[0051] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] The following describes the embodiments of the present application in conjunction with the accompanying drawings. The terms used in the implementation methods of the present application are only used to explain the specific embodiments of the present application and are not intended to limit the present application.

[0053] The embodiments of the present application are described below in conjunction with the accompanying drawings. Those skilled in the art will appreciate that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0054] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, and this is merely a way of distinguishing the objects of the same attributes when describing them in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0055] The present application can be applied in the field of machine learning technology, and specifically can be applied to a method of training a pre-built machine learning model to obtain a bad feature removal model in a scenario where a machine learning model is used to remove bad features from MRI (Magnetic Resonance Imaging) images.

[0056] This application can be applied to, but is not limited to, applications with model training capabilities or cloud services provided by cloud-side servers. The following describes each of these:

[0057] See also Figure 1 , Figure 1 A schematic diagram of a system architecture is shown. The system may include a terminal 100 and a server 200. The server 200 may provide the method provided in the embodiment of the present application to one or more terminals.

[0058] Among them, a model training application can be installed on the terminal 100. The above application and web page can provide an interface. The terminal 100 can receive relevant parameters entered by the user on the model training interface and send the above parameters to the server 200. The server 200 can obtain processing results based on the received parameters and return the processing results to the terminal 100.

[0059] It should be understood that in some optional implementations, the terminal 100 can also complete the action of obtaining the processing result based on the received parameters by itself without the need for the cooperation of the server, and the embodiments of the present application are not limited to this.

[0060] Next describe Figure 1 The product form of the mid-terminal 100;

[0061] The terminal 100 in the embodiment of the present application can be a mobile phone, a tablet computer, a wearable device, an in-vehicle device, an augmented reality (AR) / virtual reality (VR) device, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), etc., and the embodiment of the present application does not impose any restrictions on this.

[0062] The terminal 100 may include components such as a radio frequency unit, a memory, an input unit, a display unit, a camera (optional), an audio circuit (optional), a speaker (optional), a microphone (optional), a headphone jack (optional), a processor, an external interface, and a power supply. Those skilled in the art will appreciate that the aforementioned components are merely examples and do not limit the terminal or multi-function device, and may include more or fewer components, or a combination of certain components, or different components.

[0063] The input unit can be used to receive input digital or character information and generate key signal input related to user settings and function control of the portable multi-function device. Specifically, the input unit may include a touch screen (optional) and / or other input devices. Specifically, other input devices may include, but are not limited to, one or more of a physical keyboard, function keys (such as volume control keys, power keys, etc.), a trackball, a mouse, a joystick, etc.

[0064] Among them, the input device can receive input data and so on.

[0065] The display unit can be used to display information input by the user or information provided to the user, various menus of the terminal, interactive interface, file display and / or playback of any multimedia file. In an embodiment of the present application, the display unit can be used to display the interface of the model training process, processing results, etc.

[0066] Among them, the memory can be used to store software codes related to the model training method, the processor can execute the steps of the model training method, and can also schedule other units (such as the above-mentioned input unit and display unit) to implement corresponding functions.

[0067] The radio frequency unit (optional) can be used to send and receive information or receive and send signals during a call.

[0068] In this embodiment of the present application, the radio frequency unit can send data to the server 200 and receive the processing results sent by the server 200.

[0069] It should be understood that the radio frequency unit is optional and can be replaced by other communication interfaces, such as a network port.

[0070] The terminal 100 further includes a power source (such as a battery) for supplying power to various components.

[0071] The terminal 100 also includes an external interface, which may be a standard Micro USB interface or a multi-pin connector, and may be used to connect the terminal 100 to other devices for communication, or to connect a charger to charge the terminal 100.

[0072] The server 200 includes a bus, a processor, a communication interface, and a memory. The processor, the memory, and the communication interface communicate with each other via the bus.

[0073] Among them, the memory can be used to store software codes related to the model training method, the processor can execute the steps of the chip's model training method, and can also schedule other units to implement corresponding functions.

[0074] MRI images are constructed based on coil data from multiple coils of at least one slice. An MRI image reconstructed from coil data from multiple coils of a slice is a two-dimensional MRI image, and a three-dimensional MRI image is reconstructed from coil data from multiple coils of multiple slices. Specifically, an original coil image is generated based on coil data collected by each coil, and a two-dimensional MRI image or a three-dimensional MRI image is reconstructed based on multiple original coil images of at least one slice. Because coil data are complex signals, the original coil images generated from the complex signals are all complex images that include real and imaginary images. Undesirable features in complex images often have the advantages of simple feature distribution and low difficulty in constructing sample data. However, complex images contain a high degree of undesirable features, which is not conducive to their removal. Compared to complex images, amplitude images are generated from multiple complex images and contain a lower degree of undesirable features. However, the distribution of undesirable features is complex.

[0075] Undesirable features consist of noise and / or artifacts. For example, noise in complex images follows a Gaussian distribution. Reconstructed MRI images, such as two-dimensional MRI images or three-dimensional MRI images, follow a Rayleigh distribution. The properties of the Gaussian distribution indicate that predicting noise from noisy images is straightforward, and preparing a pair of clean and noisy images during machine learning training is also straightforward. However, because the signal-to-noise ratio of the original coil image is often extremely low, and the true structure is hidden within the Gaussian noise, predicting the clean image or noise component is ineffective. In other words, MRI noise exhibits different properties in amplitude images and complex images from the receiving coil. In amplitude images, noise follows a Rayleigh distribution, meaning it has both DC and AC components. In contrast, in coil images, noise follows a zero-mean Gaussian distribution. Therefore, for amplitude images, both AC and DC noise must be eliminated to effectively denoise and improve their quality.

[0076] In the existing technology, the denoising effect of using machine learning models to identify amplitude images with complex and undesirable features or to identify clean images in coil images with a very high proportion of undesirable features needs to be improved.

[0077] To address the above issues, the present invention provides a model training method and related apparatus, which aims to improve the effectiveness of machine learning models in removing undesirable features and thereby identifying clean images. The model training method of the present invention is described in detail below with reference to the accompanying drawings.

[0078] Reference Figure 2 , Figure 2 A flow chart of a model training method provided in an embodiment of the present application is shown as follows: Figure 2As shown, an embodiment of the present application provides a model training method, which includes a first iterative training and multiple target iterative trainings, each of which includes:

[0079] S201 : Acquire a clean coil image of each sample coil of each sample slice.

[0080] In this embodiment, based on the simple distribution of undesirable coil image features and their near-zero mean, at the beginning of the first iteration of training, clean coil images of each sample coil in the sample slice are acquired based on the sample coil data obtained from multiple scans of the sample object. During target iteration training, clean coil images of the sample coils in the original training data are simply acquired.

[0081] In an optional embodiment, a specific method for obtaining a clean coil image of each sample coil of a sample slice is as follows:

[0082] Based on the sample coil data of the sample object, multiple raw coil images are obtained for each sample coil in each sample slice. For each sample coil, the arithmetic mean of all raw coil images of the sample coil is calculated to obtain an average coil image of the sample coil. Based on the average coil images of each sample coil corresponding to the sample slice, an amplitude image of the sample slice is reconstructed. Based on the amplitude image of the sample slice and the sensitivity map and phase map of each sample coil, a clean coil image of each sample coil in the sample slice is obtained.

[0083] S202 : For each sample coil, generate a bad characteristic coil image of the sample coil based on the bad characteristic component of the sample coil and a clean coil image.

[0084] In this embodiment, the bad characteristic coil image is obtained by adding bad characteristic components to the clean coil image.

[0085] In an optional embodiment, in the first iterative training, the original bad feature component based on the sample coil is added to the clean coil image to obtain the original bad feature coil image.

[0086] In an optional embodiment, in the target iterative training, the original bad feature coil image is updated by the original bad feature component and the constructed new bad feature component.

[0087] Specifically, for each sample coil of each sample slice, an undesirable feature component of the sample coil is constructed as a new undesirable feature component.

[0088] For each sample coil, the target complex components of the original bad characteristic coil image of the sample coil are updated based on the target complex components of the original bad characteristic components of the sample coil and the target complex components of the new bad characteristic components to obtain the target complex components of the new bad characteristic coil image of the sample coil.

[0089] S203 : Generate a bad characteristic amplitude image of the sample slice based on the bad characteristic coil image of each sample coil of the sample slice.

[0090] In this embodiment, the bad characteristic amplitude image of the sample slice is reconstructed based on the bad characteristic coil image of each sample coil of the sample slice.

[0091] In an optional embodiment, in the first iterative training, after reconstructing the original bad feature amplitude image of the sample slice based on the original bad feature coil image of each sample coil of the sample slice based on the square sum method, the original bad feature amplitude image of the sample slice is stored for direct use in subsequent target iterative training.

[0092] In an optional embodiment, during the target iterative training, the bad characteristic amplitude image of the sample slice is updated by locally updating the bad characteristic coil image of the sample coil of the sample slice.

[0093] Specifically, for each sample slice, a local update set is determined for the sample slice. The local update set includes multiple target sample coils. Based on the new bad feature coil images of each target sample coil in the local update set, the original bad feature amplitude image of the sample slice is updated to obtain a new bad feature amplitude image of the sample slice.

[0094] S204 : Acquire a simulated coil image of each sample coil based on the defective characteristic amplitude image of the sample slice and the defective characteristic coil image of each sample coil.

[0095] In this embodiment, a simulated coil image of the sample coil is obtained by simulation based on the relationship between the defective characteristic amplitude image of the slice and the defective characteristic coil image of the sample coil.

[0096] In an optional embodiment, a specific method for obtaining a simulated coil image of each sample coil includes:

[0097] The sensitivity map and phase map are calculated based on the smoothed defect feature amplitude image of the sample slice and the defect feature coil image of each sample coil. For each sample slice, a simulated coil image of each sample coil is obtained based on the defect feature amplitude image of the sample slice and the sensitivity map and phase map of each sample coil. The real part of the simulated coil image of the sample coil is the product of the cosine value of the defect feature amplitude image of the sample slice, the sensitivity map of the sample coil, and the phase map, while the imaginary part is the product of the sine value of the defect feature amplitude image of the sample slice, the sensitivity map of the sample coil, and the phase map.

[0098] S205: Based on the training data, the machine learning model is trained. If the preset training conditions are met, a bad feature removal model is obtained.

[0099] The training data includes multiple samples and corresponding labels. The samples include the target complex component of the bad characteristic coil image of the sample coil and the target complex component of the simulated coil image. The label is the target complex component of the clean coil image of the sample coil. The target complex component is one of the real part or the imaginary part.

[0100] As can be seen from the above technical solution, an embodiment of the present application provides a model training method for obtaining clean coil images of each sample coil of each sample slice. For each sample coil, a bad feature coil image of the sample coil is generated based on the bad feature component and the clean coil image of the sample coil. A bad feature amplitude image of the sample slice is generated based on the bad feature coil images of each sample coil of the sample slice. Based on the bad feature amplitude image of the sample slice and the bad feature coil images of each sample coil, a simulated coil image of each sample coil is obtained. Based on the training data, a machine learning model is trained, and if the training completion condition is met, a bad feature removal model is obtained. The training data includes multiple samples and corresponding labels, the samples include the target complex component of the bad feature coil image of the sample coil and the target complex component of the simulated coil image, the label is the target complex component of the clean coil image of the sample coil, and the target complex component is one of the real part or the imaginary part. It can be seen that in each iterative training of this method, the target complex components of the bad feature coil image of the sample coil and the target complex components of the simulated coil image are used as sample images, and the target complex components of the clean coil image of the sample coil are used as labels, wherein the simulated coil image is simulated based on the bad feature amplitude image of the sample slice and the bad feature coil image of each sample coil. Therefore, the machine learning model can simultaneously utilize the low proportion of bad features in the amplitude image and the simple distribution characteristics of bad features in the coil image to construct training data to execute each iteration, thereby improving the model performance of the bad feature removal model to achieve the accuracy of removing bad features.

[0101] In one possible implementation, the model training method provided in the embodiments of the present application can be applied to the model training of a noise reduction model constructed based on a prediction-based machine learning model. That is, the undesirable feature removal model is a noise reduction model, which is used to remove image noise and obtain a clean coil image.

[0102] It should be noted that creating MRI magnitude images by combining coil data from all coils can overcome the low signal-to-noise ratio disadvantage of coil images. However, in MRI magnitude images, noise typically follows a Rayleigh distribution, meaning it typically contains both alternating current (AC) and direct current (DC) components. Therefore, effective denoising of the magnitude image is necessary to improve its quality. The noise reduction effects in coil and magnitude images are mutually influential. In magnitude images, the signal-to-noise ratio is higher, making it easier to predict the true fine structure in noisy images than in coil images. In contrast, the noise distribution in magnitude images follows a rather complex Rayleigh distribution. Because the offset introduced by the Rayleigh distribution depends on the noise level and the original contrast in the coil images, it is difficult to predict from the magnitude images. In coil images, the noise distribution follows a simpler normal distribution, making it easier to predict from noisy images or to prepare a pair of clean and noisy images for noise removal during machine learning training. In contrast, the signal-to-noise ratio of coil images is much lower, making it more difficult to predict the clean image or the noise component.

[0103] See also Figure 3 , Figure 3 A specific implementation flow chart of a model training method provided in an embodiment of the present application is as follows: Figure 3 As shown, the method for performing the first iterative training of the machine learning model specifically includes S301 to S311, and these steps are described in detail below.

[0104] S301 , scanning a sample object multiple times, and obtaining multiple original coil images of each sample coil in each sample slice based on the sample coil data of the sample object.

[0105] In this embodiment, the sample object is a preset magnetic resonance imaging target, such as human tissue, animal models, and non-biological materials. By repeatedly scanning the sample object using the same protocol, for each sample coil of each sample slice, the coil data of the sample coil obtained by scanning is Fourier transformed to obtain the original coil image, and the number of repeated scans is recorded as N. acq , each sample coil corresponds to N acq The original coil image.

[0106] In this embodiment, the coil image is a complex image including a real image and an imaginary image. For the sample coil C, the original coil image X acquired for the kth time isk Expressed as: in, Represents X k The real part of Represents X k The imaginary part of , i represents the imaginary unit.

[0107] S302 : For each sample coil, calculate the arithmetic mean of all original coil images of the sample coil to obtain an average coil image of the sample coil.

[0108] In this embodiment, the average coil image of the sample coil obtained by calculating the arithmetic average of multiple original coil images of the sample coil can improve the signal-to-noise ratio without changing the noise distribution.

[0109] The average coil image of sample coil j of sample slice s is denoted as X s,j , X s,j It is obtained by calculating the arithmetic mean of each original coil image of the sample coil j, that is:

[0110]

[0111] S303 : For each sample slice, reconstruct an amplitude image of the sample slice based on the average coil image of each sample coil.

[0112] In this embodiment, the number of coils is N c , the arithmetic mean of the jth sample coil of sample slice s is X s,j , then the calculation method of the amplitude image Ms of the sample slice s is:

[0113]

[0114] Where j is the coil index.

[0115] S304 : Calculate a sensitivity map and a phase map based on the amplitude image of the smoothed sample slice and the average coil image of each sample coil.

[0116] In this embodiment, the specific method for calculating the sensitivity map and the phase map based on the smoothed amplitude image of the sample slice and the average coil image of each sample coil includes:

[0117] A1. Smoothing the amplitude image of the sample slice and the amplitude image of each sample coil.

[0118] In this embodiment, each amplitude image is smoothed by a low-pass filter. Optionally, the low-pass filter is used to smooth the amplitude image using any one of the algorithms such as k-space filtering, moving average filtering or image space Gaussian kernel filtering. The amplitude image M of the sample coil j after smoothing is j，smooth and the amplitude image M of the sample slice s s,smooth As shown below:

[0119] M j,smooth =LPF(M j );

[0120] M s,smooth =LPF(Ms);

[0121] Among them, LPF represents low-pass filter, M j is the magnitude image of sample coil j, that is, M j =|X s,j |.

[0122] A2. Calculate the ratio of the amplitude image of the smoothed sample coil to the amplitude image of the sample slice to obtain the sensitivity map of the sample coil. For example, the sensitivity map S of the sample coil j of the sample slice s is: s,j The calculation method is:

[0123] S s,j =M j,smooth / M s,smooth .

[0124] A3. Use the preset angle acquisition function to obtain the original phase image p of the sample coil as follows:

[0125] p=arg(X s,j )

[0126] Wherein, arg represents the preset angle acquisition function.

[0127] A4. Based on low-pass filtering, the original phase image p of the sample coil is smoothed to obtain a smoothed phase image.

[0128] Specifically, the original cosine value cosp and the original sine value sinp of the original phase image p are smoothed to obtain the smoothed cosine value p cos and sine value p sin , calculate the cosine value p cos and sine value p sin The smooth phase diagram is obtained by summing .

[0129] p cos =LPF(cosp);

[0130] p sin=LPF(sinp);

[0131] p complex =p cos +ip sin ;

[0132] A5. Normalize the smoothed phase image to obtain the phase image of the sample coil.

[0133] Specifically, the phase diagram of the sample coil is expressed as:

[0134] p=arg(p complex ).

[0135] It should be noted that the sensitivity map and phase map for each coil can be pre-calculated and stored in a parameter database before scanning the target object. In this step, if the sensitivity map and phase map do not exist in the parameter database, they can be calculated by comparing the amplitude image of the coil and the amplitude image of the slice. Since sensitivity maps are generally smooth, the LPF has little effect on sensitivity map estimation. Due to the non-smooth nature of phase maps, low-pass filtering is required, which has little effect on sensitivity map estimation. Ideal phase maps may have edges, so the low-pass filter needs to be adjusted to avoid phase ambiguity.

[0136] S305 . For each sample slice, acquire a clean coil image of each sample coil based on the amplitude image of the sample slice and the sensitivity map and phase map of each sample coil.

[0137] In this embodiment, the real part of the clean coil image of the sample coil is the product of the amplitude image of the sample slice, the sensitivity map of the sample coil, and the cosine value of the phase map, and the imaginary part is the product of the amplitude image of the sample slice, the sensitivity map of the sample coil, and the sine value of the phase map.

[0138] That is, the clean coil image of sample coil j in sample slice s is represented as follows:

[0139]

[0140] S306 : For each sample coil, construct an original noise component of the sample coil based on a preset original seed and original intensity.

[0141] In this embodiment, the original noise component is constructed based on the original seed and the original intensity, including the real part and the imaginary part, N s,j The original noise component N of the jth sample coil is represented by s,j It can be expressed as:

[0142]

[0143] in, represents the original real noise, S307: For each sample coil, generate an original noisy coil image based on the original noise component and the clean coil image.

[0144] In this embodiment, the preset original noise component is added to the clean coil image to obtain the original noisy coil image of the sample coil. Taking the j-th sample coil as an example, the original noisy coil image of the j-th sample coil is Expressed as:

[0145]

[0146] S308 : Generate an original noisy amplitude image of the sample slice based on the original noisy coil images of each sample coil of the sample slice.

[0147] In this embodiment, the original noisy amplitude image of the sample slice s The calculation method is:

[0148]

[0149] Among them, N c represents the number of sample coils, represents the original noisy coil image of the jth sample coil.

[0150] It should be noted that the original noisy amplitude image of the sample slice is generated After that, store It can be used directly when introducing new noise components and updating the noisy amplitude image in the future, thereby reducing the amount of calculation.

[0151] S309 , recalculating the sensitivity map and the phase map based on the original noisy amplitude image of the smoothed sample slice and the original noisy coil image of each sample coil.

[0152] In this embodiment, the specific method for calculating the sensitivity map and the phase map can refer to the above S304.

[0153] The re-acquired sensitivity map is denoted as S' s,j , the phase diagram is recorded as p'.

[0154] S310 , for each sample slice, obtaining an original simulated coil image of each sample coil based on the original noisy amplitude image of the sample slice and the sensitivity map and phase map of each sample coil.

[0155] In this embodiment, the original simulated coil image of the sample coil j of the sample slice s The calculation method is:

[0156]

[0157] In this embodiment, S309-S310 are specific methods for obtaining original simulated coil images of each sample coil based on the original noisy amplitude image of the sample slice and the original noisy coil image of each sample coil, and can be referred to in S304-S305 accordingly. S311: Perform the first training of the preset machine learning model based on the original training data.

[0158] In this embodiment, the original training data is the training data used for the first iterative training. The original training data includes multiple original samples and corresponding labels. The original samples include the target complex component of the original noisy coil image of the sample coil and the target complex component of the original simulated coil image. The label is the target complex component of the clean coil image of the sample coil. The target complex component is either the real part or the imaginary part.

[0159] For example, the original training data includes multiple original samples and corresponding labels. The original samples include the real part of the original noisy coil image of the sample coil and the real part of the original simulated coil image of the sample coil. The label is the real part of the clean coil image of the sample coil.

[0160] In this embodiment, the specific method of iteratively training the machine learning model is that in each iteration, each sample is input into the machine learning model, the label is used as the target output, the loss value is calculated based on the preset loss function, and the model parameters are optimized using the preset optimization algorithm based on the loss value to enter the next iteration.

[0161] It can be seen from the above technical solution that the present application uses the target complex component of the noisy coil image and the simulated coil image as input to train the machine learning model. Since the simulated coil image is simulated based on the noisy coil image and the noisy amplitude image, the machine learning model can use the image information of the MRI coil image and the MRI amplitude image at the same time. With the help of the amplitude image, the difficulty of noise estimation can be reduced, and the zero-mean noise property of the model output image can be guaranteed.

[0162] See also Figure 4 , Figure 4 A specific implementation flow chart of another model training method provided in the embodiment of the present application is as follows: Figure 4 As shown, the method for performing target iterative training of the machine learning model specifically includes S401 to S406, and these steps are described in detail below.

[0163] S401: Determine whether a preset training completion condition is met. If not, for each sample coil in each sample slice, use a new seed and a new intensity to construct a new noise component of the sample coil.

[0164] In this embodiment, the sample coil is selected from the sample coil of the sample slice, and the new noise component of the sample coil is recorded as in, represents the real part of the new noise component, Represents the imaginary part of the new noise component.

[0165] S402 : For each sample coil, based on the target complex components of the original noise components and the target complex components of the new noise components of the sample coil, update the target complex components of the original noisy coil image to obtain the target complex components of the new noisy coil image.

[0166] In this embodiment, the real part or imaginary part of the original noisy coil image is updated to obtain the real part or imaginary part of the new noisy coil image. Taking the target complex component as the real part as an example, the original noisy coil image of the sample coil ch is updated. The real part of The new noisy coil image obtained New Real Department for:

[0167]

[0168] in, express The real part of represents the real part of the new noise component, Represents the real part of the original noise component.

[0169] S403 : For each sample slice, based on the new noisy coil images of each target sample coil of the sample slice, update the original noisy amplitude image of the sample slice to obtain a new noisy amplitude image of the sample slice.

[0170] In this embodiment, the target sample coil is a coil in the local update set. The local update set is divided according to channels. The index set of the target coil in the local update set is recorded as Including the channels corresponding to the local update set, the new noisy amplitude image The update method is:

[0171]

[0172] The new noise component N' is generated by a new random seed and intensity. In this way, multiple noise components can be updated. Indicates that r is an index set Index in. When coil ch and coil j ≠ ch are related. The noise components of these coils j≠ch are also updated.

[0173] S404 : recalculating the sensitivity map and the phase map based on the new noisy amplitude image of the smoothed sample slice and the new noisy coil image of each sample coil.

[0174] In this embodiment, the specific method for calculating the sensitivity map and the phase map can refer to the above S304.

[0175] The re-obtained sensitivity map is denoted as S” s,j , the phase diagram is recorded as p".

[0176] S405 . For each sample slice, obtain a new simulated coil image of each sample coil based on the new noisy amplitude image of the sample slice and the sensitivity map and phase map of each sample coil.

[0177] In this embodiment, the new simulated coil image of the sample coil j of the sample slice s The calculation method is:

[0178]

[0179] In this embodiment, S404 to S405 are specific methods for obtaining new simulated coil images of each sample coil based on the new noisy amplitude image of the sample slice and the new noisy coil image of each sample coil. Corresponding references may be made to S404 to S405.

[0180] S406: Based on the target training data, perform target iterative training on the preset machine learning model. If the training completion conditions are met, a noise reduction model is obtained.

[0181] In this embodiment, the target iterative training is any iterative training except the first iterative training. The target training data is the training data used for the target iterative training. The target training data includes multiple target samples and corresponding labels. The target samples include the target complex component of the new noisy coil image of the sample coil and the target complex component of the new simulated coil image. The label is the target complex component of the clean coil image of the sample coil. The target complex component is one of the real part and the imaginary part.

[0182] It can be seen from the above technical solution that an embodiment of the present application provides a model training method, which trains a machine learning model to obtain a denoising model by collecting the target complex components of the new noisy coil image and the target complex components of the new simulated coil image as sample images. Since the new simulated coil image is simulated based on the new noisy coil image and the new noisy amplitude image, the denoising model can safely remove the noise in the coil image by utilizing the high signal-to-noise ratio of the amplitude image and the zero-mean noise characteristics in the coil image, and at the same time utilize the high signal-to-noise ratio in the amplitude image to improve the image quality.

[0183] During training, a noisy coil image is created with a random seed. Accordingly, the noise component of the magnitude image is partially updated, thereby updating the simulated coil image to add noise as an enhancement. Consequently, the noise components in both the coil image and the magnitude image are updated. In the magnitude image, the noise components of some coils are updated. By pre-preparing the noise magnitude and the seed used to create the noise component, the noise in the magnitude image can be partially updated, eliminating the computationally intensive noise enhancement required in the magnitude image. In the coil image, the noise can be fully updated at each iteration, effectively increasing the noise variance since the noise in the coils is updated 100% during training.

[0184] It should be noted that the model training method provided in the embodiment of the present application can also be implemented through other specific implementation processes, for example:

[0185] In one possible implementation, the training data is fed into a machine learning model after image augmentation such as rotation, flipping, and other processing.

[0186] In one possible implementation, the method further includes storing the original noisy amplitude image of each sample slice, as well as the sensitivity map, phase map, clean coil image, original seed, original intensity, and original noisy coil image of each sample coil in a training database.

[0187] In one possible implementation, a model training method provided in an embodiment of the present application can be applied to model training of an artifact prediction model. That is, the undesirable feature prediction model is an artifact prediction model, and the artifact prediction model is used to predict artifacts in an amplitude image.

[0188] Specifically, in the implementation of the model training method for the artifact removal model, instead of using seed and intensity to create thermal noise components, other parameters are recorded to create artifacts on the coil image. Noise simulation is replaced by artifact simulation. The undesirable characteristic components are referred to as artifacts.

[0189] In one possible implementation, a model training method provided in an embodiment of the present application can be applied to a prediction model for complex undesirable features. Complex undesirable features are combinations of noise and artifacts. Noise simulation is replaced by complex undesirable feature simulation. Undesirable feature components refer to complex undesirable feature components.

[0190] The above introduces a model training method provided by an embodiment of the present application. The following will introduce a device for executing the above model training method.

[0191] See also Figure 5 , Figure 5 This is a schematic diagram of the structure of a model training device provided in an embodiment of the present application. Figure 5As shown, the model training device 50 includes a plurality of training units ( Figure 5 Only one training unit is shown). The iterative training unit 51 includes:

[0192] A clean image acquisition unit 511 is used to acquire a clean coil image of each sample coil of each sample slice;

[0193] a defective coil image acquiring unit 512, configured to generate, for each sample coil, a defective characteristic coil image of the sample coil based on the defective characteristic component of the sample coil and a clean coil image;

[0194] The bad amplitude image acquiring unit 513 is configured to generate a bad characteristic amplitude image of the sample slice based on the bad characteristic coil images of each of the sample coils of the sample slice;

[0195] A simulated coil image acquisition unit 514 is configured to acquire a simulated coil image of each of the sample coils based on the defective characteristic amplitude image of the sample slice and the defective characteristic coil image of each of the sample coils;

[0196] An iterative unit 515 is used to train a preset machine learning model based on the training data, and obtain a bad feature removal model if the training completion condition is met;

[0197] The training data includes multiple samples and corresponding labels. The samples include a target complex component of a bad characteristic coil image of the sample coil and a target complex component of a simulated coil image. The label is a target complex component of a clean coil image of the sample coil. The target complex component is one of the real part and the imaginary part.

[0198] In a possible implementation, the training unit is a first training unit for a first iterative training, and the clean image acquisition unit in the first training unit is configured to acquire a clean coil image of each sample coil of each sample slice, specifically configured to:

[0199] Scanning the sample object multiple times, and obtaining multiple original coil images of each sample coil in each sample slice based on the sample coil data of the sample object;

[0200] For each of the sample coils, calculating an arithmetic mean of all original coil images of the sample coil to obtain an average coil image of the sample coil;

[0201] reconstructing an amplitude image of the sample slice based on an average coil image of each of the sample coils corresponding to the sample slice;

[0202] A clean coil image of each of the sample coils in the sample slice is acquired based on the amplitude image of the sample slice and the sensitivity map and phase map of each of the sample coils.

[0203] In a possible implementation, the clean image acquisition unit in the first training unit is configured to acquire a clean coil image of each of the sample coils in the sample slice based on the amplitude image of the sample slice and the sensitivity map and phase map of each of the sample coils, including:

[0204] performing smoothing processing on the amplitude image of the sample slice and the amplitude image of each sample coil respectively;

[0205] calculating a ratio of the smoothed amplitude image of the sample coil to the amplitude image of the sample slice to obtain a sensitivity map of the sample coil;

[0206] Acquiring an original phase image of the sample coil using a preset angle acquisition function;

[0207] Smoothing the original phase image of the sample coil to obtain a smoothed phase image of the sample coil;

[0208] Normalizing the smoothed phase image of the sample coil to obtain the phase image of the target coil;

[0209] Based on the amplitude image of the sample slice and the sensitivity map and phase map of each sample coil, a clean coil image of each sample coil in the sample slice is obtained, where the real part of the clean coil image of the sample coil is the product of the amplitude image of the sample slice, the sensitivity map of the sample coil, and the cosine value of the phase map, and the imaginary part is the product of the amplitude image of the sample slice, the sensitivity map of the sample coil, and the sine value of the phase map.

[0210] In a possible implementation, the simulated coil image acquisition unit is configured to acquire the simulated coil image of each sample coil based on the defective characteristic amplitude image of the sample slice and the defective characteristic coil image of each sample coil, specifically for:

[0211] Calculating a sensitivity map and a phase map based on the smoothed bad characteristic amplitude image of the sample slice and the bad characteristic coil image of each of the sample coils;

[0212] For each sample slice, a simulated coil image of each sample coil is acquired based on the defective characteristic amplitude image of the sample slice and the sensitivity map and phase map of each sample coil. The real part of the simulated coil image of the sample coil is the product of the defective characteristic amplitude image of the sample slice, the sensitivity map of the sample coil, and the cosine value of the phase map. The imaginary part is the product of the defective characteristic amplitude image of the sample slice, the sensitivity map of the sample coil, and the sine value of the phase map.

[0213] In one possible implementation, the training unit is a target training unit for target iterative training, and the defective coil image acquisition unit in the target training unit is configured to, when the iterative training is any target iterative training except the first iterative training, generate, for each sample coil, a defective characteristic coil image of the sample coil based on the defective characteristic component of the sample coil and a clean coil image, specifically:

[0214] Determine whether the training completion conditions are met;

[0215] If not, constructing a bad feature component of each sample coil of each sample slice as a new bad feature component;

[0216] For each sample coil, the target complex components of the original undesirable characteristic coil image of the sample coil are updated based on the target complex components of the original undesirable characteristic components of the sample coil and the target complex components of the new undesirable characteristic components to obtain the target complex components of the new undesirable characteristic coil image of the sample coil.

[0217] In a possible implementation, the bad amplitude image acquisition unit in the target training unit is configured to generate the bad characteristic amplitude image of the sample slice based on the bad characteristic coil images of each of the sample coils of the sample slice, and is configured to:

[0218] For each of the sample slices, determining a local update set of the sample slice; wherein the local update set includes a plurality of target sample coils;

[0219] Based on the new bad characteristic coil image of each target sample coil in the local update set, the original bad characteristic amplitude image of the sample slice is updated to obtain a new bad characteristic amplitude image of the sample slice.

[0220] An electronic device is also provided in an embodiment of the present application. Figure 6, which shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present application. The electronic device in the embodiments of the present application may include but is not limited to fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 6 The electronic device shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0221] like Figure 6 As shown, the electronic device may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing device 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0222] Typically, the following devices may be connected to the I / O interface 605: a device 606 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a memory card, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Although Figure 6 The electronic device is shown with various devices, but it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed instead.

[0223] An embodiment of the present application also provides a computer program product including computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements any one of the model training methods provided in the embodiments of the present application.

[0224] A computer-readable storage medium is also provided in an embodiment of the present application. The storage medium carries one or more computer programs. When one or more computer programs are executed by an electronic device, the electronic device can implement any model training method provided in the embodiment of the present application.

[0225] It should also be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided in this application, the connection relationship between the modules indicates that there is a communication connection between them, which can be specifically implemented as one or more communication buses or signal lines.

[0226] Through the description of the above embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course can also be implemented by special hardware including application-specific integrated circuits, special CPUs, special memories, special components, etc. In general, all functions performed by computer programs can be easily implemented with corresponding hardware, and the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits or special circuits, etc. However, for the present application, software program implementation is a better implementation method in most cases. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer's floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk or optical disk, etc., and includes a number of instructions to enable a computer device (which can be a personal computer, training equipment, or network equipment, etc.) to execute the methods described in each embodiment of the present application.

[0227] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0228] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, a computer, a training device or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website, a computer, a training device or a data center. The computer-readable storage medium can be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium can be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

Claims

1. A model training method, characterized in that: The method comprises multiple iterative trainings, wherein the iterative trainings include: acquiring a clean coil image of each sample coil for each sample slice; For each of the sample coils, generating a bad characteristic coil image of the sample coil based on the bad characteristic component of the sample coil and a clean coil image; generating a bad characteristic amplitude image of the sample slice based on the bad characteristic coil image of each sample coil of the sample slice; acquiring a simulated coil image of each of the sample coils based on the defective characteristic amplitude image of the sample slice and the defective characteristic coil image of each of the sample coils; Based on the training data, the preset machine learning model is trained. If the training completion conditions are met, a bad feature removal model is obtained; The training data includes multiple samples and corresponding labels. The samples include a target complex component of a bad characteristic coil image of the sample coil and a target complex component of a simulated coil image. The label is a target complex component of a clean coil image of the sample coil. The target complex component is one of the real part and the imaginary part.

2. The model training method according to claim 1, characterized in that If the iterative training is the first iterative training, obtaining a clean coil image of each sample coil of each sample slice includes: Scanning the sample object multiple times, and obtaining multiple original coil images of each sample coil in each sample slice based on the sample coil data of the sample object; For each of the sample coils, calculating an arithmetic mean of all original coil images of the sample coil to obtain an average coil image of the sample coil; reconstructing an amplitude image of the sample slice based on an average coil image of each of the sample coils corresponding to the sample slice; A clean coil image of each of the sample coils in the sample slice is acquired based on the amplitude image of the sample slice and the sensitivity map and phase map of each of the sample coils.

3. The model training method according to claim 2, characterized in that The step of acquiring a clean coil image of each of the sample coils in the sample slice based on the amplitude image of the sample slice and the sensitivity map and phase map of each of the sample coils comprises: performing smoothing processing on the amplitude image of the sample slice and the amplitude image of each sample coil respectively; calculating a ratio of the smoothed amplitude image of the sample coil to the amplitude image of the sample slice to obtain a sensitivity map of the sample coil; Acquiring an original phase image of the sample coil using a preset angle acquisition function; Smoothing the original phase image of the sample coil to obtain a smoothed phase image of the sample coil; Normalizing the smoothed phase image of the sample coil to obtain the phase image of the target coil; Based on the amplitude image of the sample slice and the sensitivity map and phase map of each sample coil, a clean coil image of each sample coil in the sample slice is obtained, where the real part of the clean coil image of the sample coil is the product of the amplitude image of the sample slice, the sensitivity map of the sample coil, and the cosine value of the phase map, and the imaginary part is the product of the amplitude image of the sample slice, the sensitivity map of the sample coil, and the sine value of the phase map.

4. The model training method according to claim 1, characterized in that The acquiring of the simulated coil image of each sample coil based on the bad characteristic amplitude image of the sample slice and the bad characteristic coil image of each sample coil comprises: Calculating a sensitivity map and a phase map based on the smoothed bad characteristic amplitude image of the sample slice and the bad characteristic coil image of each of the sample coils; For each sample slice, a simulated coil image of each sample coil is acquired based on the defective characteristic amplitude image of the sample slice and the sensitivity map and phase map of each sample coil. The real part of the simulated coil image of the sample coil is the product of the defective characteristic amplitude image of the sample slice, the sensitivity map of the sample coil, and the cosine value of the phase map. The imaginary part is the product of the defective characteristic amplitude image of the sample slice, the sensitivity map of the sample coil, and the sine value of the phase map.

5. The model training method according to claim 1, characterized in that If the iterative training is any target iterative training except the first iterative training, generating, for each sample coil, a bad feature coil image of the sample coil based on the bad feature component of the sample coil and a clean coil image, includes: Determine whether the training completion conditions are met; If not, constructing a bad feature component of each sample coil of each sample slice as a new bad feature component; For each sample coil, the target complex components of the original undesirable characteristic coil image of the sample coil are updated based on the target complex components of the original undesirable characteristic components of the sample coil and the target complex components of the new undesirable characteristic components to obtain the target complex components of the new undesirable characteristic coil image of the sample coil.

6. The model training method according to claim 5, characterized in that The generating of the bad characteristic amplitude image of the sample slice based on the bad characteristic coil image of each sample coil of the sample slice includes: For each of the sample slices, determining a local update set of the sample slice; wherein the local update set includes a plurality of target sample coils; Based on the new bad characteristic coil image of each target sample coil in the local update set, the original bad characteristic amplitude image of the sample slice is updated to obtain a new bad characteristic amplitude image of the sample slice.

7. A model training device, characterized in that: The invention comprises a plurality of training units for performing iterative training; the iterative training units include: a clean image acquisition unit, configured to acquire a clean coil image of each sample coil of each sample slice; a defective coil image acquiring unit, configured to generate, for each of the sample coils, a defective characteristic coil image of the sample coil based on the defective characteristic component of the sample coil and a clean coil image; a bad amplitude image acquiring unit, configured to generate a bad characteristic amplitude image of the sample slice based on the bad characteristic coil images of each of the sample coils of the sample slice; a simulated coil image acquisition unit, configured to acquire a simulated coil image of each of the sample coils based on the bad characteristic amplitude image of the sample slice and the bad characteristic coil image of each of the sample coils; The iteration unit is used to train the preset machine learning model based on the training data. If the training completion conditions are met, a bad feature removal model is obtained; The training data includes multiple samples and corresponding labels. The samples include a target complex component of a bad characteristic coil image of the sample coil and a target complex component of a simulated coil image. The label is a target complex component of a clean coil image of the sample coil. The target complex component is one of the real part and the imaginary part.

8. A computer program product, characterized in that It includes computer-readable instructions, which, when executed on an electronic device, enable the electronic device to implement the model training method as described in any one of claims 1 to 6.

9. An electronic device, characterized in that: comprising at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program so that the electronic device can implement the model training method as described in any one of claims 1 to 6.

10. A computer storage medium, characterized in that The storage medium carries one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the model training method as described in any one of claims 1 to 6.

Images