A Method and System for Medical Image Super-Resolution Reconstruction Based on Knowledge Distillation

Abstract

This invention provides a method and system for medical image super-resolution reconstruction based on knowledge distillation, belonging to the field of medical image processing technology. The scheme transfers knowledge learned by the teacher model across samples and at multiple scales to the student model, thereby enhancing the student model's ability to learn key features and obtain a reconstructed image that is more consistent with the details of the original image. At the same time, the scheme makes full use of the teacher model's knowledge to guide the training of the student model, which can build a high-performance and lightweight medical image super-resolution reconstruction student model. While enhancing the visual quality and recognizability of the reconstructed image, it effectively reduces the computational resources required for model deployment.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, and in particular relates to a method and system for super-resolution reconstruction of medical images based on knowledge distillation. Background Technology

[0002] With the development of precision medicine, the resolution of medical images directly affects the early detection of lesions and surgical planning. Traditional equipment such as MRI (Magnetic Resonance Imaging) and CT (Computed Tomography) are limited by hardware costs and radiation safety, often outputting low-resolution images, resulting in a missed diagnosis rate of up to 32% for tiny tumors (smaller than 3mm). Based on the idea that super-resolution reconstruction technology can effectively overcome physical limitations through software algorithms, some researchers have considered using neural network models for super-resolution reconstruction of low-resolution medical images. However, existing methods mainly focus on improving model performance, thus using increasingly larger neural network models, ignoring the actual cost of using the model, making it difficult to apply in medical equipment with limited computing resources (e.g., image acquisition equipment).

[0003] Secondly, some researchers have considered reducing the size of the model by compressing the neural network model (e.g., knowledge distillation). However, traditional knowledge distillation methods usually only focus on the feature representation within a single sample, seriously ignoring the spatial distribution structure of sample features. This results in the student model being unable to effectively learn the cross-sample spatial relationships in the teacher model, leading to the generally poor image reconstruction effect of the student model obtained based on traditional knowledge distillation methods, which cannot meet the actual needs. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention provides a medical image super-resolution reconstruction method and system based on knowledge distillation, which effectively solves the problem that traditional reconstruction methods cannot be applied to medical devices with limited computing resources due to the large model size, and that the reconstruction effect of student model images obtained by traditional knowledge distillation methods is generally not satisfactory to actual needs.

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

[0006] Acquire low-resolution medical images to be reconstructed;

[0007] Low-resolution medical images are used as input to a pre-trained neural network-based super-resolution reconstruction model to obtain the reconstruction results of the medical images. The super-resolution reconstruction model is trained using knowledge distillation, and the training process is as follows:

[0008] Construct teacher and student models based on neural networks respectively;

[0009] Based on a pre-built medical image training set, the teacher model is trained independently to obtain a well-trained teacher model;

[0010] Based on a pre-built 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 process of the student model, the correlation matrix between all sample features in the same training batch of the teacher model and the student model drives the student model to learn the correlation representation of the teacher model.

[0011] Furthermore, the student model is trained using the following loss function:

[0012]

[0013] in, This represents the loss function used for training the student model. The model represents the student model, LR represents the low-resolution medical image, and HR represents the corresponding high-resolution medical image. Describing the L1 norm, These represent the number of samples in a batch, and the height and width of each sample feature extracted by the student model and teacher model, respectively. This indicates that the softmax function normalizes the data. Denotes the KL divergence loss function. and , represent the correlation matrices between all sample features of the student model and the teacher model in a training batch, respectively, where i, j represent the coordinate positions of elements in the correlation matrix. To adjust the parameters, R represents the set of real numbers, and BHW represents the product of the three parameters B, H, and W.

[0014] Furthermore, 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] in, This represents the feature extracted from the b-th sample by the student model. This represents the feature of the b-th sample extracted by the teacher model. This indicates that all features are concatenated along the channel direction, B represents the number of samples in a batch, R represents the set of real numbers, and HW represents the product of parameters H and W.

[0019] Furthermore, 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, the extracted feature representations of several different scales are concatenated, and the correlation matrix is ​​calculated based on the concatenation result.

[0020] Furthermore, the multi-scale feature extraction employs the spatial pyramid pooling method, which uses average pooling layers of different scales to obtain feature representations at different scales.

[0021] Furthermore, 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] in, , The correlation matrices for the optimized student model and teacher model are shown below. This is a concatenated matrix obtained by extracting features of different scales from the sample features corresponding to the student model. SPP represents the concatenation matrix obtained by extracting features of different scales from the sample features corresponding to the teacher model. This indicates that all features are concatenated along the channel direction. This represents the feature extracted from the b-th sample by the student model. Let B represent the feature of the b-th sample extracted by the teacher model, B represent the number of samples in a batch, R represent the set of real numbers, and HW represent the product of parameters H and W. These represent the height and width of each sample feature extracted by the student model and the teacher model, respectively.

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

[0029]

[0030] in, This represents the optimized loss function used for training the student model. Let represent the student model, LR represent the low-resolution medical image, HR represent the corresponding high-resolution medical image, ||·||1 represents the L1 norm, and K is the dimension of the concatenation matrix obtained based on the sample features extracted from the student model and the teacher model at different scales. This indicates the number of samples in a batch. This indicates that the softmax function normalizes the data. Denotes the KL divergence loss function. , Let be the correlation matrix representations of the optimized student model and teacher model, respectively, where i and j represent the coordinate positions of elements in the correlation matrix. To adjust the parameters, KB represents the product of parameters K and B.

[0031] Furthermore, the training of the teacher model uses low-resolution medical images from the medical image training set as input, and high-resolution medical images from the corresponding medical image training set as supervision information at the output of the teacher model. At the same time, the L1 loss function is used to supervise the training process of the model parameters.

[0032] Furthermore, the teacher model adopts the SwinIR model, and the student model adopts a subpixel convolutional neural network including several convolutional layers.

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

[0034] A medical image acquisition unit is used to acquire low-resolution medical images to be reconstructed.

[0035] A super-resolution reconstruction unit is used to obtain the reconstruction result of a medical image by using a low-resolution medical image as input to a pre-trained neural network-based super-resolution reconstruction model. The super-resolution reconstruction model is trained using knowledge distillation, and the training process is as follows: A teacher model and a student model based on a neural network are constructed separately; 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 used as the super-resolution reconstruction model; during the training of the student model, the correlation matrix between all sample features in the same training batch between the teacher model and the student model drives the student model to learn the correlation representation of the teacher model.

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

[0037] (1) The present invention provides a medical image super-resolution reconstruction method and system based on knowledge distillation. The scheme transfers the knowledge of inter-sample and multi-scale associations learned by the teacher model to the student model through cross-sample multi-scale knowledge transfer, thereby enhancing the student model's ability to learn key features and obtain a reconstructed image that is more consistent with the details of the original image. At the same time, the scheme makes full use of the knowledge of the teacher model to guide the training of the student model, and can build a high-performance and lightweight medical image super-resolution reconstruction student model, effectively reducing the computing resources required for model deployment.

[0038] (2) In the loss calculation during the training process of the student model, the proposed scheme performs multi-scale feature extraction on the sample features obtained by the teacher model and the student model in a training batch, splices the extracted feature representations of several different scales, and performs the calculation of the correlation matrix based on the splicing result. While ensuring the accuracy of model training, it effectively reduces the amount of calculation of the loss function and improves the training efficiency of the model.

[0039] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0040] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0041] Figure 1 This is a flowchart of the medical image super-resolution reconstruction method based on knowledge distillation described in this embodiment of the invention;

[0042] Figure 2 This is a flowchart illustrating the training process of the network model in the knowledge distillation-based medical image super-resolution reconstruction method described in this embodiment of the invention.

[0043] Figure 3 This is a schematic diagram of the overall network model framework of the knowledge distillation-based medical image super-resolution reconstruction method described in this embodiment of the invention.

[0044] Figure 4 This is a schematic diagram of the loss calculation optimization process described in the embodiments of the present invention;

[0045] Figure 5 This is a schematic diagram of the medical image super-resolution reconstruction results described in an embodiment of the present invention;

[0046] Figure 6 This is a schematic diagram of the medical image super-resolution reconstruction system based on knowledge distillation as described in an embodiment of the present invention. Detailed Implementation

[0047] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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 invention pertains.

[0048] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0049] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

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

[0051] Acquire low-resolution medical images to be reconstructed;

[0052] Low-resolution medical images are used as input to a pre-trained neural network-based super-resolution reconstruction model to obtain the reconstruction results of the medical images. The super-resolution reconstruction model is trained using knowledge distillation, and the training process is as follows:

[0053] Construct teacher and student models based on neural networks respectively;

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

[0055] Based on a pre-built 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 process of the student model, the correlation matrix between all sample features in the same training batch of the teacher model and the student model drives the student model to learn the correlation representation of the teacher model.

[0056] like Figure 3 The diagram illustrates the overall network architecture of a knowledge distillation-based medical image super-resolution reconstruction method. Figure 3 It can be seen that the network architecture of the scheme described in this embodiment includes a teacher model and a student model. Although both the teacher model and the student model can reconstruct medical images independently, there are significant differences in the image quality and the size of the network model. The purpose of the scheme described in this embodiment is to transfer the knowledge of the teacher model to the student model, so that the smaller student model has a stronger image reconstruction capability.

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

[0058] Phase 1: Training of Teacher Models

[0059] In the scheme described in this embodiment, the teacher model selected is the existing SwinIR model (this model is an existing model in this field, and will not be described in detail here). This model has powerful performance but large size (with 930.82K parameters, where K represents the unit kilo), which has a large demand for computing resources and is difficult to apply directly to some computing resource-constrained devices (e.g., image acquisition devices);

[0060] Specifically, in the scheme described in this embodiment, the teacher model includes several RSTB (Residual SwinTransformer Block) modules, the core module of which combines the advantages of Swin Transformer and residual connection.

[0061] During the training process of the aforementioned teacher model, its input consists of low-resolution (LR) medical images from a pre-constructed medical image training set. The model's output uses high-resolution (HR) medical images from the corresponding training set as supervisory information, and employs an L1 loss function to supervise the training process of the model's parameters. This loss function is expressed as follows:

[0062]

[0063] Where f(·) represents the teacher model (SwinIR model in this embodiment), 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 and reconstruction capabilities, which will be used to guide the learning process of the student model in the second stage.

[0064] Phase Two: Training the Student Model

[0065] In the scheme described in this embodiment, the student model adopts a sub-pixel convolutional neural network model including several convolutional layers. This model has a small number of parameters (81.87K parameters, where K represents thousands). At this stage, it is necessary to freeze the parameters of the teacher model in advance and train only the student model.

[0066] During training, firstly, similar to the teacher model, an L1 loss function is applied to the output of the student model to supervise its training. Secondly, traditional knowledge distillation methods typically focus only on the feature representation within a single sample, while neural network training is usually performed incrementally in mini-batches, with the ultimate goal of modeling the spatial distribution structure of all sample features. To enable the student network to learn the cross-sample feature correlations between multimodal features within each mini-batch of the teacher network, thereby transferring the cross-sample spatial relationships modeled by the teacher to the student, ensuring a high degree of consistency in their data distribution within the feature mapping space, the scheme described in this embodiment drives the student model to learn the correlation representation of the teacher model during the training process by using the correlation matrix between all sample features in the same training batch between the teacher and student models. The utilization of this correlation matrix is ​​specifically reflected in the use of the following distillation loss function based on cross-sample feature relationships during student model training:

[0067] ;

[0068] in, This represents the loss function used for training the student model. Let LR represent the student model, HR represent the corresponding high-resolution medical image, and ||·||1 represent the L1 norm. These represent the number of samples in a batch, and the height and width of each sample feature extracted by the student model and teacher model, respectively. This indicates that the softmax function normalizes the data. Denotes the KL divergence loss function. and , represent the correlation matrices between all sample features of the student model and the teacher model in a training batch, respectively, where i, j represent the coordinate positions of elements in the correlation matrix. To adjust the parameters, R represents the set of real numbers, and BHW represents the product of the three parameters B, H, and W.

[0069] Furthermore, 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] in, This represents the feature extracted from the b-th sample by the student model. This represents the feature of the b-th sample extracted by the teacher model. and All were derived from the original three-dimensional form ( It was reshaped into a two-dimensional matrix form. C represents the number of channels, to facilitate correlation calculation. This indicates that all features are concatenated along the channel direction, B represents the number of samples in a batch, R represents the set of real numbers, and HW represents the product of parameters H and W.

[0074] Furthermore, the above loss function needs to calculate a value of size [missing information]. The correlation matrix, whose size increases exponentially with the size of the input image, undoubtedly leads to a surge in computational load and temporary storage requirements during model training. To mitigate this problem, such as... Figure 4 As shown, the solution described in this embodiment optimizes the aforementioned loss function, specifically in the following ways:

[0075] In the loss calculation during the training process of the student model, multi-scale feature extraction is performed on the sample features obtained by the teacher model and the student model in a training batch. The extracted feature representations at several different scales are concatenated, and the correlation matrix is ​​calculated based on the concatenation result.

[0076] In practice, the multi-scale feature extraction adopts the spatial pyramid pooling method, which obtains feature representations at different scales by using average pooling layers of different scales.

[0077] Specifically, for the extracted features and First, Spatial Pyramid Pooling (SPP) is performed to reduce the input features to any spatial size. In this embodiment, different average pooling layers are used to reduce the input features to... Four different sizes (understandably, the size here is not a specific limitation; C represents the number of channels, and those skilled in the art can choose five, six, or other appropriate sizes according to actual needs) are then used to concatenate all features along the channel direction and merge them into a single feature matrix. and :

[0078] Subsequently, the features of each sample in four dimensions are merged into a single dimension. The feature matrix is ​​formed by concatenating the feature matrices corresponding to each sample along the channel direction. and :

[0079]

[0080]

[0081] in, This is a concatenated matrix obtained by extracting features of different scales from the sample features corresponding to the student model. SPP is the concatenation matrix obtained by extracting features of different scales 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 here that K is the dimension of the concatenation matrix obtained from the sample features extracted from the student model and the teacher model at different scales, which is related to the size of the selected average pooling layer. For example, in this embodiment... .

[0083] Based on the above feature matrix, the correlation matrix corresponding to the optimized student model and teacher model can be obtained, as shown below:

[0084]

[0085] Clearly, the size of the correlation matrix corresponding to the optimized student model and teacher model is significantly smaller than that before optimization. At the same time, the size of this matrix is ​​not affected by the size of the input features and is fixed to a specific dimension.

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

[0087]

[0088] in, This represents the optimized loss function used for training the student model. Let represent the student model, LR represent the low-resolution medical image, HR represent the corresponding high-resolution medical image, ||·||1 represents the L1 norm, and K is the dimension of the concatenation matrix obtained based on the sample features extracted from the student model and the teacher model at different scales. This indicates the number of samples in a batch. This indicates that the softmax function normalizes the data. Denotes the KL divergence loss function. , Let be the correlation matrix representations of the optimized student model and teacher model, respectively, where i and j represent the coordinate positions of elements in the correlation matrix. To adjust the parameters, R represents the set of real numbers, and KB represents the product of parameters K and B.

[0089] Furthermore, as shown in Table 1, the computational complexity of the two loss calculation methods is compared:

[0090] Table 1 Comparison of computational complexity between the two loss calculation methods

[0091]

[0092] To further demonstrate the effectiveness of the solution described in this embodiment, this embodiment conducted experimental verification on the IXISR (Information extraction from Images Super-Resolution) medical image dataset (this dataset is commonly used in the field, so it 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, combined Figure 5 It can be observed that the image tissue structure is clearer after processing by the scheme described in this embodiment, which significantly enhances the visual quality and anatomical recognition of the image. This can help doctors observe small lesions, blood vessel morphology and abnormal areas more accurately, and is especially beneficial for the identification of difficult pathologies such as early tumors and microcalcifications.

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

[0097] A medical image acquisition unit is used to acquire low-resolution medical images to be reconstructed.

[0098] A super-resolution reconstruction unit is used to obtain the reconstruction result of a medical image by using a low-resolution medical image as input to a pre-trained neural network-based super-resolution reconstruction model. The super-resolution reconstruction model is trained using knowledge distillation, and the training process is as follows: A teacher model and a student model based on a neural network are constructed separately; 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 used as the super-resolution reconstruction model; during the training of the student model, the correlation matrix between all sample features in the same training batch between the teacher model and the student model drives the student model to learn the correlation representation of the teacher model.

[0099] It is understood that the relevant technical details of the system described in this embodiment have been described in detail in Embodiment 1, so they will not be repeated here.

[0100] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A medical image super-resolution reconstruction method based on knowledge distillation, characterized in that, include: Acquire low-resolution medical images to be reconstructed; Low-resolution medical images are used as input to a pre-trained neural network-based super-resolution reconstruction model to obtain the reconstruction results of the medical images. The super-resolution reconstruction model is trained using knowledge distillation, and the training process is as follows: Construct teacher and student models based on neural networks respectively; Based on a pre-built medical image training set, the teacher model is trained independently to obtain a well-trained teacher model; Based on a pre-built 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 process of the student model, the correlation matrix between all sample features in the same training batch of the teacher model and the student model drives the student model to learn the correlation representation of the teacher model. In the loss calculation during the training process of the student model, multi-scale feature extraction is performed on the sample features obtained by the teacher model and the student model in a training batch. The extracted feature representations at several different scales are concatenated, and the correlation matrix is ​​calculated based on the concatenation result. The correlation matrix corresponding to the student model is represented as follows: ； ； The correlation matrix corresponding to the teacher model is represented as follows: ； ； in, , The correlation matrices for the optimized student model and teacher model are shown below. This is a concatenated matrix obtained by extracting features of different scales from the sample features corresponding to the student model. SPP represents the concatenation matrix obtained by extracting features of different scales from the sample features corresponding to the teacher model. This indicates that all features are concatenated along the channel direction. This represents the feature extracted from the b-th sample by the student model. Let B represent the feature of the b-th sample extracted by the teacher model, B represent the number of samples in a batch, R represent the set of real numbers, and HW represent the product of parameters H and W. These represent the height and width of each sample feature extracted by the student model and the teacher model, respectively.

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

3. The medical image super-resolution reconstruction method based on knowledge distillation as described in claim 2, characterized in that, The correlation matrix corresponding to the student model is represented as follows: ； The correlation matrix corresponding to the teacher model is represented as follows: ； in, This represents the feature extracted from the b-th sample by the student model. This represents the feature of the b-th sample extracted by the teacher model. This indicates that all features are concatenated along the channel direction, B represents the number of samples in a batch, R represents the 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 as described in claim 1, characterized in that, The multi-scale feature extraction employs the spatial pyramid pooling method, which uses average pooling layers of different scales to obtain feature representations at different scales.

5. The medical image super-resolution reconstruction method based on knowledge distillation as described in claim 1, characterized in that, The loss calculation during student model training is specifically represented as follows: ； in, This represents the optimized loss function used for training the student model. Let represent the student model, LR represent the low-resolution medical image, HR represent the corresponding high-resolution medical image, ||·||1 represents the L1 norm, and K is the dimension of the concatenation matrix obtained based on the sample features extracted from the student model and the teacher model at different scales. This indicates the number of samples in a batch. This indicates that the softmax function normalizes the data. Denotes the KL divergence loss function. , Let be the correlation matrix representations of the optimized student model and teacher model, respectively, where i and j represent the coordinate positions of elements in the correlation matrix. To adjust the parameters, KB represents the product of parameters K and B.

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

7. The medical image super-resolution reconstruction method based on knowledge distillation as described in claim 1, characterized in that, The teacher model uses the SwinIR model, and the student model uses a subpixel convolutional neural network that includes several convolutional layers.

8. A medical image super-resolution reconstruction system based on knowledge distillation, which is based on the medical image super-resolution reconstruction method based on knowledge distillation as described in any one of claims 1-7, characterized in that, include: A medical image acquisition unit is used to acquire low-resolution medical images to be reconstructed. A super-resolution reconstruction unit is used to obtain the reconstruction result of a medical image by using a low-resolution medical image as input to a pre-trained neural network-based super-resolution reconstruction model. The super-resolution reconstruction model is trained using knowledge distillation, and the training process is as follows: A teacher model and a student model based on a neural network are constructed separately; 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 used as the super-resolution reconstruction model; during the training of the student model, the correlation matrix between all sample features in the same training batch between the teacher model and the student model drives the student model to learn the correlation representation of the teacher model.

Images