Medical image super-resolution reconstruction method and system based on knowledge distillation

Abstract

The invention provides a medical image super-resolution reconstruction method and system based on knowledge distillation, and belongs to the technical field of medical image processing. According to the scheme, inter-sample and multi-scale associated knowledge learned by a teacher model is transmitted to a student model through cross-sample multi-scale knowledge transmission; therefore, the learning ability of a student model for key features is enhanced, and a reconstructed image which is more consistent with original image details is obtained. Meanwhile, according to the scheme, knowledge of the teacher model is fully utilized to guide training of the student model, the high-performance lightweight medical image super-resolution reconstruction student model can be constructed, and computing resources needed during model deployment are effectively reduced while the visual quality and the identification degree of the reconstructed image are enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to a medical image super-resolution reconstruction method and system based on knowledge distillation. BACKGROUND

[0002] With the development of precision medicine, the resolution of medical images directly affects early lesion detection and surgical planning. Traditional MRI (Magnetic Resonance Imaging), CT (Computed Tomography) and other devices are limited by hardware costs and radiation safety, and are often output in low-resolution images, resulting in a high missed diagnosis rate of up to 32% for small tumors (less than 3mm). Based on the idea of breaking through the physical limitations through software algorithms, some researchers consider using a neural network model to perform super-resolution reconstruction of low-resolution medical images. However, existing methods mainly focus on improving model performance, using larger and larger neural network models, ignoring the actual use cost of the model, making it difficult to apply in medical devices with limited computing resources (such as image acquisition devices).

[0003] Secondly, some researchers consider using neural network model compression (such as knowledge distillation) to reduce the size of the model, but traditional knowledge distillation methods usually only focus on the feature representation within a single sample, seriously ignoring the spatial distribution structure of sample features, resulting in the student model being unable to effectively learn the cross-sample spatial relationship in the teacher model, and thus the student model obtained based on the traditional knowledge distillation method has general image reconstruction effect, which cannot meet the actual needs. SUMMARY

[0004] To overcome the above-mentioned deficiencies of the prior art, the present application provides a medical image super-resolution reconstruction method and system based on knowledge distillation, which effectively solves the problem that the traditional reconstruction method cannot be applied in medical devices with limited computing resources due to the large size of the model, and the problem that the student model obtained by the traditional knowledge distillation method has general image reconstruction effect and cannot meet the actual needs.

[0005] According to a first aspect of an embodiment of the present application, a medical image super-resolution reconstruction method based on knowledge distillation is provided, comprising:

[0006] obtaining a low-resolution medical image to be reconstructed;

[0007] using the low-resolution medical image as the input of a pre-trained neural network-based super-resolution reconstruction model to obtain the reconstruction result of the medical image; wherein the training of the super-resolution reconstruction model adopts a knowledge distillation method, and the training process is as follows:

[0008] a teacher model and a student model based on a neural network are respectively constructed;

[0009] The teacher model is independently trained based on a pre-constructed medical image training set, and a trained teacher model is obtained.

[0010] The student model is trained based on the pre-constructed medical image training set and the trained teacher model, and the trained student model is used as a super-resolution reconstruction model. In the training process of the student model, the student model is driven to learn the correlation representation of the teacher model through the correlation matrix between all sample features in the same training batch of the teacher model and the student model.

[0011] Further, the training of the student model uses the following loss function:

[0012]

[0013] wherein, the loss function used in the training of the student model, the student model, LR represents a low-resolution medical image, and HR represents a corresponding high-resolution medical image, the L1 norm, respectively represent the number of samples in a batch, the height and width of each sample feature extracted by the student model and the teacher model, the softmax function normalization, the KL divergence loss function, and respectively represent the correlation matrix between all sample features in a training batch of the student model and the teacher model, i and j represent the coordinate positions of the elements in the correlation matrix, is an adjustment parameter, R represents a set of real numbers, and BHW represents the product of B, H and W.

[0014] Further, the correlation matrix corresponding to the student model is represented as follows:

[0015]

[0016] The correlation matrix corresponding to the teacher model is represented as follows:

[0017]

[0018] wherein, the feature of the bth sample extracted by the student model, the feature of the bth sample extracted by the teacher model, where denotes concatenating all features along the channel direction, B denotes the number of samples in a batch, R denotes the set of real numbers, and HW denotes the product of parameters H and W.

[0019] Further, in the loss calculation during the student model training process, the sample features obtained by the teacher model and the student model in a training batch are respectively subjected to multi-scale feature extraction, the extracted feature representations of several different scales are spliced, and the correlation matrix calculation is performed based on the spliced result.

[0020] Further, the multi-scale feature extraction adopts a spatial pyramid pooling method, and different scale feature representations are obtained by using different scale average pooling layers.

[0021] Further, the correlation matrix corresponding to the student model is represented as follows:

[0022] ;

[0023] ;

[0024] The correlation matrix corresponding to the teacher model is represented as follows:

[0025] ;

[0026] ;

[0027] wherein, 、 are the correlation matrix representations corresponding to the optimized student model and the teacher model respectively, is a splicing matrix obtained according to different scale features extracted from sample features corresponding to the student model, is a splicing matrix obtained according to different scale features extracted from sample features corresponding to the teacher model, SPP denotes spatial pyramid pooling, denotes concatenating all features along the channel direction, denotes the feature of the bth sample extracted by the student model, denotes the feature of the bth sample extracted by the teacher model, B denotes the number of samples in a batch, R denotes the set of real numbers, and HW denotes the product of parameters H and W, denote the height and width of each sample feature extracted by the student model and the teacher model respectively.

[0028] Further, the loss calculation during the student model training process is specifically represented as follows:

[0029]

[0030] wherein, denotes an optimized loss function adopted by student model training, denotes a student model, LR denotes a low-resolution medical image, HR denotes a corresponding high-resolution medical image, ||·||1 denotes an L1 norm, K is a dimension of a splicing matrix obtained based on different scale features extracted from sample features corresponding to the student model and the teacher model, denotes a number of samples in a batch, denotes a softmax function normalization, denotes a KL divergence loss function, 、 denote correlation matrices corresponding to the optimized student model and the teacher model respectively, i, j denote coordinate positions of elements in the correlation matrix, is an adjustment parameter, and KB denotes a product of the parameters K and B.

[0031] Further, in the training of the teacher model, a low-resolution medical image in a medical image training set is taken as input, and a corresponding high-resolution medical image in the medical image training set is taken as supervision information at an output end of the teacher model, and meanwhile, a training process of model parameters is supervised by using an L1 loss function.

[0032] Further, the teacher model adopts a SwinIR model, and the student model adopts a sub-pixel convolutional neural network including a plurality of convolutional layers.

[0033] According to a second aspect of an embodiment of the present application, a medical image super-resolution reconstruction system based on knowledge distillation is provided, comprising:

[0034] a medical image acquisition unit configured to acquire a low-resolution medical image to be reconstructed;

[0035] a super-resolution reconstruction unit configured to take the low-resolution medical image as input of a pre-trained neural network-based super-resolution reconstruction model to obtain a reconstruction result of the medical image; wherein the training of the super-resolution reconstruction model adopts a knowledge distillation manner, and a training process thereof is as follows: a teacher model and a student model based on a neural network are constructed respectively; the teacher model is independently trained based on a pre-constructed medical image training set to obtain a trained teacher model; the student model is trained based on the pre-constructed medical image training set and the trained teacher model, and the trained student model is taken as the super-resolution reconstruction model; in the training process of the student model, the student model is driven to learn a correlation representation of the teacher model through a correlation matrix between all sample features in a same training batch of the teacher model and the student model.

[0036] The above one or more technical solutions have the following beneficial effects:

[0037] (1) The scheme provides a medical image super-resolution reconstruction method and system based on knowledge distillation, which transmits the inter-sample and multi-scale correlation knowledge learned by the teacher model to the student model through cross-sample multi-scale knowledge transmission, thereby enhancing the learning ability of the student model to key features, so as to obtain a reconstructed image more consistent with the details of the original image; at the same time, the scheme fully utilizes the knowledge of the teacher model to guide the training of the student model, can construct a high-performance lightweight medical image super-resolution reconstruction student model, and effectively reduces the required computing resources when deploying the model.

[0038] (2) In the loss calculation in the student model training process, the sample features obtained by the teacher model and the student model in a training batch are subjected to multi-scale feature extraction respectively, the extracted feature representations of several different scales are spliced, and the calculation of the correlation matrix is performed based on the splicing result, thereby effectively reducing the calculation amount of the loss function while ensuring the accuracy of the model training, and improving the training efficiency of the model.

[0039] The advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0040] The drawings accompanying the specification of the application form part of the application and serve to provide further understanding of the application, the illustrative embodiments of the application and their description serve to explain the application without constituting an inappropriate limitation thereof.

[0041] Figure 1 The medical image super-resolution reconstruction method based on knowledge distillation described in the embodiments of the application is a flowchart;

[0042] Figure 2 The training flowchart of the network model in the medical image super-resolution reconstruction method based on knowledge distillation described in the embodiments of the application is a flowchart;

[0043] Figure 3 The overall network model framework schematic diagram of the medical image super-resolution reconstruction method based on knowledge distillation described in the embodiments of the application is a flowchart;

[0044] Figure 4 The loss calculation optimization processing schematic diagram described in the embodiments of the application is a flowchart;

[0045] Figure 5 The medical image super-resolution reconstruction result schematic diagram described in the embodiments of the application is a flowchart;

[0046] Figure 6 The medical image super-resolution reconstruction system structure schematic diagram described in the embodiments of the application is a flowchart. DETAILED DESCRIPTION

[0047] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0048] It should be noted that the terms used herein are merely for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application.

[0049] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0050] In one or more embodiments, as shown in Figure 1 and Figure 2 The present embodiment provides a medical image super-resolution reconstruction method based on knowledge distillation, including the following processing process:

[0051] Obtain a low-resolution medical image to be reconstructed;

[0052] Take the low-resolution medical image as the input of a pre-trained neural network-based super-resolution reconstruction model to obtain the reconstruction result of the medical image; wherein the training of the super-resolution reconstruction model adopts a knowledge distillation manner, and the training process is as follows:

[0053] Respectively construct a teacher model and a student model based on a neural network;

[0054] Based on the pre-constructed medical image training set, the teacher model is independently trained to obtain a trained teacher model;

[0055] Based on the pre-constructed medical image training set and the trained teacher model, the student model is trained, and the trained student model is used as a super-resolution reconstruction model; in the training process of the student model, the correlation matrix between the teacher model and the student model in all sample features in the same training batch drives the student model to learn the correlation representation of the teacher model.

[0056] As shown in Figure 3 , a schematic diagram of the overall network architecture of a medical image super-resolution reconstruction method based on knowledge distillation is shown, in combination with Figure 3 It can be found that the network architecture of the scheme described in the present embodiment includes a teacher model and a student model, although the separate teacher model and student model can both separately complete the reconstruction of the medical image, but there is a big difference in the image quality reconstructed by the two and the size of the network model. The purpose of the scheme described in the present embodiment is to transfer the knowledge of the teacher model to the student model, so that the small-size student model has stronger image reconstruction capability.

[0057] In a specific implementation, the model training of the scheme described in this embodiment includes two stages:

[0058] Stage one: training of the teacher model

[0059] In the scheme described in this embodiment, the teacher model selects an existing SwinIR model (which is a model in the prior art and will not be described here), which has strong performance but is large in size (has 930.82K parameters, where K represents a unit of thousand) and requires a large amount of computing resources, and is difficult to be directly applied to some devices with limited computing resources (for example, image acquisition devices).

[0060] Specifically, in the scheme described in this embodiment, the teacher model includes a plurality of RSTB (Residual SwinTransformer Block) modules, which are the core modules in the SwinIR model and combine the advantages of Swin Transformer and residual connection.

[0061] In the training process of the above-mentioned teacher model, the input is a low-resolution (LR: Low Resolution) medical image in a pre-constructed medical image training set, and the output end of the model uses a high-resolution (HR: High Resolution) medical image corresponding to the training set as supervision information, and uses an L1 loss function to supervise the training process of the parameters in the model, which is expressed as follows:

[0062]

[0063] Where f(·) represents the teacher model (in this embodiment, the SwinIR model), LR represents the low-resolution medical image, HR represents the corresponding high-resolution medical image, and ||·||1 represents the L1 norm. The main purpose of this stage is to obtain a teacher model with stronger feature expression ability and reconstruction ability to guide the learning process of the student model in the second stage.

[0064] Stage two: training of the student model

[0065] In the scheme described in this embodiment, the student model adopts a sub-pixel convolutional neural network (Sub-Pixel Convolutional Neural Network) model including a plurality of convolutional layers, which has a small number of parameters (the number of parameters is 81.87K, where K represents a unit of thousand). In this stage, the parameters of the teacher model need to be pre-frozen, and only the student model is trained.

[0066] In the training process, first, as with the teacher model, an L1 loss function is applied at the output end of the student model to supervise its training. Second, traditional knowledge distillation methods usually only focus on the feature representation within a single sample, while the training process of neural networks is usually carried out in batches (mini-batch) step by step, and the ultimate goal is to model the spatial distribution structure of all sample features. In order to enable the student network to learn the cross-sample feature correlation between the multi-modal features within each small batch of the teacher network, so as to transfer the cross-sample spatial relationship modeled by the teacher to the student, so as to ensure that the data distribution in the feature mapping space has a high degree of consistency. The scheme described in the embodiment drives the student model to learn the correlation representation of the teacher model through the correlation matrix between all sample features in the same training batch of the teacher model and the student model during the training process of the student model. The use of the correlation matrix is specifically reflected in the use of the following cross-sample-based sample feature relationship distillation loss function in the training of the student model:

[0067] ;

[0068] wherein, represents the loss function used in the training of the student model, represents the student model, LR represents a low-resolution medical image, and HR represents a corresponding high-resolution medical image, and ||·||1 represents an L1 norm, respectively represent the number of samples in a batch, the height and width of each sample feature extracted by the student model and the teacher model, represents a softmax function normalization, represents a KL divergence loss function, and respectively represent the correlation matrix between all sample features in a training batch of the student model and the teacher model, i and j represent the coordinate positions of the elements in the correlation matrix, is an adjustment parameter, R represents a set of real numbers, and BHW represents the product of the three parameters B, H and W.

[0069] Further, the correlation matrix corresponding to the student model is represented as follows:

[0070]

[0071] The correlation matrix corresponding to the teacher model is represented as follows:

[0072]

[0073] wherein, represents the feature of the bth sample extracted by the student model, represents the feature of the bth sample extracted by the teacher model, and are reshaped from the original three-dimensional form ( ) into a two-dimensional matrix form ( ), C represents the number of channels, so as to facilitate the correlation calculation, represents that all features are concatenated along the channel direction, B represents the number of samples in a batch, R represents a set of real numbers, and HW represents the product of parameters H and W.

[0074] Further, the loss function described above needs to calculate a correlation matrix with a size of , and the size of the matrix exponentially increases with the size of the input image, which undoubtedly leads to an explosive increase in the amount of calculation and the amount of temporary storage occupied by the matrix during the model training process. In order to alleviate this problem, as shown in Figure 4 , the scheme described in the embodiment optimizes the loss function, which is specifically reflected in:

[0075] In the loss calculation in the student model training process, the sample features obtained by the teacher model and the student model in a training batch are subjected to multi-scale feature extraction, the feature representations of several different scales extracted are spliced, and the calculation of the correlation matrix is performed based on the spliced results.

[0076] In specific implementation, the multi-scale feature extraction adopts a spatial pyramid pooling method, and different scale feature representations are obtained by using different scale average pooling layers.

[0077] Specifically, for the extracted features and , spatial pyramid average pooling (SPP) is first performed to reduce to an arbitrary spatial size. In the embodiment, different average pooling layers are used to reduce the input features to four different sizes (it can be understood that the sizes here are not specifically limited, and C represents the number of channels here. Those skilled in the art can select an appropriate number of different scales, such as five, six, etc., according to actual needs). Subsequently, the features of the four sizes are concatenated together along the channel direction to form a feature matrix and :

[0078] Subsequently, the four sizes of features of each sample are combined into a feature matrix with a dimension of , and the feature matrix corresponding to each sample is concatenated along the channel direction to form and :

[0079]

[0080]

[0081] wherein, is a concatenation matrix obtained according to different scale features extracted from the sample features corresponding to the student model, is a concatenation matrix obtained according to different scale features extracted from the sample features corresponding to the teacher model, SPP represents spatial pyramid pooling, B represents the number of samples in a batch, and KB represents the product of parameters K and B.

[0082] It should be noted that K is the dimension of the concatenation matrix obtained based on different scale features extracted from the sample features corresponding to the student model and the teacher model, which is related to the size of the selected average pooling layer. For example, in the present embodiment, .

[0083] Based on the above feature matrix, the correlation matrix corresponding to the optimized student model and the teacher model can be obtained, and is specifically represented as follows:

[0084]

[0085] Obviously, the size of the correlation matrix corresponding to the optimized student model and the teacher model is obviously smaller than that before optimization, and the matrix size is not affected by the size of the input feature and is fixed to a specific dimension.

[0086] Finally, the loss calculation in the student model training process is specifically represented as follows:

[0087]

[0088] wherein, represents the optimized loss function used in the student model training, represents the student model, LR represents a low-resolution medical image, HR represents a corresponding high-resolution medical image, ||·||1 represents an L1 norm, K is the dimension of the concatenation matrix obtained based on different scale features extracted from the sample features corresponding to the student model and the teacher model, represents the number of samples in a batch, represents the normalization of a softmax function, represents a KL divergence loss function, , respectively represent the correlation matrix corresponding to the optimized student model and the teacher model, i and j represent the coordinate positions of the elements in the correlation matrix, is an adjustment parameter, R represents a set of real numbers, and KB represents the product of parameters K and B.

[0089] Further, as shown in Table 1, the calculation amount of the two loss calculation methods is compared:

[0090] Table 1 Comparison of calculation amount of two loss calculation methods

[0091]

[0092] In order to further show the effectiveness of the scheme described in the embodiment, the embodiment performs experimental verification on the scheme on the IXISR (Information Extraction from Images Super-Resolution) medical image dataset (this dataset is a commonly used dataset in the field, and therefore will not be described in detail here). The experimental results are shown in Table 2:

[0093] Table 2 Experimental verification results

[0094]

[0095] At the same time, in combination with Figure 5 It can be found that the image structure after being processed by the scheme described in the embodiment is clearer, significantly enhances the visual quality and anatomical recognition of the image, and can help doctors to more accurately observe micro lesions, blood vessel morphology and abnormal areas, especially for the identification of early tumors, microcalcification and other difficult pathologies.

[0096] In one or more embodiments, corresponding to the above method, as Figure 6 shown, the embodiment provides a medical image super-resolution reconstruction system based on knowledge distillation, comprising:

[0097] a medical image acquisition unit configured to acquire a low-resolution medical image to be reconstructed;

[0098] a super-resolution reconstruction unit configured to take the low-resolution medical image as an input of a pre-trained neural network-based super-resolution reconstruction model, and obtain a reconstruction result of the medical image; wherein the training of the super-resolution reconstruction model adopts a knowledge distillation manner, and the training process is as follows: a neural network-based teacher model and a student model are constructed respectively; the teacher model is independently trained based on a pre-constructed medical image training set, and a trained teacher model is obtained; the student model is trained based on the pre-constructed medical image training set and the trained teacher model, and the trained student model is taken as the super-resolution reconstruction model; in the training process of the student model, the student model is driven to learn the correlation representation of the teacher model through the correlation matrix between all sample features of the teacher model and the student model in the same training batch.

[0099] It can be understood that the related technical details of the system described in the embodiment have been described in detail in Embodiment One, and therefore will not be described here.

[0100] The above only describes the preferred embodiments of the present disclosure and is not intended to limit the present disclosure. Those skilled in the art can make various modifications and changes to the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A medical image super-resolution reconstruction method based on knowledge distillation, characterized by: include: Acquiring a low-resolution medical image to be reconstructed; A low-resolution medical image is used as input to a pre-trained neural network-based super-resolution reconstruction model to obtain a reconstruction result of the medical image. The super-resolution reconstruction model is trained using knowledge distillation, and the training process is as follows: Construct a teacher model and a student model based on neural networks respectively; Based on the pre-built medical image training set, the teacher model is independently trained to obtain a trained teacher model; Based on a pre-constructed medical image training set and a trained teacher model, the student model is trained and used as the super-resolution reconstruction model. During the training of the student model, the correlation matrix between all sample features of the teacher model and the student model in the same training batch is used to drive the student model to learn the correlation representation of the teacher model.

2. The medical image super-resolution reconstruction method based on knowledge distillation according to claim 1, characterized in that: The student model is trained using the following loss function: ； in, represents the loss function used in student model training, represents the student model, LR represents the low-resolution medical image, HR represents the corresponding high-resolution medical image, ||·||1 represents the L1 norm, Represents the number of samples in a batch, the height and width of each sample feature extracted by the student model and the teacher model, respectively. Indicates the normalization of the softmax function, represents the KL divergence loss function, and , respectively represent the correlation matrix between all sample features of the student model and the teacher model in a training batch, i, j represent the coordinate position of the elements in the correlation matrix, To adjust the parameters, R represents a set of real numbers, and BHW represents the product of three parameters B, H and W.

3. The medical image super-resolution reconstruction method based on knowledge distillation according to claim 2, characterized in that: The correlation matrix corresponding to the student model is expressed as follows: ； The correlation matrix corresponding to the teacher model is expressed as follows: ； in, represents the features of the bth sample extracted by the student model, represents the features of the b-th sample extracted by the teacher model, Indicates that all features are concatenated along the channel direction, B represents the number of samples in a batch, R represents a set of real numbers, and HW represents the product of parameters H and W.

4. The medical image super-resolution reconstruction method based on knowledge distillation according to claim 1, characterized in that: In the loss calculation during the student model training process, multi-scale feature extraction is performed on the sample features obtained by the teacher model and the student model in a training batch, respectively, the extracted feature representations of several different scales are spliced, and the correlation matrix is ​​calculated based on the splicing results.

5. The medical image super-resolution reconstruction method based on knowledge distillation according to claim 4, characterized in that: The multi-scale feature extraction adopts a spatial pyramid pooling method, and obtains feature representations of different scales by adopting average pooling layers of different scales.

6. The medical image super-resolution reconstruction method based on knowledge distillation according to claim 4, characterized in that: The correlation matrix corresponding to the student model is expressed as follows: ； ； The correlation matrix corresponding to the teacher model is expressed as follows: ； ； in, 、 are the correlation matrices corresponding to the optimized student model and teacher model, is the concatenation matrix obtained by extracting different scale features based on the sample features corresponding to the student model, is the concatenation matrix obtained by extracting features of different scales according to the sample features corresponding to the teacher model. SPP represents spatial pyramid pooling. Indicates that all features are connected in series along the channel direction. represents the features of the bth sample extracted by the student model, represents the feature of the bth sample extracted by the teacher model, B represents the number of samples in a batch, R represents a set of real numbers, HW represents the product of parameters H and W, Represent the height and width of each sample feature extracted by the student model and the teacher model respectively.

7. The medical image super-resolution reconstruction method based on knowledge distillation according to claim 6, characterized in that: The loss calculation during student model training is specifically expressed as follows: ； in, represents the optimized loss function used in student model training, represents the student model, LR represents the low-resolution medical image, HR represents the corresponding high-resolution medical image, ||·||1 represents the L1 norm, and K is the dimension of the splicing matrix obtained by extracting different scale features based on the sample features corresponding to the student model and the teacher model. represents the number of samples in a batch, Indicates the normalization of the softmax function, represents the KL divergence loss function, 、 are the correlation matrices corresponding to the optimized student model and teacher model, i and j represent the coordinate positions of the elements in the correlation matrix, For parameter adjustment, KB represents the product of parameters K and B.

8. The medical image super-resolution reconstruction method based on knowledge distillation according to claim 1, characterized in that: In the training of the teacher model, low-resolution medical images in the medical image training set are used as input, and the corresponding high-resolution medical images in the medical image training set are used as supervision information at the output end of the teacher model. At the same time, the L1 loss function is used to supervise the training process of the model parameters.

9. The medical image super-resolution reconstruction method based on knowledge distillation according to claim 1, characterized in that: The teacher model adopts the SwinIR model, and the student model adopts a sub-pixel convolutional neural network including several convolutional layers.

10. A medical image super-resolution reconstruction system based on knowledge distillation, characterized by: include: a medical image acquisition unit, configured to acquire a low-resolution medical image to be reconstructed; A super-resolution reconstruction unit is used to use low-resolution medical images as input to a pre-trained neural network-based super-resolution reconstruction model to obtain a reconstruction result of the medical image; wherein, the training of the super-resolution reconstruction model adopts a knowledge distillation method, and its training process is as follows: constructing a neural network-based teacher model and a student model respectively; independently training the teacher model based on a pre-constructed medical image training set to obtain a trained teacher model; training the student model based on the pre-constructed medical image training set and the trained teacher model, and using the trained student model as a super-resolution reconstruction model; during the training process of the student model, the student model is driven to learn the correlation representation of the teacher model through the correlation matrix between the teacher model and all sample features of the student model in the same training batch.

Images