Magnetic Resonance Parameter Imaging Method, Device and Intelligent Terminal Based on Deep Learning

Abstract

This application relates to the field of magnetic resonance imaging, and in particular, to a magnetic resonance parameter imaging method, device, and intelligent terminal based on deep learning. The magnetic resonance parameter imaging method based on deep learning includes: determining the forward conversion relationship from multiple first low-resolution weighted images to a first high-resolution quantitative image, and determining the inverse conversion relationship from the first high-resolution quantitative image to multiple first high-resolution weighted images; constructing an integrated network based on a quantitative convolutional network and a physical model; obtaining a training data set; training the integrated network based on the training data set; and inputting multiple second low-resolution weighted images to be analyzed into the trained integrated network to obtain a super-resolution quantitative image. This application has the characteristics of improving the performance and interpretability of the trained integrated network, eliminating the cumbersome steps of first performing image super-resolution processing and then non-linear fitting processing in related technologies, and reducing the overall time consumption.

Description

Technical Field

[0001] This application relates to the field of magnetic resonance imaging technology, and in particular, to a magnetic resonance parameter imaging method, apparatus, and intelligent terminal based on deep learning. Background Art

[0002] Magnetic resonance imaging technology, as a commonly used non-invasive detection technology, can provide rich information about various tissues inside the human body, and display sectional images of the anatomy and pathology of different human structures in different grayscales. Compared with CT (Computed Tomography) imaging technology, magnetic resonance imaging technology has the advantages of no ionizing radiation and high soft tissue contrast. Different tissues in the human body can be distinguished by magnetic resonance imaging, and the differences in image signals originate from different parameters of the tissues themselves, such as longitudinal relaxation time (T1), transverse relaxation time (T2), and proton density (PD), etc. In clinical applications, the parameter settings of magnetic resonance imaging include echo time (TE), repetition time (TR), and flip angle (FA), etc. By setting different parameters, different image weightings can be obtained, such as T1-weighted or T2-weighted images. These weighted images can usually reflect the characteristics of many common diseases and can be used for disease diagnosis. However, this diagnosis method often relies on the subjective evaluation of radiologists. In such magnetic resonance images, a tissue is usually described as brighter or darker than other regions. The problem is that the absolute intensity of the image signal and its differences have no direct meaning, and clinical diagnosis only focuses on the contrast differences between different tissues.

[0003] To alleviate the above problems, there is a magnetic resonance quantitative imaging technology in related technologies. For example, a magnetic resonance imaging quantitative parameter calculation method and apparatus disclosed by the original applicant in the invention authorization announcement number CN101912262B. The magnetic resonance quantitative imaging technology changes specific imaging parameters, acquires a series of contrast-weighted images with different contrasts, and combines a non-linear fitting algorithm to obtain a magnetic resonance quantitative parameter image. Using the magnetic resonance quantitative parameter image, the quantitative relaxation parameter value of the voxel can be directly obtained. The magnetic resonance quantitative imaging technology can study the effects of diseases and treatments on organisms through quantitative values, enabling quantitative comparison of magnetic resonance images on different patients and different machines, thereby reducing the differences in different objective factors (such as different machines, different patients, different times, etc.), and providing a basis for disease grading, prognosis evaluation, etc. Therefore, it has very important reference value for both clinical and research.

[0004] The core of magnetic resonance quantitative imaging technology is to obtain corresponding magnetic resonance quantitative parameter images from a series of magnetic resonance weighted images, and then obtain corresponding quantitative relaxation parameter values. In order to obtain more valuable quantitative relaxation parameter values, high-resolution (such as a resolution of 200*100 or above) and high-quality magnetic resonance weighted images need to be acquired. However, acquiring high-resolution images takes too much time. Therefore, image super-resolution technology is usually used in the field of magnetic resonance quantitative imaging technology.

[0005] Referring to Figure 1 , when performing magnetic resonance scanning, the magnetic resonance quantitative imaging system first acquires a series of low-resolution magnetic resonance weighted images, and then through image super-resolution technology, converts a series of low-resolution magnetic resonance weighted images into a series of high-resolution magnetic resonance weighted images. This type of magnetic resonance weighted image converted from low resolution to high resolution is also called super-resolution magnetic resonance weighted image. Then, a non-linear fitting algorithm is used to process each pixel of a series of super-resolution magnetic resonance weighted images, and finally a high-resolution magnetic resonance quantitative parameter image is obtained. Such a magnetic resonance quantitative parameter image is also called a super-resolution magnetic resonance quantitative parameter image. The above method combining image super-resolution technology uses low-resolution scanning to reduce the scanning time. However, after the scanning is completed, image super-resolution processing and non-linear fitting still need to be performed in sequence, and the entire process still has the problems of long time consumption and low imaging efficiency. Summary of the Invention

[0006] One object of the present application is to provide a magnetic resonance parameter imaging method based on deep learning, which has the characteristics of better imaging effect and faster imaging efficiency.

[0007] Another object of the present application is to provide a magnetic resonance parameter imaging device based on deep learning, which has the characteristics of better imaging effect and faster imaging efficiency.

[0008] Another object of the present application is to provide an intelligent terminal, which has better imaging effect and faster imaging efficiency by executing the magnetic resonance parameter imaging method based on deep learning.

[0009] Another object of the present application is to provide a computer-readable storage medium, which has better imaging effect and faster imaging efficiency by storing the magnetic resonance parameter imaging method based on deep learning that can be loaded.

[0010] The above object one of the present application is achieved through the following technical solutions:

[0011] A magnetic resonance parameter imaging method based on deep learning, comprising:

[0012] Determine the forward conversion relationship from multiple first low-resolution weighted images to a first high-resolution quantitative image, and determine the inverse conversion relationship from the first high-resolution quantitative image to multiple first high-resolution weighted images;

[0013] Based on a quantitative convolutional network corresponding to the forward conversion relationship and a physical model corresponding to the inverse conversion relationship, construct an integrated network, wherein the quantitative convolutional network includes a quantitative loss function and the physical model includes a weighted loss function;

[0014] Obtain a training data set for which the forward conversion relationship and the inverse conversion relationship have been determined;

[0015] Based on the training data set, train the integrated network, wherein the quantitative loss function and the weighted loss function are optimized simultaneously during the training process;

[0016] Input multiple second low-resolution weighted images to be analyzed into the trained integrated network to obtain a super-resolution quantitative image.

[0017] By adopting the above technical solution, the integrated network is trained using a training data set for which the forward conversion relationship and the inverse conversion relationship have been determined. During the training process, the quantitative loss function and the weighted loss function are optimized simultaneously. The entire integrated network tends to satisfy both the forward conversion relationship and the inverse conversion relationship. Therefore, the optimization of the quantitative loss function is affected by the weighted loss function, so that the optimization of the quantitative convolutional network is constrained by the physical model, thereby improving the performance and interpretability of the trained integrated network. When performing magnetic resonance parameter imaging, since the quantitative convolutional network and the physical model are jointly completed using an integrated network, inputting the collected second low-resolution weighted images into the trained integrated network can directly obtain a super-resolution quantitative image, completing the conversion from low-quality magnetic resonance weighted images to high-quality magnetic resonance quantitative images. This method omits the cumbersome steps of first performing image super-resolution processing and then non-linear fitting processing in the related art, reducing the overall time consumption. In summary, the present technical solution has the characteristics of better imaging effect and faster imaging efficiency.

[0018] Optionally, the step of training the integrated network based on the training data set includes:

[0019] Based on the training data set, determine a training high-resolution weighted image, a training low-resolution weighted image, and a training high-resolution quantitative image; wherein the training low-resolution weighted image and the training high-resolution quantitative image satisfy the forward conversion relationship, and the training high-resolution quantitative image and the training high-resolution weighted image satisfy the inverse conversion relationship;

[0020] Input multiple of the training low-resolution weighted images into the integrated network. The quantitative convolutional network outputs a test quantitative image based on the multiple training low-resolution weighted images, and the physical model outputs multiple test weighted images based on the test quantitative image;

[0021] Determine the quantitative loss function based on the training high-resolution quantitative image and the test quantitative image;

[0022] Determine the weighted loss function based on the training high-resolution weighted image and the test weighted image;

[0023] Determine the total loss function based on the quantitative loss function and the weighted loss function;

[0024] Optimize the total loss function to train the integrated network until the integrated network is trained.

[0025] By adopting the above technical solution, the training low-resolution weighted image and the training high-resolution weighted image satisfy the forward transformation relationship. The training high-resolution quantitative image is equivalent to the ideal result output by the quantitative convolutional network based on the training low-resolution weighted image, while the test quantitative image is the actual result output by the quantitative convolutional network based on the training low-resolution weighted image. Therefore, with the training high-resolution weighted image as the label, combined with the test quantitative image, the quantitative loss function can be obtained, and the quantitative loss function can be used to reflect the calculation error of the current quantitative convolutional network. The training high-resolution weighted image and the training high-resolution quantitative image satisfy the forward transformation relationship. The training high-resolution weighted image is equivalent to the ideal result output by the physical model based on the training high-resolution weighted image, that is, in the case where the quantitative convolutional network can output an ideal test quantitative image, the ideal result that the physical model can output, while the test weighted image is the actual result output by the physical model based on the test quantitative image. Therefore, with the training high-resolution weighted image as the label, combined with the test weighted image, the weighted loss function can be obtained, and the weighted loss function can be used to reflect the calculation error of the physical model and can also reflect the calculation error of the current quantitative convolutional network. In order to obtain a quantitative convolutional network with smaller errors, it is necessary to optimize the quantitative loss function. The quantitative convolutional network and the physical model as a whole are an integrated network. When the total loss function of the integrated network is optimized, the quantitative loss function and the weighted loss function will be optimized simultaneously. Therefore, the optimization of the weighted loss function has a constraining effect on the optimization of the quantitative loss function, that is, the physical model has a constraining effect on the update of the quantitative convolutional network, making the update of the quantitative convolutional network constrained by the physical laws corresponding to the physical model, and finally making the output result of the integrated network more interpretable, and the output imaging effect more realistic and valuable for reference.

[0026] Optionally, the physical model is an exponential function model with a decaying trend.

[0027] By adopting the above technical solution, the magnetic resonance image physically conforms to the law of signal exponential decay, making the association between the training of the integrated network and the physical law of the magnetic resonance image stronger, thereby improving the imaging effect of the integrated network.

[0028] Optionally, the exponent of the physical model is related to the frequency-locking time;

[0029] In the step where the physical model outputs the test weighted image based on the test quantitative image, the physical model outputs the test weighted image based on the frequency-locking time corresponding to the training low-resolution weighted image input to the quantitative convolutional network and the test quantitative image output by the quantitative convolutional network.

[0030] By adopting the above technical solution, magnetic resonance quantitative images need to be obtained by processing a series of magnetic resonance weighted images corresponding to different frequency-locking times, and based on a certain frequency-locking time, the magnetic resonance quantitative image can be converted into the magnetic resonance weighted image corresponding to the frequency-locking time. Using the frequency-locking time corresponding to the training low-resolution weighted image input to the quantitative convolutional network and the test quantitative image output by the quantitative convolutional network, the physical model can convert the test quantitative image into the required test weighted image.

[0031] Optionally, the method for determining whether the training of the integrated network is completed includes:

[0032] Inputting multiple training low-resolution weighted images into the integrated network, and the quantitative convolutional network outputs a verification quantitative image;

[0033] Calculating the error based on the training high-resolution quantitative image and the verification quantitative image to determine the quantitative error;

[0034] Performing a numerical comparison based on the quantitative error and the quantitative threshold, and judging whether the training is completed based on the comparison result.

[0035] By adopting the above technical solution, the training high-resolution quantitative image is equivalent to the ideal result output by the quantitative convolutional network based on the training low-resolution weighted image, and also the ideal result output by the integrated network in actual application, while the verification quantitative image is the actual result output by the integrated network. Therefore, the quantitative error between the training high-resolution quantitative image and the verification quantitative image can reflect the deviation between the actual result and the ideal result of the integrated network, and based on the numerical comparison between the quantitative error and the quantitative threshold, it can be judged whether this deviation is within the acceptable error range. If it is acceptable, the training of the integrated network is completed.

[0036] Optionally, the total loss function is optimized by the Adam optimizer.

[0037] By adopting the above technical solution, the Adam optimizer has the characteristics of fast convergence speed and easy parameter adjustment, and is applicable to the optimization of the total loss function.

[0038] Optionally, the step of obtaining the training data set that has determined the forward conversion relationship and the reverse conversion relationship includes:

[0039] Obtain the original data set including multiple pieces of the second high-resolution weighted images;

[0040] Based on the original data set, determine the training high-resolution weighted images;

[0041] Perform low-resolution processing on the training high-resolution weighted images to obtain training low-resolution weighted images;

[0042] Perform pixel-by-pixel non-linear fitting processing on the training high-resolution weighted images to obtain training high-resolution quantitative images;

[0043] Based on the training high-resolution weighted images, the second training weighted images, and the training high-resolution quantitative images, determine the training data set.

[0044] By adopting the above technical solution, using the high-resolution weighted images in the original data set, low-resolution weighted images and high-resolution quantitative images can be obtained, and the forward conversion relationship is satisfied between the low-resolution weighted images and the high-resolution quantitative images, and the reverse conversion relationship is satisfied between the high-resolution quantitative images and the high-resolution weighted images, so that the training data set contains at least one set of low-resolution weighted images, high-resolution weighted images, and high-resolution quantitative images for which the forward conversion relationship and the reverse conversion relationship have been determined, facilitating the invocation of subsequent steps.

[0045] The second main object of the present invention also proposes a magnetic resonance parameter imaging device based on deep learning.

[0046] A magnetic resonance parameter imaging device based on deep learning includes:

[0047] A clustering simulation module for determining the forward conversion relationship from multiple first low-resolution weighted images to a first high-resolution quantitative image, and for determining the reverse conversion relationship from the first high-resolution quantitative image to multiple first high-resolution weighted images;

[0048] A network construction module for constructing an integrated network based on a quantitative convolutional network corresponding to the forward conversion relationship and a physical model corresponding to the reverse conversion relationship, wherein the quantitative convolutional network includes a quantitative loss function, and the physical model includes a weighted loss function;

[0049] A data acquisition module for acquiring a training data set in which the forward conversion relationship and the reverse conversion relationship have been determined;

[0050] A network training module for training the integrated network based on the training data set, wherein the quantitative loss function and the weighted loss function are optimized simultaneously during the training process;

[0051] A network imaging module for inputting multiple second low-resolution weighted images to be analyzed into the trained integrated network to obtain a super-resolution quantitative image.

[0052] The third main inventive object of the present invention further proposes an intelligent terminal.

[0053] An intelligent terminal includes a memory and a processor, and a computer program capable of being loaded and executed by the processor for the magnetic resonance parameter imaging method based on deep learning as described in any of the above technical solutions is stored on the memory.

[0054] The fourth main inventive object of the present invention further proposes a computer-readable storage medium.

[0055] A computer-readable storage medium stores a computer program capable of being loaded and executed by the processor for the magnetic resonance parameter imaging method based on deep learning as described in any of the above technical solutions. Description of the Drawings

[0056] Figure 1 is a schematic flowchart of magnetic resonance quantitative imaging in the related art.

[0057] Figure 2 is a schematic flowchart of the magnetic resonance parameter imaging method based on deep learning of the present application.

[0058] Figure 3 is a working process demonstration diagram of the quantitative convolution network and the physical model of the present application.

[0059] Figure 4 is a schematic sub-flowchart of steps S3 and S4 in the magnetic resonance parameter imaging method based on deep learning of the present application.

[0060] Figure 5 is a demonstration diagram of the optimization process of the integrated network of the present application.

[0061] Figure 6 is a working process demonstration diagram of the integrated network of the present application during actual application.

[0062] Figure 7 is a module schematic diagram of the magnetic resonance parameter imaging device based on deep learning of the present application.

[0063] In the figure, 1 is the clustering simulation module; 2 is the network construction module; 3 is the data acquisition module; 4 is the network training module; 5 is the network imaging module. Specific implementation manner

[0064] In various fields that require analysis using image content, in order to improve the acquisition efficiency of high-quality images, image super-resolution technology is usually adopted. Among them, the image obtained by resolution enhancement processing from a low-resolution image is called a super-resolution image. The existing magnetic resonance image super-resolution technologies can be divided into two categories: the first category is the multi-image super-resolution technology, which integrates multiple low-resolution images to obtain a high-resolution super-resolution image; the second category is the single-image super-resolution technology, which realizes the mapping from a single low-resolution image to a high-resolution image through learning or model methods.

[0065] Magnetic resonance quantitative imaging technology is a technology that changes specific imaging parameters, acquires a series of weighted images with different contrasts and different locking frequencies, and then combines non-linear fitting algorithms to process to obtain a quantitative parameter image. The value of each pixel point in the quantitative parameter image is the T1ρ value (quantitative relaxation parameter value) of the human tissue at the corresponding position. Therefore, magnetic resonance quantitative imaging technology can be summarized as a technology that uses the content of the quantitative parameter image to obtain the quantitative relaxation parameter value of the voxel. Magnetic resonance quantitative imaging technology can reduce the differences in results of different devices and provide a basis for disease grading, prognosis assessment, etc., and thus has very important clinical value and research value. However, the effect of magnetic resonance quantitative imaging technology is affected by the image resolution. However, it is very time-consuming to acquire high-resolution images. Due to reasons such as the patient's physical condition, images acquired at high resolution in a short time may have motion artifacts. Therefore, there is a contradiction between the acquisition of high-resolution images and the acceleration of the scanning speed. In addition, the processing from the weighted image to the quantitative parameter image requires the use of non-linear fitting algorithms, which contain a large number of iterative operations and the calculation process is also very time-consuming.

[0066] Refer to Figure 1, Currently, in order to improve the imaging efficiency of quantitative parameter images, the commonly used image super-resolution technology in the field of magnetic resonance quantitative imaging is the second type, which uses a constructed low-resolution - high-resolution image pair (LR-HR) to train a convolutional neural network, and applies the trained convolutional neural network to the conversion of the acquired low-resolution images to super-resolution images. Specifically, during magnetic resonance scanning, the system only acquires a series of low-resolution magnetic resonance weighted images, and each magnetic resonance weighted image has a different locking frequency time; then, the trained convolutional neural network is used to convert this series of low-resolution magnetic resonance weighted images into a series of super-resolution magnetic resonance weighted images; then, a non-linear fitting algorithm is applied to each pixel of this series of super-resolution magnetic resonance weighted images, and finally, a super-resolution magnetic resonance quantitative image is obtained. However, the above method reduces the scanning duration by using low-resolution scanning, but still requires image super-resolution processing and non-linear fitting in sequence after scanning, and the entire process still has the problems of long time consumption and low imaging efficiency.

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application. Additionally, the labels of the steps in this embodiment are only for convenience of description and do not represent the limitation of the execution order of the steps. In actual applications, the execution order of the steps can be adjusted as needed, or they can be executed simultaneously, and these adjustments or substitutions all fall within the protection scope of the present invention.

[0068] The following combines the specification appendix Figures 1-7 to further describe the embodiments of the present application in detail.

[0069] The embodiments of the present application provide a magnetic resonance parameter imaging method based on deep learning, and the main process of the method is described as follows.

[0070] Refer to Figure 2 and Figure 3 , S1. Determine the forward conversion relationship from multiple first low-resolution weighted images to a first high-resolution quantitative image, and determine the reverse conversion relationship from the first high-resolution quantitative image to multiple first high-resolution weighted images.

[0071] Among them, the magnetic resonance images described below include weighted images and quantitative parameter images; the weighted images refer to the images obtained by magnetic resonance scanning, and each weighted image has a corresponding frequency locking time. By setting different frequency locking times, different weighted images can be obtained; the quantitative parameter images refer to the images obtained by using a non-linear fitting algorithm pixel by pixel for a group of weighted images corresponding to different frequency locking times. Usually, one quantitative parameter image needs to be processed from five to eight weighted images. Each pixel point of the quantitative parameter image represents a quantitative relaxation parameter value. Therefore, each quantitative parameter image is equivalent to a data set of quantitative relaxation parameter values.

[0072] The first low-resolution weighted image refers to a weighted image with a lower resolution, the first high-resolution weighted image refers to a weighted image with a higher resolution, and the first high-resolution quantitative image refers to a high-resolution quantitative parameter image obtained by using a non-linear fitting algorithm pixel by pixel for a group of weighted images with a higher resolution. Among them, the above-mentioned resolution refers to the resolution within the magnetic resonance slice, and the evaluation criteria for the high and low resolution are determined according to the actual situation. In this embodiment, when the resolution is less than 200*100, it is regarded as low resolution, and when the resolution is greater than or equal to 200*100, it is regarded as high resolution.

[0073] The forward conversion relationship refers to the conversion relationship from multiple first low-resolution weighted images to the first high-resolution quantitative image, which is the conversion from multiple weighted images to a single quantitative parameter image and can be realized by a convolutional neural network with deep learning ability. The reverse conversion relationship refers to the conversion relationship from the first high-resolution quantitative image to multiple first high-resolution weighted images, which is the conversion from a single quantitative parameter image to multiple weighted images. Since the conversion from multiple weighted images to a single quantitative parameter image is completed by non-linear model fitting, and this model is usually an exponential model, the reverse conversion relationship can be realized by a function model associated with the exponential model.

[0074] S2. Construct an integrated network based on the quantitative convolutional network corresponding to the forward conversion relationship and the physical model corresponding to the reverse conversion relationship.

[0075] Among them, the quantitative convolutional network is a convolutional neural network that can realize the forward conversion relationship. Input multiple first low-resolution weighted images into the quantitative convolutional network, and the quantitative convolutional network can output the first high-resolution quantitative image. The quantitative convolutional network has learning ability and can be updated, and the quantitative convolutional network will be updated during the training process of the integrated network. In this embodiment, the number of convolutional modules in the quantitative convolutional network, the number of convolutional layers, the number of output channels of the convolutional layer, and the size of the convolutional kernel and other parameters are not limited, as long as the quantitative convolutional network can realize the forward conversion relationship, and the actual parameters used can be adaptively adjusted according to the actual situation.

[0076] The physical model is an exponential function model with a decaying trend. Since magnetic resonance images physically conform to the law of signal exponential decay, using an exponential function can make the physical model more in line with the real situation and increase the authenticity and interpretability of its output results. Since the physical model reflects the physical laws of magnetic resonance, the physical model itself does not change during the training process of the integrated network.

[0077] The physical model is specifically shown in formula (1).

[0078] (1)

[0079] Among them, S is the first high-resolution weighted image output by the physical model, S0 is the first high-resolution quantitative image input to the physical model, e is the exponential function, TSL is the lock-in time, and T1ρ is the quantitative relaxation parameter value.

[0080] It should be noted that in the process of using a non-linear model to fit a single quantitative parameter image based on multiple weighted images, the input content is the weighted images and the lock-in times corresponding to the weighted images one by one, and the output fitting result is the quantitative parameter image. In the process of using the physical model to obtain multiple weighted images based on a single quantitative parameter image, the input content is the quantitative parameter image and the lock-in time, and the output result is the weighted image corresponding to the lock-in time one by one. Non-linear model fitting is a mathematical method that uses weighted images and lock-in times to obtain a quantitative parameter image, while the physical model reflects an objective physical law of magnetic resonance quantitative imaging. Through the lock-in time and the quantitative parameter image, the weighted image is obtained conversely.

[0081] Combining the quantitative convolutional network and the physical model, an integrated network can be constructed, enabling the integrated network to simultaneously achieve forward and reverse transformation relationships. Specifically, the quantitative convolutional network and the physical model are equivalent to two modules in series in the integrated network. After multiple first low-resolution weighted images are input into the integrated network, the quantitative convolutional network outputs a first high-resolution quantitative image based on the multiple first low-resolution weighted images, and then the physical model outputs multiple first high-resolution weighted images based on the first high-resolution quantitative image.

[0082] The constructed integrated network needs to be trained first. After the calculation error of the integrated network reaches the expected standard, the integrated network will be officially put into the normal magnetic resonance parameter imaging task. Therefore, the integrated network has a training stage and an application stage.

[0083] Specifically, the quantitative convolutional network corresponds to a quantitative loss function, the physical model corresponds to a weighted loss function, and the integrated network has a total loss function, which is the integration of the quantitative loss function and the weighted loss function. The training of the integrated network is actually an optimization process of the total loss function. When optimizing the total loss function, the quantitative loss function and the weighted loss function are optimized simultaneously, so that the finally trained integrated network can not only accurately convert the weighted image to the quantitative parameter image, but also reflect the physical laws of magnetic resonance imaging, making the integrated network have better performance and interpretability. It should be noted that the physical model in the embodiments of the present application is a function model set based on the physical laws of magnetic resonance imaging. Therefore, the physical model does not change during the process of optimizing the total loss function, and only the quantitative convolutional network will change.

[0084] During the training stage of the integrated network, the integrated network outputs both the first high-resolution quantitative image and the first high-resolution weighted image. Among them, the first high-resolution quantitative image is used to calculate the quantitative loss function, which can reflect the error between the actual output and the ideal output of the quantitative convolutional network; the first high-resolution weighted image is used to calculate the weighted loss function, which can reflect the error between the actual output and the ideal output of the physical model. However, the first high-resolution weighted image is calculated from the first high-resolution quantitative image, and the first high-resolution quantitative image is actually the output of the current quantitative convolutional network. Therefore, the quantitative loss function can also verify the error of the quantitative convolutional network. And because the physical model itself symbolizes the physical law of exponential decay, the quantitative loss function can also verify the deviation between the output of the quantitative convolutional network and the physical law of exponential decay.

[0085] When optimizing the total loss function of the integrated network, the quantitative loss function and the weighted loss function are optimized simultaneously. Therefore, the optimization of the weighted loss function has a constraining effect on the optimization of the quantitative loss function, that is, the physical model has a constraining effect on the update of the quantitative convolutional network, making the update of the quantitative convolutional network constrained by the physical laws corresponding to the physical model, and finally making the output result of the integrated network more interpretable, and the output imaging effect more real and valuable for reference.

[0086] During the application stage of the integrated network, since the error degree of the integrated network at this time already meets the requirements and there is no need to use the first high-resolution weighted image output by the physical model to verify the first high-resolution quantitative image output by the quantitative convolutional network, the integrated network only needs to output the first high-resolution quantitative image.

[0087] In this embodiment, the integrated network uses the mainstream super-resolution network EDSR. In other embodiments, other super-resolution networks can also be used.

[0088] S3. Obtain a training data set for which the forward conversion relationship and the reverse conversion relationship have been determined.

[0089] Among them, the training data set contains data that simultaneously satisfies the forward conversion relationship and the reverse conversion relationship, and is used to train the integrated network.

[0090] Specifically, the training data set includes at least one first high-resolution quantitative image, multiple first low-resolution weighted images corresponding to this first high-resolution quantitative image, and multiple first high-resolution weighted images corresponding to this first high-resolution quantitative image. Moreover, the first high-resolution quantitative image can satisfy the forward conversion relationship with multiple corresponding first low-resolution weighted images, and the first high-resolution quantitative image can satisfy the reverse conversion relationship with multiple corresponding first high-resolution weighted images.

[0091] Refer to Figure 3 and Figure 4 , the sub-steps of step S3 include:

[0092] S31. Obtain the original data set.

[0093] Among them, the original data set refers to a high-resolution image data set obtained by using a magnetic resonance scan with a relatively high resolution. The original data set contains multiple first high-resolution weighted images.

[0094] S32. Based on the original data set, determine multiple training high-resolution weighted images.

[0095] Among them, the training high-resolution weighted images are used to calculate the weighted loss function in the subsequent steps. The training high-resolution weighted images are the first high-resolution weighted images extracted from the original data set. Each training high-resolution weighted image corresponds to a different lock-in time, and at least one first high-resolution quantitative image can be obtained based on each training high-resolution weighted image. Since usually one quantitative parameter image needs to be processed by using five to eight weighted images, at least five training high-resolution weighted images should be extracted.

[0096] S33. Perform low-resolution processing on multiple training high-resolution weighted images to obtain multiple training low-resolution weighted images.

[0097] Among them, the training low-resolution weighted images are used as the input of the integrated network in subsequent steps. The method of low-resolution processing is as follows: take the central part of the K-space after the discrete Fourier transform of the high-resolution image and then inverse transform it to the image domain to obtain the simulated low-resolution image. Each training low-resolution weighted image and each training high-resolution weighted image. Since the training high-resolution weighted image is actually the first high-resolution weighted image, the training low-resolution weighted image can be defined as the first low-resolution weighted image and can be regarded as the low-resolution image obtained by magnetic resonance scanning with a lower resolution.

[0098] S34. Perform pixel-by-pixel non-linear fitting processing on multiple training high-resolution weighted images to obtain training high-resolution quantitative images.

[0099] Among them, the training high-resolution quantitative images are used as the input of the physical module in subsequent steps and for calculating the quantitative loss function. Since the training high-resolution weighted image is actually the first high-resolution weighted image, the training high-resolution quantitative image is actually the first high-resolution quantitative image, which is the ideal result of magnetic resonance quantitative parameter imaging and can be regarded as the ideal result of the integrated output after training.

[0100] S35. Determine the training dataset based on the training high-resolution weighted images, training low-resolution weighted images, and training high-resolution quantitative images.

[0101] Among them, the training dataset provides a data basis for the training of the integrated network. The training dataset contains at least one training high-resolution quantitative image, multiple training high-resolution weighted images corresponding to the training high-resolution quantitative image, and multiple training low-resolution weighted images corresponding one-to-one to each training high-resolution weighted image. In this embodiment, the training dataset only contains one training high-resolution quantitative image, five to eight training high-resolution weighted images, and five to eight training low-resolution weighted images. In other embodiments, if the number of training high-resolution quantitative images is changed, the number of training high-resolution weighted images and training low-resolution weighted images is adjusted adaptively.

[0102] Each training high-resolution weighted image corresponds to a lock-in time. Since each training low-resolution weighted image is converted from the corresponding training high-resolution weighted image, each training low-resolution weighted image also corresponds to a lock-in time.

[0103] S4. Train the integrated network based on the training dataset.

[0104] Among them, in the training dataset, the training low-resolution weighted image and the training high-resolution quantitative image satisfy the forward conversion relationship, and the training high-resolution quantitative image and the training high-resolution weighted image satisfy the reverse conversion relationship. Using the training low-resolution weighted image as the input of the integrated network, with the training high-resolution quantitative image and the training high-resolution weighted image as labels respectively, and calculating with the actual output results of the integrated network, the quantitative loss function and the weighted loss function can be calculated respectively, and then the total loss function of the integrated network can be obtained. Optimizing the total loss function can train the integrated network.

[0105] Referring to Figure 4 and Figure 5 , the sub-steps of step S4 include:

[0106] S41. Based on the training dataset, determine the training high-resolution weighted image, the training low-resolution weighted image, and the training high-resolution quantitative image.

[0107] Among them, the training high-resolution weighted image, the training low-resolution weighted image, and the training high-resolution quantitative image are extracted from the training dataset. The training low-resolution weighted image and the training high-resolution quantitative image satisfy the forward conversion relationship and can be used as the input and ideal output of the quantitative convolution network. The training high-resolution quantitative image and the training high-resolution weighted image satisfy the reverse conversion relationship and can be used as the input and ideal output of the physical model.

[0108] S42. Input multiple training low-resolution weighted images into the integrated network to obtain the test quantitative image and multiple test weighted images.

[0109] Among them, the quantitative convolution network and the physical model in the integrated network work in series. First, input multiple training low-resolution weighted images into the quantitative convolution network, and the quantitative convolution network outputs the test quantitative image, which is a high-resolution quantitative parameter image. Then, input the lock-in time corresponding to the multiple training low-resolution weighted images and the test quantitative image output by the quantitative convolution network into the physical model in sequence, and the physical model outputs multiple test weighted images in sequence, and the test weighted image is a high-resolution weighted image.

[0110] S43. Based on the training high-resolution quantitative image and the test quantitative image, determine the quantitative loss function.

[0111] Among them, the training high-resolution quantitative image is equivalent to the ideal result output by the quantitative convolution network based on the training low-resolution weighted image, and the test quantitative image is the actual result output by the quantitative convolution network based on the training low-resolution weighted image. Therefore, with the training high-resolution weighted image as the label, combining with the test quantitative image, the quantitative loss function can be obtained, and the quantitative loss function can be used to reflect the calculation error of the current quantitative convolution network.

[0112] S44. Determine a weighted loss function based on a training high-resolution weighted image and a test weighted image.

[0113] Among them, the training high-resolution weighted image is equivalent to the ideal result output by the physical model based on the training high-resolution weighted image, that is, the ideal result that the physical model can output when the quantitative convolution network can output an ideal test quantitative image, while the test weighted image is the actual result output by the physical model based on the test quantitative image. Therefore, taking the training high-resolution weighted image as the label and combining the test weighted image, the weighted loss function can be obtained. The weighted loss function can reflect the calculation error of the physical model and also the calculation error of the current quantitative convolution network.

[0114] S45. Determine a total loss function based on the quantitative loss function and the weighted loss function.

[0115] Among them, by combining the quantitative loss function and the weighted loss function, the total loss function can be obtained.

[0116] S46. Optimize the total loss function to train the integrated network until the integrated network training is completed.

[0117] Among them, the quantitative loss function and the weighted loss function are optimized simultaneously, so that the optimization of the weighted loss function has a constraining effect on the optimization of the quantitative loss function, that is, the physical model has a constraining effect on the update of the quantitative convolution network, making the update of the quantitative convolution network constrained by the physical laws corresponding to the physical model. Eventually, the output result of the integrated network is more interpretable, and the output imaging effect is more realistic and valuable for reference.

[0118] In this embodiment, the total loss function is preferably the L1 loss. In other embodiments, the total loss function can also be replaced by the MSE loss, perceptual loss, etc. In this embodiment, the optimization method of the total loss function is to optimize it through the Adam optimizer. In other embodiments, the optimization algorithm of the total loss function can also adopt stochastic gradient descent, AdaGrad, etc.

[0119] In this embodiment, the specific steps for determining whether the integrated network training is completed include:

[0120] A1. Input the training low-resolution weighted image into the integrated network, and the quantitative convolution network outputs a verification quantitative image.

[0121] Among them, the integrated network in this step has been optimized for the total loss function at least once. The verification quantitative image is the actual output result of the current integrated network based on the training low-resolution weighted image, which can simulate the actual result output by the integrated network in actual applications. The verification quantitative image is a high-resolution quantitative parameter image.

[0122] A2. Calculate the error based on the training high-resolution quantitative image and the verification quantitative image to determine the quantitative error.

[0123] Among them, the training high-resolution quantitative image is the ideal result output by the quantitative convolution network based on the training low-resolution weighted image, which can simulate the ideal result output by the integrated network in actual applications. The quantitative error can reflect the deviation between the actual result and the ideal result of the integrated network.

[0124] A3. Perform a numerical comparison based on the quantitative error and the quantitative threshold, and determine whether the training is completed based on the comparison result.

[0125] Among them, the quantitative threshold is a preset error threshold. Based on the numerical comparison between the quantitative error and the quantitative threshold, it can be judged whether this deviation is within an acceptable error range. In this embodiment, if the quantitative error is greater than or equal to the quantitative threshold, it is determined that the current quantitative error is too large, and the error of the quantitative convolution network in the current integrated network does not meet the preset requirements, and it is necessary to return to step S42 to continue training the quantitative convolution network; if the quantitative error is less than the quantitative threshold, it is determined that the current quantitative error is small, the current integrated network meets the error condition, and the training of the integrated network is completed.

[0126] After the training of the integrated network is completed, the integrated network can be applied to the actual magnetic resonance parameter imaging.

[0127] Refer to Figure 2 , S5. Input multiple second low-resolution weighted images to be analyzed into the trained integrated network to obtain a super-resolution quantitative image.

[0128] Among them, the second low-resolution weighted image is a weighted image obtained by magnetic resonance scanning with a lower resolution, and the super-resolution quantitative image is a super-resolution magnetic resonance weighted image of the quantitative convolution network in the integrated network.

[0129] Refer to Figure 6 , specifically, when the system performs magnetic resonance scanning, it can only collect a series of low-resolution magnetic resonance weighted images, each magnetic resonance weighted image has a different locking frequency time, and then input the series of magnetic resonance weighted images as the second low-resolution weighted images into the integrated network, and directly use the integrated network to convert the second low-resolution weighted images into super-resolution quantitative images.

[0130] The implementation principle of a magnetic resonance parameter imaging method based on deep learning provided by this application is as follows: An integrated network is trained using a training dataset for which the forward conversion relationship and the inverse conversion relationship have been determined. During the training process, the quantitative loss function and the weighted loss function are optimized simultaneously. Therefore, the optimization of the weighted loss function has a constraining effect on the optimization of the quantitative loss function, that is, the physical model has a constraining effect on the update of the quantitative convolutional network, making the update of the quantitative convolutional network constrained by the physical laws corresponding to the physical model. Ultimately, the output result of the integrated network is more interpretable, and the output imaging effect is more realistic and valuable as a reference.

[0131] The method of this application not only reduces the scanning duration by using low-resolution scanning, but also does not require separate image super-resolution processing and non-linear fitting after scanning, effectively alleviating the problem of long parameter imaging time. Additionally, since the training of the integrated network satisfies the physical laws of magnetic resonance imaging, the imaging effect of the integrated network is better and more valuable as a reference.

[0132] This application also provides a magnetic resonance parameter imaging device based on deep learning, corresponding to the above-mentioned magnetic resonance parameter imaging method based on deep learning.

[0133] Referring to Figure 7 , the magnetic resonance parameter imaging device based on deep learning includes:

[0134] A clustering simulation module 1, configured to determine the forward conversion relationship from multiple first low-resolution weighted images to a first high-resolution quantitative image, and determine the inverse conversion relationship from the first high-resolution quantitative image to multiple first high-resolution weighted images, and send them to a network construction module 2.

[0135] A network construction module 2, configured to construct an integrated network based on a quantitative convolutional network corresponding to the forward conversion relationship and a physical model corresponding to the inverse conversion relationship, where the quantitative convolutional network includes a quantitative loss function, and the physical model includes a weighted loss function, and send it to a data acquisition module 3;

[0136] A data acquisition module 3, configured to acquire a training dataset for which the forward conversion relationship and the inverse conversion relationship have been determined, and send it to a network training module 4;

[0137] A network training module 4, configured to train the integrated network based on the training dataset, and send it to a network imaging module 5, where the quantitative loss function and the weighted loss function are optimized simultaneously during the training process;

[0138] A network imaging module 5, configured to input multiple second low-resolution weighted images to be analyzed into the trained integrated network to obtain a super-resolution quantitative image.

[0139] The magnetic resonance parameter imaging device based on deep learning provided in this embodiment can achieve the various steps of the foregoing embodiment due to the functions of its respective modules and the logical connections between them. Therefore, it can achieve the same technical effects as the foregoing method. For the principle analysis, reference can be made to the relevant descriptions of the steps of the foregoing magnetic resonance parameter imaging method based on deep learning, which will not be repeated here.

[0140] This application also provides an intelligent terminal.

[0141] An intelligent terminal includes a memory, a processor, and a computer program stored on the memory and executable on the processor. The memory stores training data, algorithm formulas, filtering mechanisms, etc. in the training model. The processor is used to provide computing and control capabilities. When the processor executes the computer program, it implements the magnetic resonance parameter imaging method based on deep learning.

[0142] For the intelligent terminal provided in this embodiment, since the computer program in its memory runs on the processor and will implement the various steps of the foregoing method, it can achieve the same technical effects as the foregoing method. For the principle analysis, reference can be made to the relevant descriptions of the foregoing method steps, which will not be repeated here.

[0143] This application also provides a computer-readable storage medium.

[0144] A computer-readable storage medium includes a memory and a processor. A computer program capable of being loaded and executed by the processor, such as the magnetic resonance parameter imaging method based on deep learning as described above, is stored on the memory. When the computer program is executed by the processor, it implements the magnetic resonance parameter imaging method based on deep learning.

[0145] For the readable storage medium provided in this embodiment, since the computer program therein is loaded and run on the processor and will implement the various steps of the foregoing method, it can achieve the same technical effects as the foregoing method. For the principle analysis, reference can be made to the relevant descriptions of the foregoing method steps, which will not be repeated here.

[0146] The computer-readable storage medium includes, for example: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs.

[0147] The above are only some embodiments of this application. It should be noted that for those of ordinary skill in the art in this technical field, without departing from the principle of this application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of this application.

Claims

1. A magnetic resonance parameter imaging method based on deep learning, characterized in that, it includes: Determine the forward conversion relationship from multiple first low-resolution weighted images to a first high-resolution quantitative image, and determine the reverse conversion relationship from the first high-resolution quantitative image to multiple first high-resolution weighted images; Based on a quantitative convolutional network corresponding to the forward conversion relationship and a physical model corresponding to the reverse conversion relationship, construct an integrated network, wherein the quantitative convolutional network includes a quantitative loss function, and the physical model includes a weighted loss function; Obtain a training data set for which the forward conversion relationship and the reverse conversion relationship have been determined; Based on the training data set, train the integrated network, wherein the quantitative loss function and the weighted loss function are optimized simultaneously during the training process; Input multiple second low-resolution weighted images to be analyzed into the trained integrated network to obtain a super-resolution quantitative image.

2. The magnetic resonance parameter imaging method based on deep learning according to claim 1, characterized in that, The step of training the integrated network based on the training data set includes: Based on the training data set, determine a training high-resolution weighted image, a training low-resolution weighted image, and a training high-resolution quantitative image; wherein, the training low-resolution weighted image and the training high-resolution quantitative image satisfy the forward conversion relationship, and the training high-resolution quantitative image and the training high-resolution weighted image satisfy the reverse conversion relationship; Input multiple training low-resolution weighted images into the integrated network, the quantitative convolutional network outputs a test quantitative image based on the multiple training low-resolution weighted images, and the physical model outputs multiple test weighted images based on the test quantitative image; Based on the training high-resolution quantitative image and the test quantitative image, determine the quantitative loss function; Based on the training high-resolution weighted image and the test weighted image, determine the weighted loss function; Based on the quantitative loss function and the weighted loss function, determine a total loss function; Optimize the total loss function to train the integrated network until the integrated network is trained.

3. The magnetic resonance parameter imaging method based on deep learning according to claim 2, characterized in that, The step that the physical model outputs multiple test weighted images based on the test quantitative image includes: Input the test quantitative image into the physical model to obtain multiple test weighted images, wherein the physical model is an exponential function model with a decay trend.

4. The magnetic resonance parameter imaging method based on deep learning according to claim 3, characterized in that: The step of inputting the test quantitative image into the physical model to obtain multiple test weighted images includes: Input the lock-in time corresponding to the training low-resolution weighted image input into the quantitative convolutional network and the test quantitative image output by the quantitative convolutional network into the physical model to obtain multiple test weighted images, wherein the exponent of the physical model is related to the lock-in time.

5. The method for magnetic resonance parameter imaging based on deep learning according to claim 2, characterized in that, the method for judging whether the integrated network training is completed includes: inputting multiple pieces of the training low-resolution weighted images into the integrated network, and the quantitative convolutional network outputs a verification quantitative image; calculating an error based on the training high-resolution quantitative image and the verification quantitative image to determine a quantitative error; performing a numerical comparison based on the quantitative error and a quantitative threshold, and judging whether the training is completed based on the comparison result.

6. The method for magnetic resonance parameter imaging based on deep learning according to claim 2, characterized in that, the step of optimizing the total loss function includes: optimizing the total loss function through an Adam optimizer.

7. The method for magnetic resonance parameter imaging based on deep learning according to claim 1, characterized in that, the step of obtaining a training data set for which the forward conversion relationship and the reverse conversion relationship have been determined includes: obtaining an original data set including multiple pieces of the first high-resolution weighted images; determining multiple pieces of training high-resolution weighted images based on the original data set; performing low-resolution processing on multiple pieces of the training high-resolution weighted images to obtain multiple pieces of training low-resolution weighted images; performing pixel-by-pixel non-linear fitting processing on multiple pieces of the training high-resolution weighted images to obtain a training high-resolution quantitative image; and determining the training data set based on the training high-resolution weighted images, the training low-resolution weighted images, and the training high-resolution quantitative image.

8. A device for magnetic resonance parameter imaging based on deep learning, characterized in that, it includes: a clustering simulation module (1) for determining a forward conversion relationship from multiple pieces of first low-resolution weighted images to a first high-resolution quantitative image, and determining a reverse conversion relationship from the first high-resolution quantitative image to multiple pieces of first high-resolution weighted images; a network construction module (2) for constructing an integrated network based on a quantitative convolutional network corresponding to the forward conversion relationship and a physical model corresponding to the reverse conversion relationship, wherein the quantitative convolutional network includes a quantitative loss function, and the physical model includes a weighted loss function; a data acquisition module (3) for obtaining a training data set for which the forward conversion relationship and the reverse conversion relationship have been determined; a network training module (4) for training the integrated network based on the training data set, wherein the quantitative loss function and the weighted loss function are optimized simultaneously during the training process; and a network imaging module (5) for inputting multiple pieces of second low-resolution weighted images to be analyzed into the trained integrated network to obtain a super-resolution quantitative image.

9. An intelligent terminal, characterized in that, it includes a memory and a processor, and a computer program capable of being loaded and executed by the processor and being the method for magnetic resonance parameter imaging based on deep learning according to any one of claims 1 to 7 is stored on the memory.

10. A computer-readable storage medium, characterized in that, A computer program is stored, which can be loaded and executed by a processor, for any one of the magnetic resonance parameter imaging methods based on deep learning as recited in claims 1 to 7.

Images