A super-resolution reconstruction method for MRI medical images

Through the combination of a hybrid intensive knowledge distillation model and a multi-value mapping model, the problems of artifacts and edge blur in super-resolution reconstruction of MRI medical images are solved, and higher quality image reconstruction is achieved.

CN117522688BActive Publication Date: 2024-11-19JIANGXI UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311467736.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-07
Publication Date
2024-11-19
Estimated Expiration
2043-11-07

AI Technical Summary

Technical Problem

Existing MRI medical image super-resolution reconstruction techniques are difficult to effectively remove artifacts, especially when merging information at boundaries, and traditional methods ignore the potential correlation between reconstruction and super-resolution tasks, resulting in image quality degradation and edge blur.

Method used

Using a hybrid intensive knowledge distillation model, the super-resolution knowledge is transmitted through the interaction between teacher model and student model, knowledge distillation technology is used to transfer super-resolution knowledge, combined with multi-value mapping model and Laplace distribution weighted fusion, improve the boundary reconstruction quality of MRI medical images, and Solve the problem of discomfort.

Benefits of technology

The boundary reconstruction quality of MRI medical images is significantly improved, and the blurring and disposition of image edges is solved, allowing clearer and realistic super-resolution reconstruction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117522688B_ABST
    Figure CN117522688B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for super-resolution reconstruction of MRI medical images, comprising: obtaining an MRI medical image with artifacts and under-sampling, constructing a hybrid dense knowledge distillation model, the hybrid dense knowledge distillation model comprising: a teacher model and a student model; training the hybrid dense knowledge distillation model, inputting the MRI medical image with artifacts and under-sampling into the trained hybrid dense knowledge distillation model, and obtaining an artifact-free fully sampled MRI medical image. The present invention finds through analysis that a hybrid dense network can effectively parse multiple components existing in the network output, and constructs a multi-value mapping model based on multiple components to reconstruct sharp boundaries, while the regularization constraints obtained from knowledge distillation effectively solve the ill-posed problem in the super-resolution task, and designs a hybrid dense knowledge distillation model based on the connection and complementarity between the analysis reconstruction and super-resolution subtasks to better complete the MRI super-resolution reconstruction task.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of image super-resolution reconstruction, and in particular to a super-resolution reconstruction method applied to MRI medical images. Background Art

[0002] Magnetic resonance imaging (MRI) has become an important auxiliary diagnostic technology in medical screening. However, the initial MRI medical images of rapid scanning imaging are often undersampled images with artifacts. How to eliminate artifacts in MRI medical images and obtain fully sampled high-definition images is one of the main research topics in this field. The MRI medical image super-resolution reconstruction task includes two subtasks, namely (1) reconstruction task, the goal is to remove artifacts in MRI; (2) super-resolution task, the goal is to upsample the undersampled (low-resolution image, LR) to obtain a fully sampled image (high-resolution image, HR).

[0003] Traditional medical image reconstruction techniques, such as sequence-based, parallel imaging and compressed sensing-based methods, have certain limitations. They increase the complexity of sequence design and hardware cost, and may lead to a decrease in image quality and produce discontinuous reconstruction results. Reconstruction methods based on convolutional neural networks (CNNs) have the ability to learn the mapping relationship of reconstructed images from a large amount of training data, which significantly improves the quality of reconstruction results. At the same time, CNN-based methods such as SRCNN and FSRCNN have also achieved better restoration effects than traditional interpolation methods in image super-resolution tasks. The EDSR network uses a residual scaling method to improve the quality of image restoration. On this basis, IMDN uses iterative refinement and multi-network fusion strategies to improve the level of detail and naturalness of the resulting images. These network models improve the super-resolution performance of the model by increasing the number of network residual blocks and the number of convolution kernels, which also induces more parameters and computational complexity. These works all enable the network model to learn the two tasks of super-resolution and reconstruction separately. This strategy may result in the loss of information that is beneficial to improving the indicators of a single task during the learning process.

[0004] Deep neural networks introduce smoothing biases that allow them to effectively fit single-valued mappings. However, they have poor fitting capabilities for multi-valued mappings, resulting in reconstructed images with blurred edges. For example, minimizing the error between predictions and targets using mean squared error (MSE) or mean absolute error (MAE) in regression helps achieve a global minimum MSE or MAE. However, it is not guaranteed to be suitable for every sample. In cases where the output has multiple possible values, images generated by single-valued mappings tend to appear blurred. Boundary reconstruction of MRI medical images is challenging due to the need for multi-valued mappings, a process that involves merging information from different targets at the boundary, such as treating one part as the foreground and the rest as the background.

[0005] In the super-resolution task, an LR image can be upsampled to multiple different HR images, so MRI super-resolution is a typical ill-posed problem. Mainstream methods usually only focus on reconstruction or super-resolution sub-tasks, while ignoring the potential correlation between them. Summary of the invention

[0006] The purpose of the present invention is to provide a super-resolution reconstruction method for MRI medical images to solve the problems existing in the above-mentioned prior art. Through analysis, it is found that a hybrid dense network (MDN) can effectively parse the multiple components existing in the network output, and constructing a multi-value mapping model based on multiple components can better reconstruct the sharp boundaries. At the same time, the regularization constraints obtained from knowledge distillation can effectively solve the ill-posed problem in the super-resolution task. On the basis of the certain connection and complementarity between the analysis and reconstruction and super-resolution subtasks, a hybrid dense knowledge distillation model is designed to better complete the MRI super-resolution reconstruction task.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A super-resolution reconstruction method for MRI medical images, comprising:

[0009] Acquire artifact-bearing and undersampled MRI medical images, and construct a hybrid dense knowledge distillation model, wherein the hybrid dense knowledge distillation model includes: a teacher model and a student model;

[0010] The hybrid dense knowledge distillation model is trained, and the artifact-bearing and under-sampled MRI medical image is input into the trained hybrid dense knowledge distillation model to obtain an artifact-free fully sampled MRI medical image.

[0011] Optionally, the teacher model includes: an encoder, a first decoder and an upsampling block; wherein the encoder, the first decoder and the upsampling block are connected in sequence.

[0012] Optionally, training the teacher model includes:

[0013] Acquire a high-resolution image, perform data enhancement processing on the high-resolution image, input the high-resolution image after data enhancement into the encoder, and acquire compact features;

[0014] An undersampled image is obtained, the compact feature is constrained based on the undersampled image, the constrained compact feature is input into the first decoder, and the teacher model output image is obtained through the upsampling block, while the weight of the first decoder is saved.

[0015] Optionally, the student model includes: a second decoder and a mixed dense block; wherein the second decoder and the mixed dense block are connected.

[0016] Optionally, training the student model includes:

[0017] Acquire an artifact-bearing and under-sampled MRI historical medical image, assign the saved weight of the first decoder to the second decoder, input the artifact-bearing and under-sampled MRI historical medical image into the second decoder, and acquire an artifact-free fully-sampled MRI historical medical image through the mixed dense block.

[0018] Optionally, assigning the saved weight of the first decoder to the second decoder comprises:

[0019] The method of transferring the knowledge of the teacher model to the student model through knowledge distillation is:

[0020]

[0021] Among them, L dis is the distillation loss, w is the weight, H′ is the height of the high-resolution image, W′ is the width of the high-resolution image, σ is the variance, M is the total number of distillation modules, m is the mth distillation module, C is the number of channels, i is the i-th channel index, j is the j-th height pixel index, k is the k-th width pixel index, and Y T is the prediction of the teacher model, and μ is the mean of the predictions of the student model.

[0022] Optionally, the mixed dense block includes: a convolutional layer, a first distillation submodule and a second distillation submodule; wherein the weights of the first distillation submodule and the second distillation submodule are obtained through the convolutional layer.

[0023] Optionally, training the student model further includes:

[0024] Inputting the upsampled MRI historical medical image into the convolutional layer to perform weight calculation to obtain a first weight and a second weight;

[0025] Input the upsampled MRI historical medical image into the first distillation submodule and the second distillation submodule respectively to calculate the mean and variance, obtain a first mean, a second mean, a first variance, and a second variance, obtain a first Laplace distribution according to the first mean and the first variance, and obtain a second Laplace distribution according to the second mean and the second variance;

[0026] According to the first weight and the second weight, weighting the first Laplace distribution and the second Laplace distribution respectively to obtain a mixed dense distribution;

[0027] The differential entropy of the mixed dense distribution is calculated to obtain a visual aleatoric uncertainty map.

[0028] Optionally, a visual aleatoric uncertainty map can be obtained by:

[0029] h=-∫p(x i ;θ)logp(x i ;θ)dx i h=-∫p(x i ;θ)logp(x i ;θ)dx i

[0030] Where p is the Laplace distribution, h is the random uncertainty value, and x i is the i-th artifact-bearing and undersampled MRI medical image, and θ is the parameter of the hybrid dense block.

[0031] The beneficial effects of the present invention are:

[0032] The present invention models each component as a corresponding distribution (such as Gaussian distribution or Laplace distribution), and then performs weighted fusion on each distribution according to a corresponding weight value to improve the quality of MRI boundary reconstruction.

[0033] The present invention transfers the super-resolution knowledge obtained from the teacher to the student for super-resolution reconstruction, effectively solving the ill-posed problem in the MRI super-resolution task.

[0034] The present invention uses the teacher model as an auxiliary task to learn the super-resolution task, and uses the student model as the main task to perform the super-resolution reconstruction task, and uses the super-resolution knowledge obtained from the teacher to guide the student to reconstruct undersampled MRI medical images with artifacts. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0036] Figure 1 A schematic diagram of a hybrid dense network according to an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of a super-resolution reconstruction method for MRI medical images according to an embodiment of the present invention;

[0038] Figure 3 A schematic diagram of a hybrid dense module according to an embodiment of the present invention;

[0039] Figure 4 Schematic diagram of a distillation module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] The present invention discloses a super-resolution reconstruction method for MRI medical images, comprising: obtaining an MRI medical image with artifacts and under-sampled, constructing a hybrid dense knowledge distillation model, the hybrid dense knowledge distillation model comprising: a teacher model and a student model; training the hybrid dense knowledge distillation model, inputting the MRI medical image with artifacts and under-sampled into the trained hybrid dense knowledge distillation model, and obtaining an artifact-free fully sampled MRI medical image.

[0043] The present invention analyzes that a hybrid dense network (MDN) can parse multiple components in the network output to establish a multi-value mapping model, which can better reconstruct the sharp boundary.

[0044] Figure 1 The aleatory uncertainty in the prediction variance is determined by evaluating the differential entropy implicit in the continuous mixture distribution. First, this embodiment models each component as a Laplace distribution with mean μ, variance σ, and weight w, resulting in two predicted Laplace distributions f1 and f2. Then f1 and f2 are weighted to obtain the mixture dense distribution f MDN To reconstruct the fuzzy boundary, where f MDN =p(x i ;θ), x i is the i-th artifact-bearing and undersampled MRI medical image, and θ is the parameter of the hybrid dense block. Finally, by calculating f MDN The aleatoric uncertainty is calculated using the differential entropy of

[0045] Convolutional neural networks show strong fitting ability for single-value mapping, but their performance will drop significantly when facing multi-value mapping. In MRI reconstruction tasks, MSE or MAE is usually used as the loss function. However, achieving global optimality does not ensure that each pixel reaches the required fitting value. Therefore, the reconstructed MRI medical images often have the problem of blurred boundaries. The current global optimal method in MRI reconstruction cannot fully capture boundary details. The present invention utilizes the MDN module to convert the single-value mapping problem into a multi-value mapping problem, and proposes a solution, that is, modeling each component as a corresponding distribution (such as a Gaussian distribution or a Laplace distribution), and then weighted fusion of each distribution according to the corresponding weight to improve the quality of MRI boundary reconstruction.

[0046] The mean μ, variance σ, and weight w of each component of the mixed dense module are all generated through the convolution layer, and the maximum likelihood estimation is combined with back propagation to determine the weight of the network. Let θ be the set of model parameters to be solved, and the mixed dense is shown in formula (1):

[0047]

[0048] Among them, the sum of weights w is equal to 1, m = 1, 2, p is the Laplace distribution, M is the total number of distillation modules, x i is the i-th artifact-bearing and undersampled MRI medical image, and θ is the parameter of the hybrid dense block.

[0049] The likelihood function P(θ) is shown in formula (2):

[0050]

[0051] Among them, y n is the nth fully sampled MRI medical image predicted by the teacher, x n is the nth artifact-bearing and undersampled MRI medical image.

[0052] Through maximum likelihood estimation, this embodiment can obtain the distillation loss function.

[0053] The consistency between different components in MDN essentially characterizes the accidental uncertainty of the edge of MRI medical images. The higher the uncertainty, the sharper the edge. The present invention captures the accidental uncertainty of the prediction variance by evaluating the differential entropy of the continuous mixed distribution, as shown in formula (3):

[0054] h=-∫p(x i ;θ)logp(x i ;θ)dx i (3)

[0055] Where p is the Laplace distribution, h is the random uncertainty value, and x i is the i-th artifact-bearing and undersampled MRI medical image, and θ is the parameter of the hybrid dense block.

[0056] The present invention proposes a hybrid dense knowledge distillation framework MUTED, which takes super-resolution task as auxiliary task and super-resolution reconstruction as main task.

[0057] Medical image super-resolution refers to recovering high-resolution details from undersampled medical images to make the images clearer and more realistic. In order to improve the super-resolution capability of the teacher network, the present invention uses the undersampled image X' to process the feature X after the encoder (Encoder) f The teacher model is supervised to encourage the decoder to learn useful features in the super-resolution task. After training, the weights of the decoder in the teacher model are assigned to the decoder in the student model. The student model takes the undersampled image X with artifacts as input and consists of a decoder and a mixed dense block (MDN block) to eliminate artifacts in MRI medical images and obtain fully sampled high-resolution images.

[0058] The present invention uses IMDN as the decoder. The hybrid dense module consists of two distillation modules and a convolutional layer that calculates the weights of different distillation modules. When pre-training the teacher, the alignment loss L is used. al and the super-resolution loss L sr . Alignment loss L al Compute compact features X f The mean absolute error (MAE) between the image X and the undersampled image X' is defined as:

[0059]

[0060] Where H and W are the height and width of the low-resolution image, respectively, and X′ ij is the pixel at the (i,j) index position of the undersampled image X', For compact feature X f The pixel at the (i, j) index position of the image. This encourages the teacher model to extract features that are useful for super-resolution tasks, resulting in higher quality images. Super-resolution loss L sr Calculate the teacher model output image Y T The mean absolute error (MSE) between the high-resolution image Y is defined as:

[0061]

[0062] Among them, H s is the height of the high-resolution image Y, W s is the width of the high-resolution image Y,ij is the pixel at the (i,j) index position of the high-resolution image Y, Output image Y for the teacher model T The pixel at the (i,j) index position.

[0063] The content of the hybrid dense knowledge distillation framework proposed in this invention is as follows Figure 2 shown.

[0064] Figure 2 This is an overview of the mixed dense distillation framework. Figure 3 In the hybrid dense distillation framework, the high-resolution image Y is used as the input of the teacher model (i.e., the auxiliary task). The teacher model consists of an encoder, a decoder, and an upsampling block. The decoder in the student model (i.e., the main task) shares the same network architecture as the decoder in the auxiliary task. The main task includes a decoder and a hybrid dense block. The hybrid dense module includes two distillation modules and a convolutional layer, such as Figure 4 Using the undersampled image X with artifacts as input, the student model is trained for super-resolution reconstruction in a hybrid dense knowledge distillation framework.

[0065] The total loss of the teacher network is the alignment loss L al and the super-resolution loss L sr The sum, determined by the parameter α t balance:

[0066] L t =L sr +α t L al (6)

[0067] In the student model, the present invention uses the reconstruction loss L rec and distillation loss L dis The reconstruction loss L rec Calculate the student output image Y S The mean absolute error (MAE) between the high-resolution image Y is defined as:

[0068]

[0069] Among them, S is the student output image Y S The pixel at the (i,j) index position.

[0070] The present invention constructs distillation loss based on hybrid dense network, calculates the corresponding mean μ, variance σ and corresponding weight w for each component in MDN, and obtains distillation loss by weighted combination of each component (Laplacian or Gaussian distribution) through weight w, which is:

[0071]

[0072] Among them, L dis is the distillation loss, w is the weight, H′ is the height of the high-resolution image, W′ is the width of the high-resolution image, σ is the variance, M is the total number of distillation modules, m is the mth distillation module, C is the number of channels, i is the i-th channel index, j is the j-th height pixel index, k is the k-th width pixel index, and Y T is the prediction of the teacher model, and μ is the mean of the predictions of the student model.

[0073] The overall loss function of the hybrid dense knowledge distillation framework MUTED is shown in formula (9), where α s As parameters to control the distillation:

[0074] L s =L rec +α s L dis (9)

[0075] This paper designs a hybrid dense knowledge distillation network MUTED for super-resolution reconstruction of medical images.

[0076] In order to evaluate the framework of the present invention, the process of generating two images with low resolution artifacts and low resolution artifacts from high resolution images needs to be close to the actual acquisition of MRI scanner. Assume that y is a complex value of the fully sampled k-space. By x=F -1 (y) gets the size H s ×W s Fully sampled high-resolution image, where F -1 is the two-dimensional inverse fast Fourier transform (FFT). In order to improve the acquisition efficiency of MRI, a binary mask operator Z is used to determine the Cartesian acquisition trajectory and use it to select a subset of k-space points. The undersampled k-space is obtained by y′=Z⊙y, where ⊙ represents the corresponding multiplication of the same-position elements. Correspondingly, the zero-filled image can be expressed as x′=F -1 (y). In order to obtain a low-resolution image X' of size h / s×w / s ​​(s is the scale factor), the outside of y is first truncated with the desired factor to obtain a low-resolution image X' without artifacts, and then a two-dimensional fast Fourier transform F is applied to it to obtain the low-resolution input data X with artifacts in the main task.

[0077] First, pre-train the teacher model;

[0078] The high-resolution image Y is used as the input of the teacher model (auxiliary task), and data augmentation is used during training, including random rotation and horizontal flipping. After the encoder, the compact feature X is obtained. f. The encoder consists of 4 convolutional layers, and the PReLU activation function is used in every two convolutional layers. In order to train the decoder to learn more super-resolution knowledge, formula (4) is used to constrain the compact features using undersampled images. The present invention uses the IMDN network backbone as the decoder, and uses a layer of Conv2d and a PixelShuffle to form an upsampling block. The teacher model parameters that obtain the highest restoration quality during the training process are saved and used to initialize the student model decoder to impart super-resolution task knowledge to the student.

[0079] Then, train the knowledge distillation framework student model;

[0080] Initialize the student model and assign the saved teacher model encoder weights to the student model encoder. The artifact-bearing and undersampled image X is used as the input of the student model (main task), and the corresponding high-resolution image Y is used as the input of the teacher model (auxiliary task). The student model includes a decoder and a mixed dense module. The mixed dense module includes two distillation modules and a convolutional layer. The two weights output by the convolutional layer constrain distillation module 1 and distillation module 2 respectively. Each distillation module includes an upsampling block and a convolutional layer for calculating the mean and variance. The module for calculating the mean has two convolutional layers and PReLU activation functions in the two convolutional layers. The module for calculating the variance has two convolutional layers and a Softplus, and PReLU activation functions are also added to the two convolutional layers. While training the student model for super-resolution reconstruction, the "knowledge" learned by the image Y under the fixed saved teacher model parameters is passed to the student model through formula (8). The upsampling result of each distillation module is multiplied by the corresponding distillation module weight to obtain the final output result. Backpropagation is performed with maximum likelihood estimation to determine the weights of the network, and the aleatory uncertainty of the prediction variance is visualized using formula (3).

[0081] Finally, the student model parameters that obtain the highest recovery quality during the training process are saved to obtain the mixed dense knowledge distillation network MUTED.

[0082] For example, this embodiment takes the classic IMDN network backbone as an example to obtain the hybrid dense knowledge distillation network MUTED designed by the present invention.

[0083] The specific network structure of the teacher model network is shown in Table 1 below.

[0084] Table 1

[0085]

[0086]

[0087] The specific network structure of the student model network is shown in Table 2 below.

[0088] Table 2

[0089]

[0090]

[0091]

[0092] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A super-resolution reconstruction method for MRI medical images, characterized in that: include: Acquire artifact-bearing and undersampled MRI medical images, and construct a hybrid dense knowledge distillation model, wherein the hybrid dense knowledge distillation model includes: a teacher model and a student model; The student model comprises: a second decoder and a mixed dense block; wherein the second decoder and the mixed dense block are connected; The teacher model includes: an encoder, a first decoder and an upsampling block; wherein the encoder, the first decoder and the upsampling block are connected in sequence; the mixed dense block includes: a convolutional layer, a first distillation submodule and a second distillation submodule; wherein the weights of the first distillation submodule and the second distillation submodule are obtained through the convolutional layer; Training the hybrid dense knowledge distillation model, inputting the artifact-bearing and under-sampled MRI medical image into the trained hybrid dense knowledge distillation model to obtain an artifact-free fully sampled MRI medical image; Training the teacher model includes: Acquire a high-resolution image, perform data enhancement processing on the high-resolution image, input the high-resolution image after data enhancement into the encoder, and acquire compact features; Acquire an undersampled image, constrain the compact feature based on the undersampled image, input the constrained compact feature into the first decoder, and obtain the teacher model output image through the upsampling block, while saving the weight of the first decoder; training the student model includes: Inputting the upsampled MRI historical medical image into the convolutional layer to perform weight calculation to obtain a first weight and a second weight; Input the upsampled MRI historical medical image into the first distillation submodule and the second distillation submodule respectively to calculate the mean and variance, obtain a first mean, a second mean, a first variance, and a second variance, obtain a first Laplace distribution according to the first mean and the first variance, and obtain a second Laplace distribution according to the second mean and the second variance; According to the first weight and the second weight, weighting the first Laplace distribution and the second Laplace distribution respectively to obtain a mixed dense distribution; Calculating the differential entropy of the mixed dense distribution to obtain a visualized accidental uncertainty map; Training the student model includes: Acquire an artifact-bearing and under-sampled MRI historical medical image, assign the saved weight of the first decoder to the second decoder, input the artifact-bearing and under-sampled MRI historical medical image into the second decoder, and acquire an artifact-free fully-sampled MRI historical medical image through the mixed dense block.

2. The method for super-resolution reconstruction of MRI medical images according to claim 1, characterized in that: Assigning the saved weight of the first decoder to the second decoder comprises: The method of transferring the knowledge of the teacher model to the student model through knowledge distillation is: Among them, L dis is the distillation loss, w is the weight, H' is the height of the high-resolution image, W' is the width of the high-resolution image, σ is the variance, M is the total number of distillation modules, m is the mth distillation module, C is the number of channels, i is the i-th channel index, j is the j-th height pixel index, k is the k-th width pixel index, and Y T is the prediction of the teacher model, and μ is the mean of the predictions of the student model.

3. The method for super-resolution reconstruction of MRI medical images according to claim 1, characterized in that: The method to obtain a visual aleatoric uncertainty map is: h=-∫p(x i ;θ)logp(x i ;θ)dx i Where p is the Laplace distribution, h is the random uncertainty value, and x i is the i-th artifact-bearing and undersampled MRI medical image, and θ is the parameter of the hybrid dense block.